import os import signal import json from collections import defaultdict, deque import ast import re import subprocess import logging from pathlib import Path from typing import List, Dict, Tuple, Any, Optional from threading import RLock from contextlib import contextmanager from functools import lru_cache import time import hashlib import pathspec import requests import numpy as np from mcp.server.fastmcp import FastMCP from langchain_ollama import OllamaEmbeddings from langchain_community.vectorstores import Chroma from rank_bm25 import BM25Okapi import javalang import sqlparse import sqlglot from enhanced_toon import EnhancedToon from graph.graph import LocalGraph os.environ["ANONYMIZED_TELEMETRY"] = "False" # ----------------------------- # Configuration # ----------------------------- logging.basicConfig( level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s", handlers=[ logging.FileHandler("mcp_codebase.log"), # Log to file logging.StreamHandler() # Also log to console ] ) logger = logging.getLogger("rag-mcp") OLLAMA_BASE_URL = "http://127.0.0.1:11434" os.environ["OLLAMA_HOST"] = OLLAMA_BASE_URL os.environ["OLLAMA_API_BASE"] = OLLAMA_BASE_URL CODEBASE_PATH = Path("./working_repo") CODEBASE_PATH.mkdir(exist_ok=True) VECTOR_DB_PATH = Path(os.environ.get("VECTOR_DB_PATH", "./chroma_db")) BM25_INDEX_PATH = Path(os.environ.get("BM25_INDEX_PATH", "./bm25_index.json")) RAGIGNORE_PATH = CODEBASE_PATH / ".ragignore" # Models / behavior EMBEDDING_MODEL = "bge-m3" # Ollama local embedding model RERANKER_MODEL_OLLAMA = "dengcao/Qwen3-Reranker-8B:Q4_K_M" # prompt-based reranker via Ollama ENABLE_RERANK = True # toggle reranking RERANK_TOP_N = 10 # Hybrid weights (vector first: 80%, BM25: 20%) VECTOR_WEIGHT = 0.8 BM25_WEIGHT = 0.2 # Embedding batch configuration EMBEDDING_BATCH_SIZE = 200 # Contextual weighting configuration CONTEXTUAL_WEIGHTS = { # SQL weights "sql_schema": 1.3, # CREATE TABLE, ALTER TABLE, schema definitions "sql_function": 1.2, # Function/procedure definitions "sql_view": 1.15, # VIEW definitions "sql_index": 1.1, # INDEX definitions # Python weights "python_class": 1.25, # Class definitions (enhanced) "python_function": 1.15, # Function definitions "python_method": 1.2, # Method definitions (enhanced) "python_async": 1.1, # Async functions # Go weights "go_type": 1.25, # Type definitions (struct, interface) "go_function": 1.15, # Function definitions "go_method": 1.2, # Method definitions (enhanced) "go_struct": 1.3, # Struct definitions (enhanced) # Java weights "java_class": 1.3, # Class definitions (enhanced) "java_interface": 1.25, # Interface definitions (enhanced) "java_method": 1.2, # Method definitions (enhanced) "java_constructor": 1.15, # Constructor definitions (enhanced) "java_enum": 1.2, # Enum definitions (enhanced) "java_field": 1.1, # Field definitions (enhanced) # Rust weights (new) "rust_struct": 1.3, # Struct definitions "rust_enum": 1.25, # Enum definitions "rust_trait": 1.35, # Trait definitions (very important in Rust) "rust_impl": 1.2, # Implementation blocks "rust_function": 1.15, # Function definitions "rust_method": 1.2, # Method definitions "rust_async": 1.1, # Async functions "rust_unsafe": 1.15, # Unsafe blocks/functions # TypeScript/JavaScript weights (for Svelte scripts) "typescript_interface": 1.25, # Interface definitions "typescript_class": 1.2, # Class definitions "typescript_function": 1.15, # Function definitions "typescript_type": 1.2, # Type definitions # Svelte weights (new) "svelte_component": 1.4, # Component definitions (very important) "svelte_script": 1.1, # Script blocks "svelte_style": 1.05, # Style blocks "svelte_markup": 1.15, # Markup/template "svelte_prop": 1.25, # Component props (enhanced) "svelte_reactive": 1.2, # Reactive declarations # Default weight for unspecified types "default": 1.0 } SUPPORTED_EXTENSIONS = { ".py", ".go", ".rs", ".ts", ".tsx", ".js", ".jsx", ".svelte", ".sql", ".pgsql", ".sh", ".bash", ".zsh", ".rb", ".java", ".cpp", ".c", ".h", ".hpp" } # MCP mcp = FastMCP("codebase-rag") # Global state vectorstore: Optional[Chroma] = None bm25: Optional[BM25Okapi] = None bm25_corpus: List[str] = [] chunks_metadata: List[Dict[str, Any]] = [] embeddings: Optional[OllamaEmbeddings] = None _startup_lock = RLock() index_build_time: float = 0.0 # ----------------------------- # Thread-Safe Context Manager # ----------------------------- @contextmanager def safe_search(): """Thread-safe context manager for search operations.""" with _startup_lock: if vectorstore is None or bm25 is None: raise RuntimeError("Indexes not ready. Please wait for initialization or rebuild.") yield # ----------------------------- # File Hashing for Cache Invalidation # ----------------------------- def get_file_hash(filepath: Path) -> str: """Generate hash of file content + mtime for cache key.""" try: stat = filepath.stat() content_sample = filepath.read_bytes()[:1024] # First 1KB for speed hash_input = f"{stat.st_mtime}:{stat.st_size}:{content_sample}".encode() return hashlib.md5(hash_input).hexdigest() except Exception: return "" # ----------------------------- # Contextual Weight Calculator # ----------------------------- def calculate_contextual_weight(metadata: Dict[str, Any]) -> float: """ Calculate contextual weight multiplier based on chunk metadata. Higher weights for schema-defining statements, function definitions, etc. """ language = metadata.get("language", "").lower() chunk_type = metadata.get("type", "").lower() block_type = metadata.get("block_type", "").lower() # SQL-specific weights if language == "sql": if chunk_type in ("create", "alter", "create table", "alter table"): return CONTEXTUAL_WEIGHTS.get("sql_schema", 1.0) elif chunk_type in ("function", "procedure", "create function", "create procedure"): return CONTEXTUAL_WEIGHTS.get("sql_function", 1.0) elif chunk_type in ("view", "create view"): return CONTEXTUAL_WEIGHTS.get("sql_view", 1.0) elif chunk_type in ("index", "create index"): return CONTEXTUAL_WEIGHTS.get("sql_index", 1.0) # Python-specific weights (enhanced) elif language == "python": if chunk_type == "classdef": return CONTEXTUAL_WEIGHTS.get("python_class", 1.0) elif chunk_type in ("functiondef", "asyncfunctiondef"): if metadata.get("is_method"): return CONTEXTUAL_WEIGHTS.get("python_method", 1.0) elif chunk_type == "asyncfunctiondef" or metadata.get("is_async"): return CONTEXTUAL_WEIGHTS.get("python_async", 1.0) else: return CONTEXTUAL_WEIGHTS.get("python_function", 1.0) # Go-specific weights (enhanced) elif language == "go": if chunk_type in ("struct", "interface"): return CONTEXTUAL_WEIGHTS.get("go_struct", 1.0) elif chunk_type == "type": return CONTEXTUAL_WEIGHTS.get("go_type", 1.0) elif chunk_type == "method": return CONTEXTUAL_WEIGHTS.get("go_method", 1.0) elif chunk_type == "func": return CONTEXTUAL_WEIGHTS.get("go_function", 1.0) # Java-specific weights (enhanced) elif language == "java": if chunk_type == "class": return CONTEXTUAL_WEIGHTS.get("java_class", 1.0) elif chunk_type == "interface": return CONTEXTUAL_WEIGHTS.get("java_interface", 1.0) elif chunk_type == "method": return CONTEXTUAL_WEIGHTS.get("java_method", 1.0) elif chunk_type == "constructor": return CONTEXTUAL_WEIGHTS.get("java_constructor", 1.0) elif chunk_type == "enum": return CONTEXTUAL_WEIGHTS.get("java_enum", 1.0) elif chunk_type == "field": return CONTEXTUAL_WEIGHTS.get("java_field", 1.0) # Rust-specific weights (new) elif language == "rust": if chunk_type == "struct": return CONTEXTUAL_WEIGHTS.get("rust_struct", 1.0) elif chunk_type == "enum": return CONTEXTUAL_WEIGHTS.get("rust_enum", 1.0) elif chunk_type == "trait": return CONTEXTUAL_WEIGHTS.get("rust_trait", 1.0) elif chunk_type == "impl": return CONTEXTUAL_WEIGHTS.get("rust_impl", 1.0) elif chunk_type == "function": if metadata.get("is_async"): return CONTEXTUAL_WEIGHTS.get("rust_async", 1.0) elif metadata.get("is_unsafe"): return CONTEXTUAL_WEIGHTS.get("rust_unsafe", 1.0) elif metadata.get("is_method", False): return CONTEXTUAL_WEIGHTS.get("rust_method", 1.0) else: return CONTEXTUAL_WEIGHTS.get("rust_function", 1.0) # TypeScript/JavaScript weights (for Svelte scripts) elif language == "typescript": if chunk_type == "interface": return CONTEXTUAL_WEIGHTS.get("typescript_interface", 1.0) elif chunk_type == "class": return CONTEXTUAL_WEIGHTS.get("typescript_class", 1.0) elif chunk_type == "function": return CONTEXTUAL_WEIGHTS.get("typescript_function", 1.0) elif chunk_type == "type": return CONTEXTUAL_WEIGHTS.get("typescript_type", 1.0) # Svelte-specific weights (new) elif language == "svelte": if chunk_type == "component": return CONTEXTUAL_WEIGHTS.get("svelte_component", 1.0) elif chunk_type == "prop": return CONTEXTUAL_WEIGHTS.get("svelte_prop", 1.0) elif chunk_type == "reactive": return CONTEXTUAL_WEIGHTS.get("svelte_reactive", 1.0) elif block_type == "markup": return CONTEXTUAL_WEIGHTS.get("svelte_markup", 1.0) elif block_type == "script": return CONTEXTUAL_WEIGHTS.get("svelte_script", 1.0) elif block_type == "style": return CONTEXTUAL_WEIGHTS.get("svelte_style", 1.0) # CSS weights (for Svelte styles) elif language == "css": if block_type == "style": return CONTEXTUAL_WEIGHTS.get("svelte_style", 1.0) return CONTEXTUAL_WEIGHTS.get("default", 1.0) def apply_contextual_weights_to_embeddings(embeddings_list: List[List[float]], metadata_list: List[Dict[str, Any]]) -> List[List[float]]: """ Apply contextual weights to embedding vectors by scaling them. This biases the vector space without changing the embedding model. """ weighted_embeddings = [] if len(embeddings_list) != len(metadata_list): raise ValueError( f"Embedding/metadata mismatch: {len(embeddings_list)} embeddings vs {len(metadata_list)} metadata. " "This will break downstream systems (Chroma)." ) for embedding, metadata in zip(embeddings_list, metadata_list): weight = calculate_contextual_weight(metadata) # Convert to numpy for easier manipulation emb_array = np.array(embedding) # Scale the embedding vector by the weight # This effectively increases the magnitude, making it more "important" weighted_emb = emb_array * weight # Optionally normalize to maintain consistent vector magnitudes # Comment out if you want the raw weighted vectors norm = np.linalg.norm(weighted_emb) if norm > 0: weighted_emb = weighted_emb / norm * np.linalg.norm(emb_array) weighted_embeddings.append(weighted_emb.tolist()) if weight != 1.0: logger.debug(f"Applied weight {weight:.2f} to {metadata.get('file')}:{metadata.get('type')}") return weighted_embeddings # ----------------------------- # Utility: .ragignore # ----------------------------- def load_ragignore(base_path: Path) -> pathspec.PathSpec: if RAGIGNORE_PATH.exists(): with open(RAGIGNORE_PATH, "r", encoding="utf-8") as f: return pathspec.PathSpec.from_lines("gitwildmatch", f) default_patterns = [ ".git/", "node_modules/", "__pycache__/", "*.pyc", "*.so", "*.dll", "*.exe", "venv/", ".venv/", "dist/", "build/", "*.log", "*.tmp", ".DS_Store" ] return pathspec.PathSpec.from_lines("gitwildmatch", default_patterns) # ----------------------------- # Python AST chunking with caching # ----------------------------- @lru_cache(maxsize=1000) def parse_python_file_cached(filepath_str: str, file_hash: str) -> tuple: """Cached Python AST parsing with enhanced metadata for graph relationships. ALWAYS returns a tuple of (text, metadata_tuple) entries. If parsing fails or produces no chunks, return a single fallback chunk containing the whole file text and minimal metadata. """ filepath = Path(filepath_str) chunks = [] logger.info(f"🐍 Parsing Python file: {filepath}") try: code = filepath.read_text(encoding="utf-8") tree = ast.parse(code, filename=str(filepath)) lines = code.splitlines(keepends=True) # Track imports for module relationships imports = [] module_imports = {} # Maps names to modules for better tracking # First pass: Extract all imports for node in ast.walk(tree): if isinstance(node, ast.Import): for alias in node.names: import_name = alias.asname if alias.asname else alias.name imports.append(alias.name) module_imports[import_name] = alias.name elif isinstance(node, ast.ImportFrom): module = node.module or "" for alias in node.names: import_name = alias.asname if alias.asname else alias.name full_import = f"{module}.{alias.name}" if module else alias.name imports.append(full_import) module_imports[import_name] = full_import logger.info(f" Found {len(imports)} imports") # Second pass: Process declarations decl_count = 0 for node in ast.walk(tree): chunk_text = None metadata = None # FUNCTION DEFINITIONS if isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef)): docstring = ast.get_docstring(node) or "" start_line = node.lineno - 1 end_line = getattr(node, "end_lineno", start_line + 1) # Guard the end_line if end_line <= start_line: end_line = start_line + 1 chunk_code = "".join(lines[start_line:end_line]) metadata = { "file": str(filepath), "name": node.name, "type": "function" if isinstance(node, ast.FunctionDef) else "async_function", "line": node.lineno, "language": "python", } # Extract function metadata func_meta = _extract_function_metadata(node, code) metadata.update(func_meta) # Include imports if imports: metadata["imports"] = imports # Build chunk text chunk_text = f"File: {filepath}\nType: {metadata['type']}\nName: {node.name}\n" # Add decorators if func_meta.get("decorators"): chunk_text += f"Decorators: {', '.join(func_meta['decorators'])}\n" # Add arguments args = _format_arguments(node.args) chunk_text += f"Arguments: {args}\n" # Add return type returns = _extract_return_annotation(node) if returns: chunk_text += f"Returns: {returns}\n" # Add async/generator info if isinstance(node, ast.AsyncFunctionDef): chunk_text += "Async: yes\n" if func_meta.get("is_generator"): chunk_text += "Generator: yes\n" # Add docstring if docstring: chunk_text += f"Docstring:\n{docstring}\n\n" chunk_text += f"Code:\n{chunk_code}" # CLASS DEFINITIONS elif isinstance(node, ast.ClassDef): docstring = ast.get_docstring(node) or "" start_line = node.lineno - 1 end_line = getattr(node, "end_lineno", start_line + 1) if end_line <= start_line: end_line = start_line + 1 chunk_code = "".join(lines[start_line:end_line]) metadata = { "file": str(filepath), "name": node.name, "type": "class", "line": node.lineno, "language": "python", } # Extract class metadata class_meta = _extract_class_metadata(node, code) metadata.update(class_meta) # Include imports if imports: metadata["imports"] = imports # Build chunk text chunk_text = f"File: {filepath}\nType: class\nName: {node.name}\n" # Add decorators if class_meta.get("decorators"): chunk_text += f"Decorators: {', '.join(class_meta['decorators'])}\n" # Add base classes if class_meta.get("bases"): chunk_text += f"Base Classes: {', '.join(class_meta['bases'])}\n" # Add methods if class_meta.get("methods"): chunk_text += f"Methods: {', '.join(class_meta['methods'])}\n" # Add properties if class_meta.get("properties"): chunk_text += f"Properties: {', '.join(class_meta['properties'])}\n" # Add class variables if class_meta.get("class_variables"): chunk_text += f"Class Variables: {', '.join(class_meta['class_variables'])}\n" # Add docstring if docstring: chunk_text += f"Docstring:\n{docstring}\n\n" chunk_text += f"Code:\n{chunk_code}" # MODULE-LEVEL ASSIGNMENTS (constants, globals) elif isinstance(node, ast.Assign) and node.lineno: # Only process top-level assignments (not inside functions/classes) if _is_module_level(node, tree): for target in node.targets: if isinstance(target, ast.Name): # Check if it looks like a constant (UPPER_CASE) if target.id.isupper(): start_line = node.lineno - 1 end_line = getattr(node, "end_lineno", start_line + 1) if end_line <= start_line: end_line = start_line + 1 chunk_code = "".join(lines[start_line:end_line]) metadata = { "file": str(filepath), "name": target.id, "type": "constant", "line": node.lineno, "language": "python", } # Try to extract value try: if hasattr(ast, 'unparse'): value_str = ast.unparse(node.value) metadata["value"] = value_str except Exception: pass chunk_text = f"File: {filepath}\nType: constant\nName: {target.id}\n" chunk_text += f"Code:\n{chunk_code}" # ANNOTATED ASSIGNMENTS (type-annotated variables) elif isinstance(node, ast.AnnAssign) and node.lineno: if _is_module_level(node, tree) and isinstance(node.target, ast.Name): start_line = node.lineno - 1 end_line = getattr(node, "end_lineno", start_line + 1) if end_line <= start_line: end_line = start_line + 1 chunk_code = "".join(lines[start_line:end_line]) metadata = { "file": str(filepath), "name": node.target.id, "type": "variable", "line": node.lineno, "language": "python", } # Extract type annotation if node.annotation: try: if hasattr(ast, 'unparse'): type_str = ast.unparse(node.annotation) metadata["var_type"] = type_str except Exception: pass chunk_text = f"File: {filepath}\nType: variable\nName: {node.target.id}\n" if metadata.get("var_type"): chunk_text += f"Type: {metadata['var_type']}\n" chunk_text += f"Code:\n{chunk_code}" # Add the chunk if we created one if chunk_text and metadata: chunks.append({ "text": chunk_text, "metadata": metadata }) decl_count += 1 logger.info(f" Parsed {decl_count} declarations from {filepath}") except SyntaxError as e: logger.warning(f"Python syntax error in {filepath}: {e}") except Exception as e: logger.warning(f"Failed to parse Python file {filepath}: {e}") # Return parsed chunks or fallback if chunks: logger.info(f"βœ… Successfully parsed {len(chunks)} chunks from {filepath}") return tuple((c["text"], tuple(c["metadata"].items())) for c in chunks) # Fallback: emit a whole-file chunk logger.warning(f"❌ No chunks created from Python parsing for {filepath}. Emitting fallback whole-file chunk.") try: full_code = filepath.read_text(encoding="utf-8") except Exception: full_code = "" fallback_meta = { "file": str(filepath), "name": filepath.stem, "type": "module", "line": 1, "language": "python", } return ((full_code, tuple(fallback_meta.items())),) def _is_module_level(node, tree): """Check if a node is at module level (not inside a function or class).""" for parent in ast.walk(tree): if isinstance(parent, (ast.FunctionDef, ast.AsyncFunctionDef, ast.ClassDef)): for child in ast.walk(parent): if child is node: return False return True def _extract_import_info(node): """Extract import information from Import or ImportFrom nodes.""" if isinstance(node, ast.Import): return [alias.name for alias in node.names] elif isinstance(node, ast.ImportFrom): module = node.module or "" return [f"{module}.{alias.name}" if module else alias.name for alias in node.names] return [] def _extract_function_metadata(node, code): """Extract enhanced metadata from function nodes.""" metadata = {} # Extract decorators if node.decorator_list: decorators = [] for dec in node.decorator_list: try: if hasattr(ast, 'unparse'): decorators.append(ast.unparse(dec)) elif isinstance(dec, ast.Name): decorators.append(dec.id) elif isinstance(dec, ast.Call) and isinstance(dec.func, ast.Name): decorators.append(dec.func.id) except Exception: pass if decorators: metadata["decorators"] = decorators # Extract parameters with type hints parameters = [] if node.args.args: for arg in node.args.args: param_info = arg.arg if arg.annotation: try: if hasattr(ast, 'unparse'): param_info += f": {ast.unparse(arg.annotation)}" except Exception: pass parameters.append(param_info) if parameters: metadata["parameters"] = parameters # Check for special function types if node.args.vararg: metadata["has_varargs"] = True if node.args.kwarg: metadata["has_kwargs"] = True # Check if it's a generator (contains yield) for child in ast.walk(node): if isinstance(child, (ast.Yield, ast.YieldFrom)): metadata["is_generator"] = True break # Check for property decorator if any(isinstance(dec, ast.Name) and dec.id == "property" for dec in node.decorator_list): metadata["is_property"] = True # Check for staticmethod/classmethod for dec in node.decorator_list: if isinstance(dec, ast.Name): if dec.id == "staticmethod": metadata["is_static"] = True elif dec.id == "classmethod": metadata["is_classmethod"] = True # Extract return type annotation if node.returns: try: if hasattr(ast, 'unparse'): metadata["return_type"] = ast.unparse(node.returns) except Exception: pass return metadata def _extract_class_metadata(node, code): """Extract enhanced metadata from class nodes.""" metadata = {} # Extract decorators if node.decorator_list: decorators = [] for dec in node.decorator_list: try: if hasattr(ast, 'unparse'): decorators.append(ast.unparse(dec)) elif isinstance(dec, ast.Name): decorators.append(dec.id) except Exception: pass if decorators: metadata["decorators"] = decorators # Extract base classes bases = [] for base in node.bases: try: if hasattr(ast, 'unparse'): bases.append(ast.unparse(base)) elif isinstance(base, ast.Name): bases.append(base.id) except Exception: pass if bases: metadata["bases"] = bases # Extract methods and properties methods = [] properties = [] static_methods = [] class_methods = [] class_variables = [] for item in node.body: if isinstance(item, (ast.FunctionDef, ast.AsyncFunctionDef)): # Check decorators is_property = any( isinstance(dec, ast.Name) and dec.id == "property" for dec in item.decorator_list ) is_static = any( isinstance(dec, ast.Name) and dec.id == "staticmethod" for dec in item.decorator_list ) is_classmethod = any( isinstance(dec, ast.Name) and dec.id == "classmethod" for dec in item.decorator_list ) if is_property: properties.append(item.name) elif is_static: static_methods.append(item.name) elif is_classmethod: class_methods.append(item.name) else: methods.append(item.name) elif isinstance(item, ast.AnnAssign) and isinstance(item.target, ast.Name): # Class variable with type annotation class_variables.append(item.target.id) elif isinstance(item, ast.Assign): # Class variable without type annotation for target in item.targets: if isinstance(target, ast.Name): class_variables.append(target.id) if methods: metadata["methods"] = methods if properties: metadata["properties"] = properties if static_methods: metadata["static_methods"] = static_methods if class_methods: metadata["class_methods"] = class_methods if class_variables: metadata["class_variables"] = class_variables # Check for metaclass for keyword in node.keywords: if keyword.arg == "metaclass": try: if hasattr(ast, 'unparse'): metadata["metaclass"] = ast.unparse(keyword.value) elif isinstance(keyword.value, ast.Name): metadata["metaclass"] = keyword.value.id except Exception: pass return metadata def _format_arguments(args): """Format function arguments with type hints.""" arg_strs = [] # Regular arguments for arg in args.args: arg_str = arg.arg if arg.annotation: try: if hasattr(ast, 'unparse'): arg_str += f": {ast.unparse(arg.annotation)}" except Exception: pass arg_strs.append(arg_str) # *args if args.vararg: vararg_str = f"*{args.vararg.arg}" if args.vararg.annotation: try: if hasattr(ast, 'unparse'): vararg_str += f": {ast.unparse(args.vararg.annotation)}" except Exception: pass arg_strs.append(vararg_str) # **kwargs if args.kwarg: kwarg_str = f"**{args.kwarg.arg}" if args.kwarg.annotation: try: if hasattr(ast, 'unparse'): kwarg_str += f": {ast.unparse(args.kwarg.annotation)}" except Exception: pass arg_strs.append(kwarg_str) return ", ".join(arg_strs) if arg_strs else "None" def _extract_return_annotation(node): """Extract return type annotation from function node.""" if not node.returns: return None try: if hasattr(ast, 'unparse'): return ast.unparse(node.returns) except Exception: pass return None def parse_python_file(filepath: Path) -> List[Dict]: """Parse Python file with caching.""" file_hash = get_file_hash(filepath) cached_result = parse_python_file_cached(str(filepath), file_hash) return [{"text": text, "metadata": dict(meta)} for text, meta in cached_result] @lru_cache(maxsize=1000) def parse_sql_file_cached(filepath_str: str, file_hash: str) -> tuple: """Cached SQL parsing. Returns tuple for hashability.""" filepath = Path(filepath_str) try: sql_code = filepath.read_text(encoding="utf-8") except Exception as e: logger.warning(f"parse_sql_file: cannot read {filepath}: {e}") return tuple() chunks: List[Dict[str, Any]] = [] statements = [s for s in sqlparse.split(sql_code) if s and s.strip()] lines = sql_code.splitlines() search_pos = 0 for stmt in statements: start_idx = sql_code.find(stmt, search_pos) if start_idx == -1: start_idx = sql_code.find(stmt) if start_idx == -1: start_line = len(lines) else: start_line = sql_code[:start_idx].count("\n") + 1 search_pos = start_idx + len(stmt) leading_comments = include_leading_comments(lines, start_line) stmt_clean = stmt.strip() meta: Dict[str, Any] = { "file": str(filepath), "name": None, "type": None, "line": start_line, "language": "sql" } try: # Special handling for CREATE TYPE ENUM (PostgreSQL) enum_match = re.match( r'CREATE\s+TYPE\s+(\w+)\s+AS\s+ENUM\s*\((.*?)\)', stmt_clean, re.IGNORECASE | re.DOTALL ) if enum_match: enum_name = enum_match.group(1) enum_values = enum_match.group(2) meta["type"] = "create type enum" meta["name"] = enum_name meta["enum_values"] = ",".join([v.strip().strip("'\"") for v in enum_values.split(",") if v.strip()]) chunk_text = f"File: {filepath}\nType: {meta['type']}\nName: {meta['name']}\n" if leading_comments: chunk_text += f"\nComments:\n{leading_comments}\n\n" chunk_text += f"SQL:\n{stmt_clean}" chunks.append({"text": chunk_text, "metadata": meta}) continue # Special handling for CREATE EXTENSION extension_match = re.match( r'CREATE\s+EXTENSION\s+(?:IF\s+NOT\s+EXISTS\s+)?(\w+)', stmt_clean, re.IGNORECASE ) if extension_match: extension_name = extension_match.group(1) meta["type"] = "create extension" meta["name"] = extension_name chunk_text = f"File: {filepath}\nType: {meta['type']}\nName: {meta['name']}\n" if leading_comments: chunk_text += f"\nComments:\n{leading_comments}\n\n" chunk_text += f"SQL:\n{stmt_clean}" chunks.append({"text": chunk_text, "metadata": meta}) continue # Special handling for CREATE TRIGGER trigger_match = re.match( r'CREATE\s+(?:OR\s+REPLACE\s+)?TRIGGER\s+(\w+)', stmt_clean, re.IGNORECASE ) if trigger_match: trigger_name = trigger_match.group(1) meta["type"] = "create trigger" meta["name"] = trigger_name chunk_text = f"File: {filepath}\nType: {meta['type']}\nName: {meta['name']}\n" if leading_comments: chunk_text += f"\nComments:\n{leading_comments}\n\n" chunk_text += f"SQL:\n{stmt_clean}" chunks.append({"text": chunk_text, "metadata": meta}) continue # Try sqlglot for standard SQL parsed = sqlglot.parse_one(stmt_clean, read="postgres") stmt_type = getattr(parsed, "key", None) or parsed.token_type if hasattr(parsed, "token_type") else None meta["type"] = str(stmt_type).lower() if stmt_type else "statement" tables = [t.this for t in parsed.find_all(sqlglot.exp.Table)] meta["tables"] = ",".join([str(t) for t in tables if isinstance(t, str)]) ctes = [] for cte in parsed.find_all(sqlglot.exp.CTE): try: alias = cte.alias_or_name if alias: ctes.append(alias) except Exception: pass meta["ctes"] = ",".join(ctes) funcs = [] for f in parsed.find_all(sqlglot.exp.Func): try: name = f.name if name: funcs.append(name) except Exception: pass meta["functions"] = ",".join(funcs) if meta.get("tables") and meta["tables"]: meta["name"] = meta["tables"].split(",")[0] elif ctes: meta["name"] = ctes[0] elif funcs: meta["name"] = funcs[0] chunk_text = f"File: {filepath}\nType: {meta.get('type')}\nName: {meta.get('name')}\n" if leading_comments: chunk_text += f"\nComments:\n{leading_comments}\n\n" chunk_text += f"SQL:\n{stmt_clean}" chunks.append({"text": chunk_text, "metadata": meta}) continue except Exception: pass # Fallback: simple heuristics if sqlglot fails try: parsed_tok = sqlparse.parse(stmt_clean)[0] first_token = parsed_tok.token_first(skip_cm=True) stmt_type = first_token.value.upper() if first_token else "UNKNOWN" except Exception: stmt_type = "UNKNOWN" meta["type"] = stmt_type.lower() chunk_text = f"File: {filepath}\nType: {meta['type']}\nName: {meta.get('name')}\n" if leading_comments: chunk_text += f"\nComments:\n{leading_comments}\n\n" chunk_text += f"SQL:\n{stmt_clean}" chunks.append({"text": chunk_text, "metadata": meta}) return tuple((c["text"], tuple(c["metadata"].items())) for c in chunks) def parse_sql_file(filepath: Path) -> List[Dict[str, Any]]: """Parse SQL file with caching.""" file_hash = get_file_hash(filepath) cached_result = parse_sql_file_cached(str(filepath), file_hash) return [{"text": text, "metadata": dict(meta)} for text, meta in cached_result] def include_leading_comments(lines: List[str], stmt_start_line: int, max_context_lines: int = 8) -> str: """ Collect contiguous comment lines immediately above stmt_start_line (1-based). Supports single-line comments (--), and block comments (/* ... */). Returns a string containing the comment block (no leading comment markers). """ comments: List[str] = [] idx = stmt_start_line - 2 # convert to 0-based index of the line above the statement collected = 0 # Collect single-line comments (-- ...) and blank lines (include blanks between comments) while idx >= 0 and collected < max_context_lines: raw = lines[idx] stripped = raw.strip() if stripped.startswith("--"): # remove leading -- and optional space comments.insert(0, stripped[2:].lstrip()) idx -= 1 collected += 1 continue # block comment end if stripped.endswith("*/"): # gather entire block block_lines = [] while idx >= 0: block_line = lines[idx].rstrip() block_lines.insert(0, block_line) if block_line.strip().startswith("/*"): break idx -= 1 # remove /* and */ markers and join cleaned = [] for bl in block_lines: s = bl.strip() if s.startswith("/*"): s = s[2:].lstrip() if s.endswith("*/"): s = s[:-2].rstrip() cleaned.append(s) comments = cleaned + comments break # blank line - allow if we already have comments (so comments may be separated by one blank) if stripped == "": if comments: comments.insert(0, "") idx -= 1 collected += 1 continue else: break # otherwise stop when non-comment found break # join into paragraph if comments: # remove possible leading/trailing empty lines while comments and comments[0] == "": comments.pop(0) while comments and comments[-1] == "": comments.pop() return "\n".join(comments).strip() return "" # ----------------------------- # Go AST via helper binary with caching # ----------------------------- # Replace the parse_go_file_cached function in mcp_codebase.py (around line 685) # The issue is that the Go binary outputs lowercase JSON keys, but Python expects uppercase @lru_cache(maxsize=1000) def parse_go_file_cached(filepath_str: str, file_hash: str) -> tuple: """Cached Go parsing with enhanced metadata for graph relationships. ALWAYS returns a tuple of (text, metadata_tuple) entries. If the Go helper fails or produces no chunks, return a single fallback chunk containing the whole file text and minimal metadata. """ filepath = Path(filepath_str) chunks = [] logger.info(f"🐹 Parsing Go file: {filepath}") go_helper = Path("./tools/parse_go_ast") if not go_helper.is_file(): logger.warning(f"Go parser binary not found: {go_helper}. Using fallback regex parsing for {filepath}") return _parse_go_fallback(filepath_str, file_hash) try: code = filepath.read_text(encoding="utf-8") lines = code.splitlines(keepends=True) proc = subprocess.run( [str(go_helper), str(filepath)], capture_output=True, text=True, check=True, timeout=20 ) decls = json.loads(proc.stdout) logger.info(f" AST helper found {len(decls)} declarations") for i, d in enumerate(decls): # Use lowercase keys (matching the Go JSON output) start_line = max(0, d.get("start_line", 1) - 1) end_line = d.get("end_line", start_line + 1) # Guard the end_line to be at least start_line+1 if end_line <= start_line: end_line = start_line + 1 # Make slice safe even if end_line > len(lines) chunk_code = "".join(lines[start_line:end_line]) name = d.get("name", "") typ = d.get("type", "unknown") doc_comment = d.get("doc_comment", "") receiver = d.get("receiver", "") full_name = d.get("full_name", name) fields = d.get("fields", []) methods = d.get("methods", []) # Skip package declarations (they don't contribute to code graph) if typ == "package": continue # Skip empty or invalid declarations if not name or typ == "unknown": continue metadata = { "file": str(filepath), "name": name, "type": typ, "line": start_line + 1, "language": "go", } # Build comprehensive chunk text chunk_text = f"File: {filepath}\nType: {typ}\nName: {name}\n" # Add receiver information for methods if receiver: metadata["receiver"] = receiver # Extract receiver type without pointer for type matching receiver_base = receiver.replace('*', '').strip() metadata["receiver_base_type"] = receiver_base chunk_text += f"Receiver: {receiver}\n" # Add full name if different from name (methods have qualified names) if full_name and full_name != name: metadata["full_name"] = full_name # Add documentation if doc_comment: metadata["docstring"] = doc_comment chunk_text += f"Doc:\n{doc_comment}\n\n" # Add fields for structs if fields: metadata["fields"] = fields chunk_text += f"Fields: {', '.join(fields)}\n" # Add methods for interfaces if methods: metadata["interface_methods"] = methods chunk_text += f"Methods: {', '.join(methods)}\n" chunk_text += f"Code:\n{chunk_code}" logger.info(f" Declaration {i+1}: {typ} {name}") chunks.append({ "text": chunk_text, "metadata": metadata }) except subprocess.CalledProcessError as e: stderr = e.stderr if e.stderr else "unknown error" logger.warning(f"Go AST helper failed for {filepath}: {stderr}") except json.JSONDecodeError as e: logger.warning(f"Go AST helper output invalid JSON for {filepath}: {e}") except Exception as e: logger.warning(f"Go AST helper failed for {filepath}: {e}") # If parser produced chunks, return them in the (text, metadata_tuple) format if chunks: logger.info(f"βœ… Successfully parsed {len(chunks)} chunks from {filepath}") return tuple((c["text"], tuple(c["metadata"].items())) for c in chunks) # Fallback to regex-based parsing logger.warning(f"⚠️ Go AST helper produced no chunks for {filepath}. Trying regex fallback.") return _parse_go_fallback(filepath_str, file_hash) def _parse_go_fallback(filepath_str: str, file_hash: str) -> tuple: """Fallback regex-based Go parser when AST helper fails or is unavailable.""" filepath = Path(filepath_str) chunks = [] logger.info(f" Using regex-based Go parsing for {filepath}") try: content = filepath.read_text(encoding="utf-8") lines = content.splitlines(keepends=True) # Helper function to find matching closing brace def find_brace_block_end(start_line: int) -> int: """Find the line with the matching closing brace.""" brace_count = 0 for i in range(start_line, len(lines)): for char in lines[i]: if char == '{': brace_count += 1 elif char == '}': brace_count -= 1 if brace_count == 0: return i + 1 # Include closing brace line return len(lines) # Helper function to extract preceding comments def extract_comments(start_line: int) -> str: """Extract comments immediately before a declaration.""" comments = [] for i in range(max(0, start_line - 10), start_line): line = lines[i].strip() if line.startswith('//'): comments.append(line[2:].strip()) elif line.startswith('/*'): # Multi-line comment comment_lines = [] for j in range(i, start_line): comment_lines.append(lines[j].strip()) if '*/' in lines[j]: break comment_text = ' '.join(comment_lines) comment_text = comment_text.replace('/*', '').replace('*/', '').strip() comments.append(comment_text) elif line and not line.startswith('package') and not line.startswith('import'): # Non-empty, non-comment line before declaration - stop looking break return '\n'.join(comments) if comments else "" # Pattern to match Go function declarations (including methods) func_pattern = re.compile( r'^func\s+(?:\(([^)]+)\)\s+)?(\w+)\s*\([^)]*\)(?:\s*\([^)]*\)|\s+[\w\[\].*]+)?\s*\{', re.MULTILINE ) for match in func_pattern.finditer(content): receiver_part = match.group(1) # Receiver (optional) func_name = match.group(2) start_pos = match.start() start_line = content[:start_pos].count('\n') end_line = find_brace_block_end(start_line) func_code = "".join(lines[start_line:end_line]) comments = extract_comments(start_line) # Parse receiver if present receiver_type = "" receiver_base_type = "" if receiver_part: # Extract type from receiver (e.g., "u *User" -> "*User") parts = receiver_part.strip().split() if len(parts) >= 2: receiver_type = parts[-1] elif len(parts) == 1: receiver_type = parts[0] receiver_base_type = receiver_type.replace('*', '').strip() metadata = { "file": str(filepath), "name": func_name, "type": "method" if receiver_type else "function", "line": start_line + 1, "language": "go", } chunk_text = f"File: {filepath}\nType: {metadata['type']}\nName: {func_name}\n" # Add receiver information for methods if receiver_type: metadata["receiver"] = receiver_type metadata["receiver_base_type"] = receiver_base_type chunk_text += f"Receiver: {receiver_type}\n" if comments: metadata["docstring"] = comments chunk_text += f"Comments:\n{comments}\n\n" chunk_text += f"Code:\n{func_code}" chunks.append({ "text": chunk_text, "metadata": metadata }) # Pattern to match struct declarations struct_pattern = re.compile(r'^type\s+(\w+)\s+struct\s*\{', re.MULTILINE) for match in struct_pattern.finditer(content): struct_name = match.group(1) start_pos = match.start() start_line = content[:start_pos].count('\n') end_line = find_brace_block_end(start_line) struct_code = "".join(lines[start_line:end_line]) comments = extract_comments(start_line) # Extract field names from struct field_names = [] struct_body_match = re.search(r'struct\s*\{([^}]+)\}', struct_code, re.DOTALL) if struct_body_match: field_lines = struct_body_match.group(1).split('\n') for line in field_lines: line = line.strip() if not line or line.startswith('//'): continue # Match field declarations: FieldName Type or FieldName, FieldName2 Type field_match = re.match(r'(\w+(?:\s*,\s*\w+)*)\s+[\w\[\]*.]+', line) if field_match: # Handle multiple fields on one line field_str = field_match.group(1) fields = [f.strip() for f in field_str.split(',')] field_names.extend(fields) else: # Check for embedded fields (just a type, no name) embedded_match = re.match(r'(\*?[\w.]+)\s*(?://|$)', line) if embedded_match and not line.startswith('type'): field_names.append(f"embedded:{embedded_match.group(1)}") metadata = { "file": str(filepath), "name": struct_name, "type": "struct", "line": start_line + 1, "language": "go", } chunk_text = f"File: {filepath}\nType: struct\nName: {struct_name}\n" if comments: metadata["docstring"] = comments chunk_text += f"Comments:\n{comments}\n\n" if field_names: metadata["fields"] = field_names chunk_text += f"Fields: {', '.join(field_names)}\n" chunk_text += f"Code:\n{struct_code}" chunks.append({ "text": chunk_text, "metadata": metadata }) # Pattern to match interface declarations interface_pattern = re.compile(r'^type\s+(\w+)\s+interface\s*\{', re.MULTILINE) for match in interface_pattern.finditer(content): interface_name = match.group(1) start_pos = match.start() start_line = content[:start_pos].count('\n') end_line = find_brace_block_end(start_line) interface_code = "".join(lines[start_line:end_line]) comments = extract_comments(start_line) # Extract method signatures from interface method_names = [] interface_body_match = re.search(r'interface\s*\{([^}]+)\}', interface_code, re.DOTALL) if interface_body_match: method_lines = interface_body_match.group(1).split('\n') for line in method_lines: line = line.strip() if not line or line.startswith('//'): continue # Match method declarations: MethodName(params) returnType method_match = re.match(r'(\w+)\s*\([^)]*\)', line) if method_match: method_names.append(method_match.group(1)) metadata = { "file": str(filepath), "name": interface_name, "type": "interface", "line": start_line + 1, "language": "go", } chunk_text = f"File: {filepath}\nType: interface\nName: {interface_name}\n" if comments: metadata["docstring"] = comments chunk_text += f"Comments:\n{comments}\n\n" if method_names: metadata["interface_methods"] = method_names chunk_text += f"Methods: {', '.join(method_names)}\n" chunk_text += f"Code:\n{interface_code}" chunks.append({ "text": chunk_text, "metadata": metadata }) # Pattern to match type aliases and other type declarations type_alias_pattern = re.compile(r'^type\s+(\w+)\s+(?!struct|interface)(.+?)(?:\n|$)', re.MULTILINE) for match in type_alias_pattern.finditer(content): type_name = match.group(1) type_def = match.group(2).strip() start_pos = match.start() start_line = content[:start_pos].count('\n') # For simple type aliases, just capture the line type_code = lines[start_line] if start_line < len(lines) else match.group(0) comments = extract_comments(start_line) metadata = { "file": str(filepath), "name": type_name, "type": "type", "line": start_line + 1, "language": "go", } chunk_text = f"File: {filepath}\nType: type alias\nName: {type_name}\n" if comments: metadata["docstring"] = comments chunk_text += f"Comments:\n{comments}\n\n" chunk_text += f"Definition: {type_def}\n" chunk_text += f"Code:\n{type_code}" chunks.append({ "text": chunk_text, "metadata": metadata }) # Pattern to match const and var declarations const_var_pattern = re.compile(r'^(const|var)\s+(?:\(|(\w+))', re.MULTILINE) for match in const_var_pattern.finditer(content): decl_type = match.group(1) # 'const' or 'var' single_name = match.group(2) # Name if single declaration start_pos = match.start() start_line = content[:start_pos].count('\n') if single_name: # Single declaration var_code = lines[start_line] if start_line < len(lines) else match.group(0) comments = extract_comments(start_line) metadata = { "file": str(filepath), "name": single_name, "type": decl_type, "line": start_line + 1, "language": "go", } chunk_text = f"File: {filepath}\nType: {decl_type}\nName: {single_name}\n" if comments: metadata["docstring"] = comments chunk_text += f"Comments:\n{comments}\n\n" chunk_text += f"Code:\n{var_code}" chunks.append({ "text": chunk_text, "metadata": metadata }) else: # Block declaration (const ( ... ) or var ( ... )) end_line = find_brace_block_end(start_line) if '(' in lines[start_line] else start_line + 1 # For simplicity in fallback, just capture the whole block block_code = "".join(lines[start_line:end_line]) comments = extract_comments(start_line) # Extract individual names from block names = re.findall(r'^\s*(\w+)\s+', block_code, re.MULTILINE) if names: for name in names: metadata = { "file": str(filepath), "name": name, "type": decl_type, "line": start_line + 1, "language": "go", } chunk_text = f"File: {filepath}\nType: {decl_type}\nName: {name}\n" if comments: metadata["docstring"] = comments chunk_text += f"Code:\n{block_code}" chunks.append({ "text": chunk_text, "metadata": metadata }) except Exception as e: logger.warning(f"Regex-based Go parsing failed for {filepath}: {e}") # Return parsed chunks or final fallback if chunks: logger.info(f"βœ… Regex fallback parsed {len(chunks)} chunks from {filepath}") return tuple((c["text"], tuple(c["metadata"].items())) for c in chunks) # Final fallback: emit a whole-file chunk logger.warning(f"❌ No chunks created from Go parsing for {filepath}. Emitting fallback whole-file chunk.") try: full_code = filepath.read_text(encoding="utf-8") except Exception: full_code = "" fallback_meta = { "file": str(filepath), "name": filepath.name, "type": "file", "line": 1, "language": "go", } return ((full_code, tuple(fallback_meta.items())),) def find_brace_block_end_go(lines: List[str], start_line: int) -> int: """Find the end of a Go brace block starting at start_line.""" brace_count = 0 in_block = False for i in range(start_line, len(lines)): line = lines[i] for char in line: if char == '{': brace_count += 1 in_block = True elif char == '}': brace_count -= 1 if in_block and brace_count == 0: return i return len(lines) - 1 def extract_go_comments(lines: List[str], func_line_index: int, max_lines: int = 10) -> str: """Extract leading comments above a Go function/struct/interface.""" comments = [] idx = func_line_index - 1 while idx >= 0 and len(comments) < max_lines: line = lines[idx].strip() if line.startswith('//'): # Remove // and trim comments.insert(0, line[2:].strip()) idx -= 1 elif line.startswith('/*') or '*/' in line: # Block comment - collect entire block if '*/' in line and not line.startswith('/*'): # End of block, go backwards to find start block_lines = [line] idx -= 1 while idx >= 0: block_line = lines[idx].strip() block_lines.insert(0, block_line) if block_line.startswith('/*'): break idx -= 1 # Clean up block comment markers block_text = ' '.join(block_lines) block_text = block_text.replace('/*', '').replace('*/', '').replace('*', '').strip() comments.insert(0, block_text) idx -= 1 else: # Single line block comment clean = line.replace('/*', '').replace('*/', '').strip() comments.insert(0, clean) idx -= 1 elif line == '': # Allow one blank line if comments: idx -= 1 else: break else: break # Clean up while comments and comments[0] == '': comments.pop(0) while comments and comments[-1] == '': comments.pop(-1) return '\n'.join(comments) def parse_go_file(filepath: Path) -> List[Dict]: """Parse Go file with caching.""" file_hash = get_file_hash(filepath) cached_result = parse_go_file_cached(str(filepath), file_hash) if not cached_result: return chunk_text_file(filepath, "go") return [{"text": text, "metadata": dict(meta)} for text, meta in cached_result] # ----------------------------- # Java parsing (javalang + manual body extraction) with caching # ----------------------------- @lru_cache(maxsize=1000) def parse_java_file_cached(filepath_str: str, file_hash: str) -> tuple: """Cached Java parsing with enhanced metadata for graph relationships. ALWAYS returns a tuple of (text, metadata_tuple) entries. If parsing fails or produces no chunks, return a single fallback chunk containing the whole file text and minimal metadata. """ filepath = Path(filepath_str) chunks = [] logger.info(f"β˜• Parsing Java file: {filepath}") try: code = filepath.read_text(encoding="utf-8") lines = code.splitlines(keepends=True) try: tree = javalang.parse.parse(code) except javalang.parser.JavaSyntaxError as e: logger.warning(f"Java syntax error in {filepath}: {e}. Using fallback chunk.") fallback_meta = { "file": str(filepath), "name": filepath.name, "type": "file", "line": 1, "language": "java", } return ((code, tuple(fallback_meta.items())),) def find_brace_block_end(start_idx: int) -> int: """Find the closing brace for a block starting at start_idx.""" brace_count = 0 for i in range(start_idx, len(lines)): for char in lines[i]: if char == '{': brace_count += 1 elif char == '}': brace_count -= 1 if brace_count == 0: return i + 1 # Include the closing brace line return len(lines) # If no closing brace found, take rest of file def extract_docstring(node) -> str | None: """Extract documentation/Javadoc from a node.""" if hasattr(node, 'documentation') and node.documentation: return node.documentation.strip() return None def get_visibility(modifiers) -> str: """Extract visibility from modifiers set.""" if not modifiers: return "package-private" if 'public' in modifiers: return "public" if 'protected' in modifiers: return "protected" if 'private' in modifiers: return "private" return "package-private" def extract_imports(tree) -> list[str]: """Extract all imports from the compilation unit.""" imports = [] if hasattr(tree, 'imports') and tree.imports: for imp in tree.imports: imports.append(imp.path) return imports def extract_annotations(node) -> list[str]: """Extract annotations from a node.""" annotations = [] if hasattr(node, 'annotations') and node.annotations: for anno in node.annotations: if hasattr(anno, 'name'): annotations.append(anno.name) return annotations def type_to_string(type_obj) -> str: """Convert a javalang type object to string.""" if not type_obj: return "void" if hasattr(type_obj, 'name'): base_type = type_obj.name # Handle generic types if hasattr(type_obj, 'arguments') and type_obj.arguments: args = [type_to_string(arg.type) if hasattr(arg, 'type') else str(arg) for arg in type_obj.arguments] return f"{base_type}<{', '.join(args)}>" # Handle array dimensions if hasattr(type_obj, 'dimensions') and type_obj.dimensions: return base_type + "[]" * len(type_obj.dimensions) return base_type return str(type_obj) # Extract package and imports once package_name = tree.package.name if hasattr(tree, 'package') and tree.package else None imports = extract_imports(tree) decl_count = 0 for path, node in tree: metadata = None chunk_text = None start_line = 0 end_line = 0 # CLASS DECLARATION if isinstance(node, javalang.tree.ClassDeclaration): start_line = node.position.line - 1 if node.position else 0 end_line = find_brace_block_end(start_line) chunk_text = "".join(lines[start_line:end_line]) extends_class = node.extends.name if node.extends else None implements_interfaces = [] if hasattr(node, 'implements') and node.implements: implements_interfaces = [type_to_string(impl) for impl in node.implements] metadata = { "file": str(filepath), "name": node.name, "type": "class", "line": start_line + 1, "language": "java", "visibility": get_visibility(node.modifiers), "package": package_name, "imports": imports, "extends": extends_class, "implements": implements_interfaces, "annotations": extract_annotations(node), } # Extract generics if hasattr(node, 'type_parameters') and node.type_parameters: metadata["generics"] = [tp.name for tp in node.type_parameters] # Check for abstract/final/static modifiers if node.modifiers: if 'abstract' in node.modifiers: metadata["is_abstract"] = True if 'final' in node.modifiers: metadata["is_final"] = True if 'static' in node.modifiers: metadata["is_static"] = True docstring = extract_docstring(node) if docstring: metadata["docstring"] = docstring # METHOD DECLARATION elif isinstance(node, javalang.tree.MethodDeclaration): if not hasattr(node, 'position') or not node.position: continue start_line = node.position.line - 1 end_line = find_brace_block_end(start_line) chunk_text = "".join(lines[start_line:end_line]) # Find enclosing class class_name = None for p in reversed(path): if isinstance(p, javalang.tree.ClassDeclaration): class_name = p.name break return_type = type_to_string(node.return_type) parameters = [] if node.parameters: for param in node.parameters: param_type = type_to_string(param.type) param_str = f"{param.name}: {param_type}" if param.varargs: param_str = f"{param.name}: {param_type}..." parameters.append(param_str) metadata = { "file": str(filepath), "name": node.name, "type": "method", "class": class_name, "line": start_line + 1, "language": "java", "visibility": get_visibility(node.modifiers), "package": package_name, "return_type": return_type, "parameters": parameters, "annotations": extract_annotations(node), } # Extract generics if hasattr(node, 'type_parameters') and node.type_parameters: metadata["generics"] = [tp.name for tp in node.type_parameters] # Check modifiers if node.modifiers: if 'abstract' in node.modifiers: metadata["is_abstract"] = True if 'final' in node.modifiers: metadata["is_final"] = True if 'static' in node.modifiers: metadata["is_static"] = True if 'synchronized' in node.modifiers: metadata["is_synchronized"] = True if 'native' in node.modifiers: metadata["is_native"] = True docstring = extract_docstring(node) if docstring: metadata["docstring"] = docstring # FIELD DECLARATION elif isinstance(node, javalang.tree.FieldDeclaration): if not hasattr(node, 'position') or not node.position: continue start_line = node.position.line - 1 end_line = start_line + 1 chunk_text = lines[start_line] if start_line < len(lines) else "" field_type = type_to_string(node.type) # Process each declarator (there can be multiple fields on one line) for declarator in node.declarators: field_metadata = { "file": str(filepath), "name": declarator.name, "type": "field", "line": start_line + 1, "language": "java", "visibility": get_visibility(node.modifiers), "package": package_name, "field_type": field_type, "annotations": extract_annotations(node), } # Check modifiers if node.modifiers: if 'static' in node.modifiers: field_metadata["is_static"] = True if 'final' in node.modifiers: field_metadata["is_final"] = True if 'volatile' in node.modifiers: field_metadata["is_volatile"] = True if 'transient' in node.modifiers: field_metadata["is_transient"] = True docstring = extract_docstring(node) if docstring: field_metadata["docstring"] = docstring field_text = f"File: {filepath}\nType: Field\nName: {declarator.name}\n" field_text += f"Field Type: {field_type}\nVisibility: {field_metadata['visibility']}\n" if field_metadata.get("annotations"): field_text += f"Annotations: {', '.join(field_metadata['annotations'])}\n" field_text += f"Code:\n{chunk_text}" chunks.append({ "text": field_text, "metadata": field_metadata }) decl_count += 1 # Skip the standard processing since we handled fields inline continue # CONSTRUCTOR DECLARATION elif isinstance(node, javalang.tree.ConstructorDeclaration): if not hasattr(node, 'position') or not node.position: continue start_line = node.position.line - 1 end_line = find_brace_block_end(start_line) chunk_text = "".join(lines[start_line:end_line]) # Find enclosing class class_name = None for p in reversed(path): if isinstance(p, javalang.tree.ClassDeclaration): class_name = p.name break parameters = [] if node.parameters: for param in node.parameters: param_type = type_to_string(param.type) parameters.append(f"{param.name}: {param_type}") metadata = { "file": str(filepath), "name": f"{class_name}", "type": "constructor", "class": class_name, "line": start_line + 1, "language": "java", "visibility": get_visibility(node.modifiers), "package": package_name, "parameters": parameters, "annotations": extract_annotations(node), } docstring = extract_docstring(node) if docstring: metadata["docstring"] = docstring # ENUM DECLARATION elif isinstance(node, javalang.tree.EnumDeclaration): start_line = node.position.line - 1 if node.position else 0 end_line = find_brace_block_end(start_line) chunk_text = "".join(lines[start_line:end_line]) # Extract enum constants variants = [] if hasattr(node, 'body') and node.body: for item in node.body: if isinstance(item, javalang.tree.EnumConstantDeclaration): variants.append(item.name) implements_interfaces = [] if hasattr(node, 'implements') and node.implements: implements_interfaces = [type_to_string(impl) for impl in node.implements] metadata = { "file": str(filepath), "name": node.name, "type": "enum", "line": start_line + 1, "language": "java", "visibility": get_visibility(node.modifiers), "package": package_name, "implements": implements_interfaces, "variants": variants, "annotations": extract_annotations(node), } docstring = extract_docstring(node) if docstring: metadata["docstring"] = docstring # INTERFACE DECLARATION elif isinstance(node, javalang.tree.InterfaceDeclaration): start_line = node.position.line - 1 if node.position else 0 end_line = find_brace_block_end(start_line) chunk_text = "".join(lines[start_line:end_line]) extends_interfaces = [] if hasattr(node, 'extends') and node.extends: extends_interfaces = [type_to_string(ext) for ext in node.extends] # Extract method signatures methods = [] if hasattr(node, 'body') and node.body: for item in node.body: if isinstance(item, javalang.tree.MethodDeclaration): methods.append(item.name) metadata = { "file": str(filepath), "name": node.name, "type": "interface", "line": start_line + 1, "language": "java", "visibility": get_visibility(node.modifiers), "package": package_name, "extends": extends_interfaces, "methods": methods, "annotations": extract_annotations(node), } # Extract generics if hasattr(node, 'type_parameters') and node.type_parameters: metadata["generics"] = [tp.name for tp in node.type_parameters] docstring = extract_docstring(node) if docstring: metadata["docstring"] = docstring # ANNOTATION DECLARATION elif isinstance(node, javalang.tree.AnnotationDeclaration): start_line = node.position.line - 1 if node.position else 0 end_line = find_brace_block_end(start_line) chunk_text = "".join(lines[start_line:end_line]) metadata = { "file": str(filepath), "name": node.name, "type": "annotation", "line": start_line + 1, "language": "java", "visibility": get_visibility(node.modifiers), "package": package_name, } docstring = extract_docstring(node) if docstring: metadata["docstring"] = docstring # Build chunk text and add to chunks list if metadata and chunk_text is not None: text_parts = [f"File: {filepath}"] text_parts.append(f"Type: {metadata['type']}") text_parts.append(f"Name: {metadata['name']}") if metadata.get('visibility'): text_parts.append(f"Visibility: {metadata['visibility']}") if metadata.get('package'): text_parts.append(f"Package: {metadata['package']}") if metadata.get('extends'): text_parts.append(f"Extends: {metadata['extends']}") if metadata.get('implements'): text_parts.append(f"Implements: {', '.join(metadata['implements'])}") if metadata.get('return_type'): text_parts.append(f"Returns: {metadata['return_type']}") if metadata.get('parameters'): text_parts.append(f"Parameters: {', '.join(metadata['parameters'])}") if metadata.get('annotations'): text_parts.append(f"Annotations: {', '.join(metadata['annotations'])}") if metadata.get('generics'): text_parts.append(f"Generics: {', '.join(metadata['generics'])}") text_parts.append(f"Code:\n{chunk_text}") chunks.append({ "text": "\n".join(text_parts), "metadata": metadata }) decl_count += 1 logger.info(f" Parsed {decl_count} declarations from {filepath}") except Exception as e: logger.warning(f"Failed to parse Java file {filepath}: {e}") # Return parsed chunks or fallback if chunks: logger.info(f"βœ… Successfully parsed {len(chunks)} chunks from {filepath}") return tuple((c["text"], tuple(c["metadata"].items())) for c in chunks) # Fallback: emit a whole-file chunk logger.warning(f"❌ No chunks created from Java AST for {filepath}. Emitting fallback whole-file chunk.") try: full_code = filepath.read_text(encoding="utf-8") except Exception: full_code = "" fallback_meta = { "file": str(filepath), "name": filepath.name, "type": "file", "line": 1, "language": "java", } return ((full_code, tuple(fallback_meta.items())),) def parse_java_file(filepath: Path) -> List[Dict]: """Parse Java source file with caching.""" fileHash = hash(filepath.read_bytes()) cached_result = parse_java_file_cached(str(filepath), fileHash) return [{"text": text, "metadata": dict(meta)} for text, meta in cached_result] # ----------------------------- # Svelte support with caching # ----------------------------- @lru_cache(maxsize=500) def parse_svelte_file_cached(filepath_str: str, file_hash: str) -> tuple: """Cached Svelte parsing with enhanced multi-language graph relationships. ALWAYS returns a tuple of (text, metadata_tuple) entries. If parsing fails or produces no chunks, return a single fallback chunk containing the whole file text and minimal metadata. """ filepath = Path(filepath_str) chunks = [] logger.info(f"🎨 Parsing Svelte file: {filepath}") try: text = filepath.read_text(encoding="utf-8") lines = text.splitlines(keepends=True) # Extract component name from filename for graph relationships component_name = filepath.stem if component_name and component_name[0].islower(): component_name = component_name[0].upper() + component_name[1:] # PascalCase # Track component props and context for graph relationships component_props = [] component_stores = [] component_imports = [] reactive_declarations = [] # Parse script blocks script_matches = list(re.finditer(r"]*?))?>(.*?)", text, flags=re.DOTALL)) script_idx = 0 for script_match in script_matches: script_attrs = script_match.group(1) or "" script_content = script_match.group(2) script_start_pos = script_match.start(2) script_start_line = text[:script_start_pos].count("\n") + 1 # Determine script context is_module = 'context="module"' in script_attrs or "context='module'" in script_attrs is_typescript = 'lang="ts"' in script_attrs or "lang='ts'" in script_attrs script_type = "script_module" if is_module else "script" script_lang = "typescript" if is_typescript else "javascript" logger.info(f" Processing {script_type} block (lang: {script_lang})") # Try TypeScript/JavaScript parser ts_helper = Path("./tools/parse_ts.js") parsed_with_helper = False if ts_helper.is_file(): tmp = None try: # Create temporary file for parsing tmp = filepath.parent / f".tmp_{filepath.name}_{script_idx}.ts" tmp.write_text(script_content, encoding="utf-8") proc = subprocess.run( ["node", str(ts_helper), str(tmp)], capture_output=True, text=True, check=True, timeout=15 ) decls = json.loads(proc.stdout) script_lines = script_content.splitlines(keepends=True) logger.info(f" TS helper found {len(decls)} declarations") for i, d in enumerate(decls): start_rel = max(0, d.get("start_line", 1) - 1) end_rel = d.get("end_line", start_rel + 1) # Guard the end_line if end_rel <= start_rel: end_rel = start_rel + 1 chunk_code = "".join(script_lines[start_rel:end_rel]) doc = d.get("doc_comment", "").strip() abs_start = script_start_line + start_rel decl_name = d.get("name", "") decl_type = d.get("type", "unknown") # Skip invalid declarations if not decl_name or decl_type == "unknown": continue # Analyze declaration for Svelte-specific patterns is_exported = "export" in chunk_code is_prop = False is_reactive = False is_store = False if decl_type == "variable" and not is_module: # Check for prop: export let prop = ... if re.search(r'\bexport\s+let\s+' + re.escape(decl_name), chunk_code): component_props.append(decl_name) is_prop = True decl_type = "prop" # Check for reactive declaration: $: reactiveVar = ... if chunk_code.strip().startswith('$:'): reactive_declarations.append(decl_name) is_reactive = True decl_type = "reactive" # Check for store: $storeName if decl_name.startswith('$'): component_stores.append(decl_name) is_store = True # Track imports if "import" in chunk_code and "from" in chunk_code: import_match = re.search(r'from\s+["\']([^"\']+)["\']', chunk_code) if import_match: component_imports.append(import_match.group(1)) # Build chunk text chunk_text = f"File: {filepath}\nType: {decl_type}\nName: {decl_name}\n" chunk_text += f"Component: {component_name}\nScript Type: {script_type}\n" if doc: chunk_text += f"Doc:\n{doc}\n\n" chunk_text += f"Code:\n{chunk_code}" metadata = { "file": str(filepath), "name": decl_name, "type": decl_type, "line": abs_start, "language": script_lang, "block_type": script_type, "component_name": component_name, } # Add Svelte-specific metadata if is_prop: metadata["is_exported"] = True metadata["component_prop"] = True if is_reactive: metadata["is_reactive"] = True if is_store: metadata["is_store"] = True if is_exported and not is_prop: metadata["is_exported"] = True # Function-specific metadata if decl_type == "function": if "createEventDispatcher" in chunk_code: metadata["is_event_dispatcher"] = True if "dispatch(" in chunk_code: metadata["dispatches_events"] = True if doc: metadata["docstring"] = doc chunks.append({ "text": chunk_text, "metadata": metadata }) logger.info(f" Declaration {i+1}: {decl_type} {decl_name}") parsed_with_helper = True except subprocess.CalledProcessError as e: logger.warning(f" TS parser subprocess failed for {filepath}: {e.stderr}") except json.JSONDecodeError as e: logger.warning(f" TS parser output invalid JSON for {filepath}: {e}") except Exception as e: logger.warning(f" TS parser failed for {filepath}: {e}") finally: if tmp and tmp.exists(): tmp.unlink(missing_ok=True) # Fallback: Create basic script chunk if parser failed if not parsed_with_helper: logger.info(f" Using fallback script chunk") chunks.append(_make_script_chunk( filepath, script_content, script_start_line, script_type, component_name, script_lang )) script_idx += 1 # Parse style blocks style_matches = list(re.finditer(r"]*?))?>(.*?)", text, flags=re.DOTALL)) for style_idx, style_match in enumerate(style_matches): style_attrs = style_match.group(1) or "" style_content = style_match.group(2) style_start_pos = style_match.start(2) style_start_line = text[:style_start_pos].count("\n") + 1 # Determine style language is_scss = 'lang="scss"' in style_attrs or "lang='scss'" in style_attrs is_less = 'lang="less"' in style_attrs or "lang='less'" in style_attrs is_scoped = "scoped" not in style_attrs # Svelte styles are scoped by default style_lang = "scss" if is_scss else ("less" if is_less else "css") # Extract CSS classes and selectors css_classes = list(set(re.findall(r'\.([a-zA-Z][\w-]*)\s*[{,:]', style_content))) css_ids = list(set(re.findall(r'#([a-zA-Z][\w-]*)\s*[{,:]', style_content))) css_selectors = list(set(re.findall(r'^([a-zA-Z][\w-]*)\s*{', style_content, re.MULTILINE))) # Extract CSS custom properties (variables) css_vars = list(set(re.findall(r'--([a-zA-Z][\w-]*)', style_content))) chunk_text = f"File: {filepath}\nType: style\nComponent: {component_name}\n" chunk_text += f"Language: {style_lang}\nScoped: {is_scoped}\n" if css_classes: chunk_text += f"CSS Classes: {', '.join(css_classes)}\n" if css_ids: chunk_text += f"CSS IDs: {', '.join(css_ids)}\n" if css_vars: chunk_text += f"CSS Variables: {', '.join(css_vars)}\n" chunk_text += f"Style Content:\n{style_content}" chunks.append({ "text": chunk_text, "metadata": { "file": str(filepath), "name": f"style_block_{style_idx}", "type": "style", "line": style_start_line, "language": style_lang, "block_type": "style", "component_name": component_name, "css_classes": css_classes, "css_ids": css_ids, "css_selectors": css_selectors, "css_variables": css_vars, "is_scoped": is_scoped, } }) # Parse markup (template) markup = text # Remove script blocks markup = re.sub(r"]*>.*?", "", markup, flags=re.DOTALL) # Remove style blocks markup = re.sub(r"]*>.*?", "", markup, flags=re.DOTALL) markup = markup.strip() if markup: # Extract Svelte-specific markup features used_components = list(set(re.findall(r'<([A-Z][a-zA-Z0-9]*)', markup))) prop_bindings = re.findall(r'\b(\w+)={([^}]+)}', markup) event_handlers = list(set(re.findall(r'on:(\w+)=', markup))) # Extract slots slots = list(set(re.findall(r' dict: """Create a fallback script chunk when TS parser fails.""" chunk_text = f"File: {filepath}\nType: {script_type}\nComponent: {component_name}\n" chunk_text += f"Language: {language}\nCode:\n{script_content}" # Extract basic info with regex exports = re.findall(r'export\s+(?:let|const|function|class)\s+(\w+)', script_content) imports = re.findall(r'from\s+["\']([^"\']+)["\']', script_content) metadata = { "file": str(filepath), "name": script_type, "type": script_type, "line": start_line, "language": language, "block_type": script_type, "component_name": component_name, } if exports: metadata["exports"] = exports if imports: metadata["imports"] = imports return { "text": chunk_text, "metadata": metadata } def parse_svelte_file(filepath: Path) -> List[Dict]: """Parse Svelte file with caching.""" file_hash = get_file_hash(filepath) cached_result = parse_svelte_file_cached(str(filepath), file_hash) if not cached_result: return chunk_text_file(filepath, "svelte") return [{"text": text, "metadata": dict(meta)} for text, meta in cached_result] # ----------------------------- # Heuristic chunking for other languages # ----------------------------- def chunk_text_file(filepath: Path, language: str) -> List[Dict]: try: lines = filepath.read_text(encoding="utf-8").splitlines() except Exception as e: logger.warning(f"chunk_text_file read failed {filepath}: {e}") return [] chunks = [] current_chunk = [] start_line = 0 for i, ln in enumerate(lines): stripped = ln.strip() if stripped.startswith(("def ", "func ", "fn ", "class ", "struct ", "impl ", "export ", "const ", "let ")) or stripped.startswith(("//", "#", "--")) or stripped.endswith("{"): if current_chunk: text = "\n".join(current_chunk) chunks.append({ "text": f"File: {filepath}\nLanguage: {language}\n\n{text}", "metadata": {"file": str(filepath), "name": f"chunk_{len(chunks)}", "type": "code_block", "line": start_line + 1, "language": language} }) current_chunk = [] start_line = i current_chunk.append(ln) if current_chunk: text = "\n".join(current_chunk) chunks.append({ "text": f"File: {filepath}\nLanguage: {language}\n\n{text}", "metadata": {"file": str(filepath), "name": f"chunk_{len(chunks)}", "type": "code_block", "line": start_line + 1, "language": language} }) return chunks def parse_rust_file(filepath: Path) -> List[Dict]: """Parse Rust file with caching.""" file_hash = get_file_hash(filepath) cached_result = parse_rust_file_cached(str(filepath), file_hash) return [{"text": text, "metadata": dict(meta)} for text, meta in cached_result] @lru_cache(maxsize=1000) def parse_rust_file_cached(filepath_str: str, file_hash: str) -> tuple: """Fixed Rust parsing that properly handles imports as metadata, not separate entities.""" filepath = Path(filepath_str) chunks = [] logger.info(f"πŸ”§ Parsing Rust file: {filepath}") rust_helper = Path("./tools/target/release/parse_rust_ast") if not rust_helper.is_file(): logger.warning(f"Rust parser binary not found: {rust_helper}. Using fallback.") return create_fallback_chunk(filepath) try: code = filepath.read_text(encoding="utf-8") lines = code.splitlines(keepends=True) proc = subprocess.run( [str(rust_helper), str(filepath)], capture_output=True, text=True, check=True, timeout=20 ) decls = json.loads(proc.stdout) logger.info(f" AST helper found {len(decls)} declarations") # Collect imports first to attach to actual code entities file_imports = [] code_decls = [] for d in decls: if d.get("type_") == "import": file_imports.extend(d.get("parameters", [])) else: code_decls.append(d) logger.info(f" Found {len(file_imports)} imports and {len(code_decls)} code declarations") # Process actual code entities (functions, structs, etc.) for i, d in enumerate(code_decls): start_line = max(1, d.get("start_line", 1)) end_line = max(start_line, d.get("end_line", start_line + 1)) # Ensure we don't exceed file bounds if start_line > len(lines): start_line = len(lines) if end_line > len(lines): end_line = len(lines) chunk_code = "".join(lines[start_line-1:end_line]) metadata = { "file": str(filepath), "name": d.get("name", "") or filepath.stem, "type": d.get("type_", "") or "unknown", "line": start_line, "language": "rust", "imports": file_imports, # Attach imports to ALL code entities in this file } # Enhanced metadata extraction for code entities enhanced_fields = [ "visibility", "is_async", "is_unsafe", "generics", "traits", "fields", "methods", "return_type", "parameters", "docstring", "calls" # calls from the specific entity ] for field in enhanced_fields: if field in d and d[field] is not None: metadata[field] = d[field] logger.info(f" Code {i+1}: {metadata.get('type', 'unknown')} {metadata.get('name', 'unnamed')}") # Build rich chunk text for CODE entities chunk_text = build_rust_code_chunk_text(metadata, chunk_code) chunks.append({ "text": chunk_text, "metadata": metadata }) # If we have no code declarations but have imports, create a file-level chunk if not chunks and file_imports: logger.info(" Creating file-level chunk for imports-only file") chunks.append(create_file_level_chunk(filepath, file_imports)) except Exception as e: logger.warning(f"Rust AST helper failed for {filepath}: {e}") return create_fallback_chunk(filepath) if chunks: logger.info(f"βœ… Successfully parsed {len(chunks)} chunks from {filepath}") return tuple((c["text"], tuple(c["metadata"].items())) for c in chunks) logger.warning(f"❌ No chunks created from Rust AST for {filepath}. Emitting fallback.") return create_fallback_chunk(filepath) def build_rust_code_chunk_text(metadata: dict, code: str) -> str: """Build comprehensive Rust code chunk text.""" parts = [ f"File: {metadata.get('file', '')}", f"Type: {metadata.get('type', 'unknown')}", f"Name: {metadata.get('name', '')}", ] # Add enhanced metadata for CODE entities if metadata.get("visibility"): parts.append(f"Visibility: {metadata['visibility']}") if metadata.get("is_async"): parts.append("Async: yes") if metadata.get("is_unsafe"): parts.append("Unsafe: yes") if metadata.get("generics"): parts.append(f"Generics: {', '.join(metadata['generics'])}") if metadata.get("traits"): parts.append(f"Implements: {', '.join(metadata['traits'])}") if metadata.get("fields"): parts.append(f"Fields: {', '.join(metadata['fields'])}") if metadata.get("imports"): parts.append(f"Imports: {', '.join(metadata['imports'])}") if metadata.get("calls"): parts.append(f"Calls: {', '.join(metadata['calls'])}") if metadata.get("docstring"): parts.append(f"Doc: {metadata['docstring']}") parts.append(f"Code:\n{code}") return "\n".join(parts) def create_file_level_chunk(filepath: Path, imports: list) -> dict: """Create a chunk for files that only have imports.""" try: full_code = filepath.read_text(encoding="utf-8") except Exception: full_code = "" metadata = { "file": str(filepath), "name": filepath.stem, "type": "module", # Treat import-only files as modules "line": 1, "language": "rust", "imports": imports, } chunk_text = f"File: {filepath}\nType: module\nName: {filepath.stem}\nImports: {', '.join(imports)}\nCode:\n{full_code}" return {"text": chunk_text, "metadata": metadata} def create_fallback_chunk(filepath: Path) -> tuple: """Create a minimal fallback chunk with proper metadata.""" try: full_code = filepath.read_text(encoding="utf-8") except Exception: full_code = "" fallback_meta = { "file": str(filepath), "name": filepath.stem, # Use stem instead of full name "type": "file", "line": 1, "language": "rust", } chunk_text = f"File: {filepath}\nType: file\nName: {filepath.stem}\nCode:\n{full_code}" return ((chunk_text, tuple(fallback_meta.items())),) @lru_cache(maxsize=1000) def parse_shell_file_cached(filepath_str: str) -> tuple: """Cached shell script parsing. Returns tuple for hashability.""" filepath = Path(filepath_str) try: content = filepath.read_text(encoding="utf-8") chunks = [] lines = content.splitlines(keepends=True) patterns = [ r'^(\w+)\s*\(\s*\)\s*\{', r'^function\s+(\w+)\s*\{', r'^function\s+(\w+)\s*\(\s*\)\s*\{', r'^def\s+(\w+)\s*\(\s*\)\s*\{', ] for i, line in enumerate(lines): line_text = line.strip() for pattern in patterns: match = re.match(pattern, line_text) if match: func_name = match.group(1) start_line = i end_line = find_shell_function_end(lines, i) func_code = "".join(lines[start_line:end_line + 1]) # Extract leading comments leading_comments = extract_shell_comments(lines, start_line) chunk_text = f"File: {filepath}\nType: function\nName: {func_name}\n" if leading_comments: chunk_text += f"Comments:\n{leading_comments}\n\n" chunk_text += f"Code:\n{func_code}" chunks.append({ "text": chunk_text, "metadata": { "file": str(filepath), "name": func_name, "type": "function", "line": start_line + 1, "language": "shell", "has_braces": "{" in line_text } }) break # Move to next line after finding a match # If no functions found, fall back to basic chunking return tuple((c["text"], tuple(c["metadata"].items())) for c in chunks) if chunks else tuple() except Exception as e: logger.warning(f"parse_shell_file failed {filepath}: {e}") return tuple() def parse_shell_file(filepath: Path) -> List[Dict]: """Enhanced shell script parsing with caching.""" cached_result = parse_shell_file_cached(str(filepath)) if not cached_result: return chunk_text_file(filepath, "shell") return [{"text": text, "metadata": dict(meta)} for text, meta in cached_result] def extract_shell_comments(lines: List[str], function_line_index: int, max_lines: int = 10) -> str: """Extract leading comments above a shell function.""" comments = [] idx = function_line_index - 1 while idx >= 0 and len(comments) < max_lines: line = lines[idx].strip() if line.startswith('#'): comments.insert(0, line[1:].strip()) # Remove the # and trim idx -= 1 elif line == '': # Allow one blank line between comments if comments: comments.insert(0, "") idx -= 1 else: break else: break # Clean up: remove leading/trailing empty lines while comments and comments[0] == "": comments.pop(0) while comments and comments[-1] == "": comments.pop(-1) return "\n".join(comments) def find_shell_function_end(lines: List[str], start_line_index: int) -> int: """ Find the end of a shell function by tracking brace nesting. Args: lines: List of file lines start_line_index: Starting line index (0-based) of the function Returns: Line index (0-based) where the function ends """ brace_count = 0 in_function = False for i in range(start_line_index, len(lines)): line = lines[i] # Count opening and closing braces for char in line: if char == '{': brace_count += 1 in_function = True elif char == '}': brace_count -= 1 # If we've returned to brace_count 0 and we were in a function, this is the end if in_function and brace_count == 0: return i # Special case: shell functions without braces (single line) if not in_function and i > start_line_index: # Look for function end patterns stripped = line.strip() if (stripped.startswith(('function ', 'def ')) or re.match(r'^\w+\(\s*\)', stripped) or stripped.endswith(')') and '()' not in line): # New function starting, so previous one ended return i - 1 # If we never found the end, return the last line return len(lines) - 1 # ----------------------------- # Detect language helper # ----------------------------- def detect_language(filepath: Path) -> str: ext = filepath.suffix.lower() lang_map = { ".py": "python", ".go": "go", ".rs": "rust", ".ts": "typescript", ".tsx": "typescript", ".js": "javascript", ".jsx": "javascript", ".svelte": "svelte", ".java": "java", ".sql": "sql", ".pgsql": "sql", ".sh": "shell" } return lang_map.get(ext, "unknown") # ----------------------------- # Indexing codebase # ----------------------------- def index_codebase(codebase_path: Path) -> Tuple[List[str], List[Dict[str, Any]]]: ragignore = load_ragignore(codebase_path) all_texts = [] all_metadata = [] logger.info(f"Indexing codebase at {codebase_path}") for filepath in codebase_path.rglob("*"): if not filepath.is_file(): continue rel = filepath.relative_to(codebase_path) if ragignore.match_file(str(rel)): continue if filepath.suffix.lower() not in SUPPORTED_EXTENSIONS: continue language = detect_language(filepath) if language == "python": chunks = parse_python_file(filepath) elif language == "go": chunks = parse_go_file(filepath) elif language == "java": chunks = parse_java_file(filepath) elif language == "svelte": chunks = parse_svelte_file(filepath) elif language == "sql": chunks = parse_sql_file(filepath) elif language == "shell": chunks = parse_shell_file(filepath)# elif language == "rust": chunks = parse_rust_file(filepath) else: chunks = chunk_text_file(filepath, language) for c in chunks: all_texts.append(c["text"]) all_metadata.append(clean_metadata_for_chroma(c["metadata"])) logger.info(f"Indexed {len(all_texts)} chunks") return all_texts, all_metadata def clean_metadata_for_chroma(metadata: Dict[str, Any]) -> Dict[str, Any]: """Clean metadata to only include str, int, float, bool values (no None).""" cleaned = {} for key, value in metadata.items(): if value is None: cleaned[key] = "" # Convert None to empty string elif isinstance(value, (str, int, float, bool)): cleaned[key] = value elif isinstance(value, list): # Convert lists to comma-separated strings cleaned[key] = ",".join(str(v) for v in value if v is not None) else: # Convert other types to strings cleaned[key] = str(value) return cleaned def build_indexes(): global vectorstore, bm25, bm25_corpus, chunks_metadata, embeddings, index_build_time logger.info("Building indexes with contextual weighting and batch embedding...") start_time = time.time() # ------------------------------------------------- # Step 1: Parse codebase and build enhanced graph # ------------------------------------------------- texts, metadatas = index_codebase(CODEBASE_PATH) if len(texts) != len(metadatas): logger.error( "Indexing mismatch: %d texts vs %d metadata entries. Aborting index build.", len(texts), len(metadatas), ) raise ValueError(f"Indexing produced {len(texts)} texts but {len(metadatas)} metadata entries") logger.info("=== Sample metadata inspection ===") for i, m in enumerate(metadatas[:5]): logger.info(f"Chunk {i}: type={m.get('type')}, name={m.get('name')}") logger.info(f" imports={m.get('imports')}") logger.info(f" calls={m.get('calls')}") logger.info(f" parameters={m.get('parameters')}") # Initialize graph graph = LocalGraph(root_path=str(CODEBASE_PATH)) graph.clear() # Track all created nodes for second-pass relationships created_nodes = {} # Track language-specific features for cross-language relationships component_registry = {} interface_registry = {} trait_registry = {} class_registry = {} package_registry = {} function_registry = {} logger.info("=== Phase 1: Creating nodes ===") # ------------------------------------------------- # PHASE 1: Create all nodes with rich metadata # ------------------------------------------------- for m in metadatas: fpath = m.get("file") lang = m.get("language", "unknown") node_type = m.get("type", "unknown") name = m.get("name", "") # Skip empty/invalid nodes if not name or node_type == "unknown" or not fpath: continue # Normalize metadata if "implements_traits" in m and m["implements_traits"]: m["implements"] = m["implements_traits"] # Create file node file_node_id = f"file::{fpath}" if file_node_id not in graph.graph.nodes: graph.add_node( file_node_id, type="File", name=fpath, path=fpath, lang=lang, ) created_nodes[file_node_id] = { "type": "File", "name": fpath, "path": fpath, "lang": lang, } # Create node ID with language prefix node_id = f"{lang}::{node_type}::{fpath}::{name}" # Base attributes for all nodes base_attrs = { "name": name, "file": fpath, "line": m.get("line"), "lang": lang } # Add docstring/documentation if available if m.get("docstring"): base_attrs["docstring"] = m.get("docstring") # Store for second-pass processing created_nodes[node_id] = base_attrs.copy() created_nodes[node_id]["type"] = node_type created_nodes[node_id]["metadata"] = m # ===== FUNCTIONS & METHODS ===== if node_type in ["function", "method", "async_function", "constructor"]: if node_type == "method" and "class" in m: class_name = m['class'] node_id = f"{lang}::{node_type}::{fpath}::{class_name}.{name}" created_nodes[node_id] = base_attrs.copy() created_nodes[node_id]["type"] = node_type created_nodes[node_id]["class_name"] = class_name created_nodes[node_id]["metadata"] = m graph.add_node( node_id, **base_attrs, type=node_type, class_name=class_name, return_type=m.get("return_type"), parameters=m.get("parameters", []), visibility=m.get("visibility"), is_static=m.get("is_static", False), is_abstract=m.get("is_abstract", False), is_async=m.get("is_async", False), is_generator=m.get("is_generator", False), decorators=m.get("decorators", []), annotations=m.get("annotations", []) ) class_node_id = f"{lang}::class::{fpath}::{class_name}" graph.add_edge(class_node_id, node_id, "contains") # Register in function_registry for cross-file calls function_registry[name] = { "node_id": node_id, "lang": lang, "file": fpath, "type": node_type } else: graph.add_node( node_id, **base_attrs, type=node_type, return_type=m.get("return_type"), parameters=m.get("parameters", []), visibility=m.get("visibility"), is_async=m.get("is_async", False), is_generator=m.get("is_generator", False), is_exported=m.get("is_exported", False), is_property=m.get("is_property", False), decorators=m.get("decorators", []), annotations=m.get("annotations", []), generics=m.get("generics", []) ) graph.add_edge(file_node_id, node_id, "contains") # ===== CLASSES ===== elif node_type == "class": graph.add_node( node_id, **base_attrs, type=node_type, extends=m.get("extends"), implements=m.get("implements", []), methods=m.get("methods", []), properties=m.get("properties", []), fields=m.get("fields", []), static_methods=m.get("static_methods", []), class_methods=m.get("class_methods", []), class_variables=m.get("class_variables", []), visibility=m.get("visibility"), is_abstract=m.get("is_abstract", False), is_final=m.get("is_final", False), decorators=m.get("decorators", []), annotations=m.get("annotations", []), generics=m.get("generics", []), metaclass=m.get("metaclass") ) graph.add_edge(file_node_id, node_id, "contains") class_registry[name] = { "node_id": node_id, "lang": lang, "file": fpath, "metadata": m } # ===== STRUCTS (Rust/Go) ===== elif node_type == "struct": graph.add_node( node_id, **base_attrs, type=node_type, fields=m.get("fields", []), visibility=m.get("visibility"), generics=m.get("generics", []), derives=m.get("derives", []) ) graph.add_edge(file_node_id, node_id, "contains") class_registry[name] = { "node_id": node_id, "lang": lang, "file": fpath, "metadata": m } # ===== INTERFACES (TypeScript/Java/Go) ===== elif node_type == "interface": graph.add_node( node_id, **base_attrs, type=node_type, extends=m.get("extends", []), methods=m.get("methods", []), properties=m.get("properties", []), interface_methods=m.get("interface_methods", []), visibility=m.get("visibility"), generics=m.get("generics", []) ) graph.add_edge(file_node_id, node_id, "contains") interface_registry[name] = { "node_id": node_id, "lang": lang, "file": fpath, "metadata": m } # ===== TRAITS (Rust) ===== elif node_type == "trait": graph.add_node( node_id, **base_attrs, type=node_type, methods=m.get("methods", []), visibility=m.get("visibility"), generics=m.get("generics", []) ) graph.add_edge(file_node_id, node_id, "contains") trait_registry[name] = { "node_id": node_id, "lang": lang, "file": fpath, "metadata": m } # ===== RUST IMPL BLOCKS ===== elif node_type == "impl": graph.add_node( node_id, **base_attrs, type=node_type, traits=m.get("traits", []), methods=m.get("methods", []), generics=m.get("generics", []), is_unsafe=m.get("is_unsafe", False) ) graph.add_edge(file_node_id, node_id, "contains") # ===== ENUMS ===== elif node_type == "enum": graph.add_node( node_id, **base_attrs, type=node_type, variants=m.get("variants", []), members=m.get("members", []), fields=m.get("fields", []), visibility=m.get("visibility") ) graph.add_edge(file_node_id, node_id, "contains") # ===== TYPE ALIASES ===== elif node_type == "type": graph.add_node( node_id, **base_attrs, type=node_type, type_definition=m.get("type_definition"), generics=m.get("generics", []) ) graph.add_edge(file_node_id, node_id, "contains") # ===== MODULES (Rust/Go/Python packages) ===== elif node_type == "module": graph.add_node( node_id, **base_attrs, type=node_type, visibility=m.get("visibility") ) graph.add_edge(file_node_id, node_id, "contains") # Create symbolic module node for cross-references module_symbol_id = f"module_symbol::{lang}::{name}" if module_symbol_id not in graph.graph.nodes: graph.add_node( module_symbol_id, type=name, name=name, file=fpath, lang=lang ) created_nodes[module_symbol_id] = { "type": name, "name": name, "file": fpath, "lang": lang } graph.add_edge(node_id, module_symbol_id, "defines") package_registry[name] = { "node_id": node_id, "symbol_id": module_symbol_id, "lang": lang, "file": fpath } # ===== SVELTE COMPONENTS ===== elif node_type == "markup" and lang == "svelte": component_name = m.get("component_name", name) component_node_id = f"svelte::component::{fpath}::{component_name}" graph.add_node( component_node_id, **base_attrs, type="component", component_name=component_name, props=m.get("component_props", []), slots=m.get("slots", []), has_default_slot=m.get("has_default_slot", False), events=m.get("event_handlers", []), used_components=m.get("used_components", []), stores=m.get("stores", []), template_stores=m.get("template_stores", []), use_directives=m.get("use_directives", []), transitions=m.get("transitions", []), bind_directives=m.get("bind_directives", []) ) graph.add_edge(file_node_id, component_node_id, "contains") component_registry[component_name] = { "node_id": component_node_id, "file": fpath, "metadata": m } node_id = component_node_id # ===== SVELTE SCRIPTS ===== elif node_type in ["script", "script_module"] and lang in ["typescript", "javascript"]: pass # ===== VARIABLES & CONSTANTS ===== elif node_type in ["variable", "constant", "prop", "reactive", "var", "const"]: graph.add_node( node_id, **base_attrs, type=node_type, var_type=m.get("var_type"), var_kind=m.get("var_kind"), is_exported=m.get("is_exported", False), is_reactive=m.get("is_reactive", False), is_store=m.get("is_store", False), value=m.get("value") ) graph.add_edge(file_node_id, node_id, "contains") # ===== STYLES (CSS/SCSS/LESS) ===== elif node_type == "style": graph.add_node( node_id, **base_attrs, type=node_type, css_classes=m.get("css_classes", []), css_ids=m.get("css_ids", []), css_variables=m.get("css_variables", []), css_selectors=m.get("css_selectors", []), is_scoped=m.get("is_scoped", True), component_name=m.get("component_name") ) graph.add_edge(file_node_id, node_id, "contains") # ===== IMPORTS ===== elif node_type == "import": import_path = m.get("import_path") or m.get("full_name") or m.get("parameters", [name])[0] if m.get("parameters") else name if isinstance(import_path, list): import_path = import_path[0] if import_path else name import_node_id = f"{lang}::import::{fpath}::{import_path}" graph.add_node( import_node_id, **base_attrs, type=node_type, import_path=import_path, imported_names=m.get("imported_names", []) ) graph.add_edge(file_node_id, import_node_id, "contains") # Create symbolic nodes - DON'T pass created_nodes, collect new nodes separately new_nodes = _create_import_symbolic_nodes(graph, import_node_id, file_node_id, import_path, lang) # Add new nodes after iteration completes for new_node_id, new_node_data in new_nodes.items(): if new_node_id not in created_nodes: created_nodes[new_node_id] = new_node_data node_id = import_node_id # ===== FIELDS (Java standalone fields) ===== elif node_type == "field": graph.add_node( node_id, **base_attrs, type=node_type, field_type=m.get("field_type"), visibility=m.get("visibility"), is_static=m.get("is_static", False), is_final=m.get("is_final", False), is_volatile=m.get("is_volatile", False), annotations=m.get("annotations", []) ) graph.add_edge(file_node_id, node_id, "contains") # ===== ANNOTATIONS (Java) ===== elif node_type == "annotation": graph.add_node( node_id, **base_attrs, type=node_type, visibility=m.get("visibility") ) graph.add_edge(file_node_id, node_id, "contains") # ===== GENERIC FALLBACK ===== else: graph.add_node(node_id, **base_attrs, type=node_type) graph.add_edge(file_node_id, node_id, "contains") logger.info(f"Phase 1 complete: Created {len(created_nodes)} nodes") logger.info(f" Classes: {len(class_registry)}, Interfaces: {len(interface_registry)}") logger.info(f" Traits: {len(trait_registry)}, Components: {len(component_registry)}") logger.info(f" Packages/Modules: {len(package_registry)}") # ------------------------------------------------- # PHASE 2: Build relationships # ------------------------------------------------- logger.info("=== Phase 2: Building relationships ===") # CRITICAL FIX: Create a snapshot of the items to iterate over # This prevents "dictionary changed size during iteration" errors nodes_snapshot = list(created_nodes.items()) for node_id, node_data in nodes_snapshot: m = node_data.get("metadata", {}) if not m: continue node_type = node_data.get("type") fpath = node_data.get("file") lang = node_data.get("lang") name = node_data.get("name") # ===== IMPORTS & DEPENDENCIES ===== imports_raw = m.get("imports", []) if node_type == "import" and "parameters" in m: imports_raw = m["parameters"] if isinstance(imports_raw, str): imports = [imp.strip() for imp in imports_raw.split(",") if imp.strip()] else: imports = imports_raw if isinstance(imports_raw, list) else [] if imports: file_node_id = f"file::{fpath}" # Collect new nodes instead of modifying created_nodes during iteration new_nodes = _add_import_edges_enhanced(graph, file_node_id, node_id, imports, lang) for new_node_id, new_node_data in new_nodes.items(): if new_node_id not in created_nodes: created_nodes[new_node_id] = new_node_data # ===== FUNCTION CALLS ===== calls_raw = m.get("calls", []) if isinstance(calls_raw, str): calls = [c.strip() for c in calls_raw.split(",") if c.strip()] else: calls = calls_raw if isinstance(calls_raw, list) else [] if calls: # Use the enhanced _add_call_edges that handles cross-file graph._add_call_edges(node_id, calls, fpath) # ALSO create direct edges to functions we can find in registries for cal in calls: # Search function_registry if we had one, but for now search created_nodes for candidate_id, candidate_data in created_nodes.items(): if (candidate_data.get('type') in ['function', 'method'] and candidate_data.get('name') == cal and candidate_data.get('file') != fpath): # Cross-file only graph.add_edge(node_id, candidate_id, "calls") # ===== CLASS/STRUCT INHERITANCE ===== if node_type in ["class", "struct"]: extends = m.get("extends") if extends: if isinstance(extends, list): extends_list = extends else: extends_list = [extends] for parent_name in extends_list: for class_name, class_info in class_registry.items(): if class_name == parent_name: graph.add_edge(node_id, class_info["node_id"], "extends") implements = m.get("implements", []) if isinstance(implements, str): implements = [implements] for interface_name in implements: for iface_name, iface_info in interface_registry.items(): if iface_name == interface_name: graph.add_edge(node_id, iface_info["node_id"], "implements") for trait_name, trait_info in trait_registry.items(): if trait_name == interface_name: graph.add_edge(node_id, trait_info["node_id"], "implements") # ===== INTERFACE INHERITANCE ===== if node_type == "interface": extends_list = m.get("extends", []) if isinstance(extends_list, str): extends_list = [extends_list] for parent_interface in extends_list: for iface_name, iface_info in interface_registry.items(): if iface_name == parent_interface: graph.add_edge(node_id, iface_info["node_id"], "extends") # ===== RUST IMPL BLOCKS ===== if node_type == "impl": traits = m.get("traits", []) if isinstance(traits, str): traits = [traits] for trait_name in traits: for t_name, t_info in trait_registry.items(): if t_name == trait_name: graph.add_edge(node_id, t_info["node_id"], "implements") impl_name = m.get("name", "") for class_name, class_info in class_registry.items(): if class_name == impl_name or class_name in impl_name: graph.add_edge(node_id, class_info["node_id"], "implements_for") # ===== GO METHOD RECEIVERS ===== if lang == "go" and node_type in ["method", "function"]: receiver = m.get("receiver") receiver_base = m.get("receiver_base_type") if receiver_base: for struct_name, struct_info in class_registry.items(): if struct_name == receiver_base and struct_info["lang"] == "go": graph.add_edge(struct_info["node_id"], node_id, "has_method") # ===== METHOD RETURN TYPES & PARAMETERS ===== if node_type in ["function", "method", "async_function", "constructor"]: return_type = m.get("return_type") if return_type and return_type not in ['void', '()', 'Result', 'None', '', 'any', 'unknown']: clean_return = _clean_type_string(return_type) for type_name, type_info in {**class_registry, **interface_registry, **trait_registry}.items(): if type_name == clean_return: graph.add_edge(node_id, type_info["node_id"], "returns") parameters = m.get("parameters", []) if isinstance(parameters, list): for param in parameters: param_type = _extract_param_type(param) if param_type: for type_name, type_info in {**class_registry, **interface_registry, **trait_registry}.items(): if type_name == param_type: graph.add_edge(node_id, type_info["node_id"], "uses_parameter") # ===== SVELTE COMPONENT RELATIONSHIPS ===== if node_type == "component" or (node_type == "markup" and lang == "svelte"): used_components = m.get("used_components", []) for comp_name in used_components: if comp_name in component_registry: comp_info = component_registry[comp_name] graph.add_edge(node_id, comp_info["node_id"], "uses_component") props = m.get("component_props", []) for prop_name in props: for iface_name, iface_info in interface_registry.items(): if f"{m.get('component_name')}Props" == iface_name: graph.add_edge(node_id, iface_info["node_id"], "uses_props_type") # ===== FIELD TYPE RELATIONSHIPS ===== if node_type in ["struct", "class"]: fields = m.get("fields", []) for field in fields: field_type = _extract_field_type(field) if field_type: for type_name, type_info in {**class_registry, **interface_registry, **trait_registry}.items(): if type_name == field_type: graph.add_edge(node_id, type_info["node_id"], "has_field_type") logger.info("Phase 2 complete: Built intra-language relationships") # ------------------------------------------------- # PHASE 3: Advanced relationship building # ------------------------------------------------- logger.info("=== Phase 3: Advanced relationships ===") type_connections = build_type_relationships(graph, created_nodes) logger.info(f"Added {type_connections} type relationships") bevy_connections = build_bevy_relationships(graph, created_nodes) logger.info(f"Added {bevy_connections} Bevy-specific relationships") cross_lang_connections = build_cross_language_relationships( graph, created_nodes, component_registry, interface_registry, class_registry ) logger.info(f"Added {cross_lang_connections} cross-language relationships") graph.save() graph.to_json() graph.to_toon() logger.info(f"βœ… Graph built with {graph.graph.number_of_nodes()} nodes and {graph.graph.number_of_edges()} edges.") # ------------------------------------------------- # PHASE 4: Embedding/index pipeline # ------------------------------------------------- logger.info("=== Phase 4: Building embeddings ===") os.environ["OLLAMA_HOST"] = OLLAMA_BASE_URL embeddings = OllamaEmbeddings(model=EMBEDDING_MODEL, base_url=OLLAMA_BASE_URL) logger.info(f"Generating embeddings for {len(texts)} documents in batches of {EMBEDDING_BATCH_SIZE}...") base_embeddings = batch_embed_documents(texts) logger.info("Applying contextual weights...") weighted_embeddings = apply_contextual_weights_to_embeddings(base_embeddings, metadatas) logger.info("Creating vector store (Chroma) with weighted embeddings...") class WeightedEmbeddingFunction: def __init__(self, texts, weighted_embs): self.lookup = defaultdict(deque) for t, emb in zip(texts, weighted_embs): self.lookup[t].append(emb) def embed_documents(self, docs): result = [] for d in docs: if not self.lookup[d]: raise ValueError(f"No embedding left for text {repr(d[:80])}") result.append(self.lookup[d].popleft()) return result def embed_query(self, q): return embeddings.embed_query(q) weighted_emb_func = WeightedEmbeddingFunction(texts=texts, weighted_embs=weighted_embeddings) vectorstore = Chroma.from_texts( texts=texts, embedding=weighted_emb_func, persist_directory=str(VECTOR_DB_PATH), metadatas=metadatas ) logger.info("Building BM25 index...") tokenized = [t.lower().split() for t in texts] bm25 = BM25Okapi(tokenized) bm25_corpus = texts chunks_metadata = metadatas BM25_INDEX_PATH.write_text( json.dumps({"corpus": texts, "metadata": metadatas}, indent=2), encoding="utf-8" ) index_build_time = time.time() elapsed = index_build_time - start_time logger.info(f"βœ… Index build complete. Total time: {elapsed:.1f}s") # ------------------------------------------------- # Helper Functions # ------------------------------------------------- def _add_import_edges_enhanced(graph, file_node_id, importing_node_id, imports, lang): """Enhanced import edge creation that resolves Rust paths to actual files.""" new_nodes = {} for imp_path in imports: if not imp_path or not isinstance(imp_path, str): continue # Skip standard library and external crates for now if any(imp_path.startswith(prefix) for prefix in ['std::', 'core::', 'alloc::', 'bevy::', 'serde::']): continue # Try to resolve relative paths within the codebase resolved_path = resolve_rust_import(imp_path, graph.root_path) if resolved_path: # Create edge to the actual file target_file_node_id = f"file::{resolved_path}" if not graph.graph.has_edge(importing_node_id, target_file_node_id): graph.add_edge(importing_node_id, target_file_node_id, "imports") else: # Create symbolic node for external/unresolved imports symbolic_id = f"rust::external::{imp_path}" if symbolic_id not in graph.graph.nodes: graph.add_node( symbolic_id, type="external", name=imp_path, file="", lang="rust", is_external=True ) new_nodes[symbolic_id] = { "type": "external", "name": imp_path, "file": "", "lang": "rust" } if not graph.graph.has_edge(importing_node_id, symbolic_id): graph.add_edge(importing_node_id, symbolic_id, "imports") return new_nodes def resolve_rust_import(import_path: str, root_path: str) -> Optional[str]: """Try to resolve a Rust import path to an actual file path.""" # Remove any trailing ::* or other modifiers clean_path = import_path.replace('::*', '').replace('*', '') # Convert :: to path separators potential_path = clean_path.replace('::', '/') # Common Rust file extensions to try extensions = ['.rs', ''] # Common source directories in Rust projects base_dirs = ['src', 'crates', 'workspace'] for base_dir in base_dirs: for ext in extensions: # Try as direct file candidate = Path(root_path) / base_dir / f"{potential_path}{ext}" if candidate.exists(): return str(candidate) # Try with mod.rs in directory candidate = Path(root_path) / base_dir / potential_path / f"mod{ext}" if candidate.exists(): return str(candidate) # Try with lib.rs in directory candidate = Path(root_path) / base_dir / potential_path / f"lib{ext}" if candidate.exists(): return str(candidate) return None def _clean_type_string(type_str: str) -> str: """Clean up type strings for matching across all languages.""" if not type_str: return "" type_str = type_str.strip() # Remove common Rust wrappers type_str = re.sub(r'Option<(.+?)>', r'\1', type_str) type_str = re.sub(r'Result<(.+?)(?:,|>).*', r'\1', type_str) type_str = re.sub(r'Vec<(.+?)>', r'\1', type_str) type_str = re.sub(r'Box<(.+?)>', r'\1', type_str) type_str = re.sub(r'Arc<(.+?)>', r'\1', type_str) type_str = re.sub(r'Rc<(.+?)>', r'\1', type_str) # Remove TypeScript/JavaScript wrappers type_str = re.sub(r'Promise<(.+?)>', r'\1', type_str) type_str = re.sub(r'Array<(.+?)>', r'\1', type_str) type_str = re.sub(r'Observable<(.+?)>', r'\1', type_str) # Remove Java wrappers type_str = re.sub(r'Optional<(.+?)>', r'\1', type_str) type_str = re.sub(r'List<(.+?)>', r'\1', type_str) type_str = re.sub(r'Set<(.+?)>', r'\1', type_str) type_str = re.sub(r'Map<.+?,\s*(.+?)>', r'\1', type_str) # Remove Go slices and pointers type_str = re.sub(r'\[\](.+)', r'\1', type_str) type_str = re.sub(r'\*(.+)', r'\1', type_str) # Remove Rust references and mutability type_str = type_str.replace('&', '').replace('mut ', '').strip() # Remove generic parameters for simpler matching type_str = re.sub(r'<.*?>', '', type_str).strip() # Remove array brackets type_str = re.sub(r'\[.*?\]', '', type_str).strip() return type_str def _extract_param_type(param: str) -> str: """Extract type from parameter string for all languages.""" if not param or not isinstance(param, str): return "" if ':' in param: parts = param.split(':') if len(parts) < 2: return "" param_type = parts[-1].strip() elif ' ' in param: parts = param.strip().split() if parts[0] and parts[0][0].isupper() or parts[0] in ['int', 'string', 'bool', 'float', 'double', 'long']: param_type = parts[0] else: param_type = parts[-1] if parts else "" else: return "" param_type = param_type.split('=')[0].strip() param_type = param_type.split('`')[0].strip() return _clean_type_string(param_type) def _extract_field_type(field: str) -> str: """Extract type from field string for all languages.""" if not field or not isinstance(field, str): return "" if field.startswith('embedded:'): return _clean_type_string(field[9:]) if ':' in field: parts = field.split(':') if len(parts) < 2: return "" field_type = parts[-1].strip() elif ' ' in field: parts = field.strip().split() if len(parts) < 2: return "" if parts[0] and parts[0][0].isupper() or parts[0] in ['int', 'string', 'bool', 'float', 'double', 'long']: field_type = parts[0] else: field_type = parts[1] if len(parts) > 1 else parts[0] else: return "" field_type = re.sub(r'`.*?`', '', field_type).strip() field_type = re.sub(r'".*?"', '', field_type).strip() return _clean_type_string(field_type) def build_cross_language_relationships(graph, created_nodes, component_registry, interface_registry, class_registry): """Build relationships across different programming languages.""" connections = 0 for ts_name, ts_info in interface_registry.items(): if ts_info["lang"] in ["typescript", "javascript"]: for rust_name, rust_info in class_registry.items(): if rust_info["lang"] == "rust" and ts_name == rust_name: if not graph.graph.has_edge(ts_info["node_id"], rust_info["node_id"]): graph.add_edge(ts_info["node_id"], rust_info["node_id"], "mirrors") connections += 1 for java_name, java_info in class_registry.items(): if java_info["lang"] == "java": for go_name, go_info in class_registry.items(): if go_info["lang"] == "go" and java_name == go_name: if not graph.graph.has_edge(java_info["node_id"], go_info["node_id"]): graph.add_edge(java_info["node_id"], go_info["node_id"], "mirrors") connections += 1 for comp_name, comp_info in component_registry.items(): props_name = f"{comp_name}Props" for iface_name, iface_info in interface_registry.items(): if iface_name == props_name or iface_name == comp_name: if not graph.graph.has_edge(comp_info["node_id"], iface_info["node_id"]): graph.add_edge(comp_info["node_id"], iface_info["node_id"], "uses_type") connections += 1 for class_name, class_info in class_registry.items(): if class_name == comp_name or class_name == props_name: if not graph.graph.has_edge(comp_info["node_id"], class_info["node_id"]): graph.add_edge(comp_info["node_id"], class_info["node_id"], "uses_type") connections += 1 for py_name, py_info in class_registry.items(): if py_info["lang"] == "python": for ts_name, ts_info in interface_registry.items(): if ts_info["lang"] == "typescript" and py_name == ts_name: if not graph.graph.has_edge(py_info["node_id"], ts_info["node_id"]): graph.add_edge(py_info["node_id"], ts_info["node_id"], "api_type") connections += 1 return connections def build_type_relationships(graph, created_nodes): """Build enhanced type relationships within language boundaries.""" connections = 0 type_map = {} for node_id, node_data in created_nodes.items(): node_type = node_data.get("type") name = node_data.get("name") lang = node_data.get("lang") if node_type in ["class", "struct", "interface", "trait", "enum", "type"] and name: key = f"{lang}::{name}" if key not in type_map: type_map[key] = [] type_map[key].append(node_id) for node_id, node_data in created_nodes.items(): m = node_data.get("metadata", {}) if not m: continue lang = node_data.get("lang") node_type = node_data.get("type") # Field types fields = m.get("fields", []) if isinstance(fields, list): for field in fields: field_type = _extract_field_type(field) if field_type: type_key = f"{lang}::{field_type}" if type_key in type_map: for target_id in type_map[type_key]: if not graph.graph.has_edge(node_id, target_id): graph.add_edge(node_id, target_id, "has_field_of_type") connections += 1 # Return types return_type = m.get("return_type") if return_type: clean_return = _clean_type_string(return_type) if clean_return: type_key = f"{lang}::{clean_return}" if type_key in type_map: for target_id in type_map[type_key]: if not graph.graph.has_edge(node_id, target_id): graph.add_edge(node_id, target_id, "returns_type") connections += 1 # Parameter types parameters = m.get("parameters", []) if isinstance(parameters, list): for param in parameters: param_type = _extract_param_type(param) if param_type: type_key = f"{lang}::{param_type}" if type_key in type_map: for target_id in type_map[type_key]: if not graph.graph.has_edge(node_id, target_id): graph.add_edge(node_id, target_id, "accepts_type") connections += 1 # Generic type parameters generics = m.get("generics", []) if isinstance(generics, list): for generic in generics: generic_parts = generic.split(':') if len(generic_parts) > 1: constraint = generic_parts[1].strip() type_key = f"{lang}::{constraint}" if type_key in type_map: for target_id in type_map[type_key]: if not graph.graph.has_edge(node_id, target_id): graph.add_edge(node_id, target_id, "constrained_by") connections += 1 return connections def build_bevy_relationships(graph, created_nodes): """Build Bevy-specific relationships (ECS patterns).""" connections = 0 systems = [] components = [] resources = [] bundles = [] for node_id, node_data in created_nodes.items(): m = node_data.get("metadata", {}) if not m: continue lang = node_data.get("lang") if lang != "rust": continue node_type = node_data.get("type") name = node_data.get("name") if node_type in ["function", "method"]: params = m.get("parameters", []) has_query = any("Query" in str(p) for p in params) has_commands = any("Commands" in str(p) for p in params) has_res = any("Res" in str(p) or "ResMut" in str(p) for p in params) if has_query or has_commands or has_res: systems.append(node_id) if node_type == "struct": derives = m.get("derives", []) if "Component" in str(derives): components.append(node_id) if "Resource" in str(derives): resources.append(node_id) if "Bundle" in str(derives): bundles.append(node_id) for system_id in systems: system_data = created_nodes.get(system_id, {}) m = system_data.get("metadata", {}) params = m.get("parameters", []) for param in params: param_str = str(param) query_match = re.findall(r'Query<[^>]*?([A-Z]\w+)', param_str) for comp_name in query_match: for comp_id in components: comp_data = created_nodes.get(comp_id, {}) if comp_data.get("name") == comp_name: if not graph.graph.has_edge(system_id, comp_id): graph.add_edge(system_id, comp_id, "queries_component") connections += 1 res_match = re.findall(r'Res(?:Mut)?<([A-Z]\w+)', param_str) for res_name in res_match: for res_id in resources: res_data = created_nodes.get(res_id, {}) if res_data.get("name") == res_name: if not graph.graph.has_edge(system_id, res_id): graph.add_edge(system_id, res_id, "uses_resource") connections += 1 for bundle_id in bundles: bundle_data = created_nodes.get(bundle_id, {}) m = bundle_data.get("metadata", {}) fields = m.get("fields", []) for field in fields: field_type = _extract_field_type(field) for comp_id in components: comp_data = created_nodes.get(comp_id, {}) if comp_data.get("name") == field_type: if not graph.graph.has_edge(bundle_id, comp_id): graph.add_edge(bundle_id, comp_id, "bundles_component") connections += 1 return connections def _create_import_symbolic_nodes(graph, import_node_id, file_node_id, import_path, lang): """Create rich symbolic nodes for imports. Returns dict of new nodes to add.""" new_nodes = {} if not import_path: return new_nodes # Normalize path separators based on language if lang == "python": parts = import_path.replace('.', '::').split('::') elif lang in ["javascript", "typescript"]: if import_path.startswith('.'): return new_nodes parts = import_path.replace('/', '::').replace('@', '').split('::') elif lang == "rust": parts = import_path.split('::') elif lang == "go": parts = import_path.split('/') elif lang == "java": parts = import_path.split('.') else: parts = import_path.replace('.', '::').replace('/', '::').split('::') parts = [p.strip() for p in parts if p.strip()] if not parts: return new_nodes # Create package/crate node package_name = parts[0] package_node_id = f"package::{lang}::{package_name}" if package_node_id not in graph.graph.nodes: graph.add_node( package_node_id, type=package_name, name=package_name, file=import_path, lang=lang, is_external=True ) new_nodes[package_node_id] = { "type": package_name, "name": package_name, "file": import_path, "lang": lang } if not graph.graph.has_edge(file_node_id, package_node_id): graph.add_edge(file_node_id, package_node_id, "imports") # Create intermediate module/namespace nodes current_path = package_name parent_node = package_node_id for part in parts[1:]: current_path = f"{current_path}::{part}" module_node_id = f"symbol::{lang}::{current_path}" if module_node_id not in graph.graph.nodes: graph.add_node( module_node_id, type=part, name=current_path, file=import_path, lang=lang, is_symbol=True ) new_nodes[module_node_id] = { "type": part, "name": current_path, "file": import_path, "lang": lang } if not graph.graph.has_edge(parent_node, module_node_id): graph.add_edge(parent_node, module_node_id, "contains") parent_node = module_node_id if parent_node != package_node_id and not graph.graph.has_edge(import_node_id, parent_node): graph.add_edge(import_node_id, parent_node, "imports") return new_nodes def _extract_parameter_type(param): """Extract type from parameter string, handling multiple formats""" if not param: return None # Handle "type: name" format (TypeScript/Python) if ':' in param: parts = param.split(':') if len(parts) >= 2: return parts[1].strip() # Handle "Type name" format (Rust/Java/C++) if ' ' in param: parts = param.split() if len(parts) >= 2: return parts[0].strip() # Handle complex types with generics if '<' in param and '>' in param: # Extract the base type before generics base_type = param.split('<')[0].strip() return base_type return param.strip() def _clean_type_name(type_name): """Clean type names by removing common modifiers""" if not type_name: return "" # Remove common modifiers modifiers = ['&', 'mut ', 'const ', 'static ', 'pub ', 'private ', 'protected '] cleaned = type_name for mod in modifiers: cleaned = cleaned.replace(mod, '') # Remove trailing references/pointers cleaned = cleaned.rstrip('*&') return cleaned.strip() def _is_meaningful_type(type_name): """Check if a type name is meaningful (not primitive, void, etc.)""" if not type_name: return False meaningless_types = [ 'void', 'int', 'i32', 'i64', 'f32', 'f64', 'bool', 'str', 'String', 'char', 'u8', 'u16', 'u32', 'u64', 'usize', 'isize', 'self', '&self', '&mut self', 'mut self', 'Self', 'Result', 'Option', 'Vec', 'String' ] cleaned = _clean_type_name(type_name) return (cleaned and cleaned not in meaningless_types and not cleaned.startswith('&') and len(cleaned) > 1) # ----------------------------- # Build / load indexes with contextual weighting and batch embedding # ----------------------------- def batch_embed_documents(texts: List[str]) -> List[List[float]]: """Embed documents in batches for better performance.""" all_embeddings = [] total_batches = (len(texts) + EMBEDDING_BATCH_SIZE - 1) // EMBEDDING_BATCH_SIZE for i in range(0, len(texts), EMBEDDING_BATCH_SIZE): batch = texts[i:i + EMBEDDING_BATCH_SIZE] batch_num = i // EMBEDDING_BATCH_SIZE + 1 logger.info(f"Embedding batch {batch_num}/{total_batches} ({len(batch)} docs)...") batch_embeddings = embeddings.embed_documents(batch) all_embeddings.extend(batch_embeddings) return all_embeddings def load_indexes(): global vectorstore, bm25, bm25_corpus, chunks_metadata, embeddings, index_build_time logger.info("Loading indexes from disk...") embeddings = OllamaEmbeddings(model=EMBEDDING_MODEL, base_url=OLLAMA_BASE_URL) vectorstore = Chroma(persist_directory=str(VECTOR_DB_PATH), embedding_function=embeddings) data = json.loads(BM25_INDEX_PATH.read_text(encoding="utf-8")) bm25_corpus = data["corpus"] chunks_metadata = data["metadata"] bm25 = BM25Okapi([t.lower().split() for t in bm25_corpus]) index_build_time = BM25_INDEX_PATH.stat().st_mtime logger.info("Indexes loaded.") # ----------------------------- # Analytics Helper Functions # ----------------------------- def count_by_language(metadata_list: List[Dict[str, Any]]) -> Dict[str, int]: """Count chunks by programming language.""" counts = {} for meta in metadata_list: lang = meta.get("language", "unknown") counts[lang] = counts.get(lang, 0) + 1 return counts def count_by_type(metadata_list: List[Dict[str, Any]]) -> Dict[str, int]: """Count chunks by type (function, class, etc.).""" counts = {} for meta in metadata_list: typ = meta.get("type", "unknown") counts[typ] = counts.get(typ, 0) + 1 return counts def calculate_avg_chunk_size(corpus: List[str]) -> float: """Calculate average chunk size in characters.""" if not corpus: return 0.0 return sum(len(text) for text in corpus) / len(corpus) # ----------------------------- # Hybrid search 80/20 + normalization # ----------------------------- def hybrid_search(query: str, k: int = 30) -> List[Tuple[str, Dict[str, Any], float]]: with safe_search(): # Vector results (Chroma returns (doc, score) where lower score might be better depending on backend) vector_results = vectorstore.similarity_search_with_score(query, k=50) # Normalize vector scores so that higher is better in [0,1] vec_values = [s for _, s in vector_results] if vector_results else [1.0] vmin, vmax = (min(vec_values), max(vec_values)) if vector_results else (0.0, 1.0) vector_scores = {} for doc, s in vector_results: # If Chroma returns distance-like (smaller better), invert; otherwise adapt # We'll normalize via (vmax - s) / (vmax - vmin + eps) eps = 1e-12 denom = (vmax - vmin) + eps norm = (vmax - s) / denom vector_scores[doc.page_content] = norm # BM25 tokenized_query = query.lower().split() bm25_scores = bm25.get_scores(tokenized_query) top_idxs = sorted(range(len(bm25_scores)), key=lambda i: bm25_scores[i], reverse=True)[:50] max_b = max((bm25_scores[i] for i in top_idxs), default=1.0) bm25_doc_scores = {bm25_corpus[i]: (bm25_scores[i] / max_b if max_b > 0 else 0.0) for i in top_idxs} # Fusion 80/20 all_docs = set(vector_scores.keys()) | set(bm25_doc_scores.keys()) fused = {} for d in all_docs: v = vector_scores.get(d, 0.0) b = bm25_doc_scores.get(d, 0.0) fused[d] = VECTOR_WEIGHT * v + BM25_WEIGHT * b # Sort by fused score descending sorted_items = sorted(fused.items(), key=lambda x: x[1], reverse=True) # Build accurate mapping from document text to its true metadata text_to_meta = {text: meta for text, meta in zip(bm25_corpus, chunks_metadata)} results = [] for doc_text, score in sorted_items[:k]: meta = text_to_meta.get(doc_text, {}) results.append((doc_text, meta, score)) return results # ----------------------------- # Enhanced Reranker with Batch Processing # ----------------------------- def rerank_with_ollama_enhanced(query: str, candidates: List[str], top_k: int = 5) -> List[Tuple[str, float]]: """Enhanced Qwen3 Reranker with batch processing and better scoring.""" if not candidates: return [] # Use domain-specific instruction for code search instruction = "Given a technical code search query, retrieve relevant code implementations, function definitions, or examples that directly address the query requirements" batch_scores = [] # Process in smaller batches to avoid timeouts batch_size = 5 for i in range(0, len(candidates), batch_size): batch = candidates[i:i + batch_size] batch_prompts = [] for chunk in batch: system_prompt = '<|im_start|>system\nJudge whether the Document meets the requirements based on the Query and the Instruct provided. Note that the answer can only be "yes" or "no".<|im_end|>\n' user_prompt = ( f'<|im_start|>user\n' f': {instruction}\n' f': {query}\n' f': {chunk[:8000]}\n' # Limit document length f'<|im_end|>\n' f'<|im_start|>assistant\n\n\n\n\n' ) batch_prompts.append(system_prompt + user_prompt) # Score batch for prompt, chunk in zip(batch_prompts, batch): try: resp = requests.post( f"{OLLAMA_BASE_URL}/api/generate", json={ "model": RERANKER_MODEL_OLLAMA, "prompt": prompt, "max_tokens": 3, "temperature": 0.0, "stream": False }, timeout=20 ) j = resp.json() response_text = j.get("response", "").strip().lower() # Enhanced scoring with confidence levels score = parse_reranker_response(response_text) batch_scores.append((chunk, score)) except requests.exceptions.Timeout: logger.warning("Reranker request timeout, assigning default score") batch_scores.append((chunk, 0.0)) except Exception as e: logger.warning(f"Rerank call failed: {e}") batch_scores.append((chunk, 0.0)) # Sort by score descending and return top_k batch_scores.sort(key=lambda x: x[1], reverse=True) return batch_scores[:top_k] def parse_reranker_response(response_text: str) -> float: """Parse Qwen3 Reranker response and convert to confidence score.""" response_lower = response_text.strip().lower() # Exact matches from the model if response_lower == "yes": return 1.0 elif response_lower == "no": return 0.0 # Handle variations and partial matches yes_indicators = ["yes", "relevant", "correct", "matches", "appropriate", "suitable"] no_indicators = ["no", "irrelevant", "incorrect", "unrelated", "inappropriate"] yes_count = sum(1 for indicator in yes_indicators if indicator in response_lower) no_count = sum(1 for indicator in no_indicators if indicator in response_lower) if yes_count > no_count: return 0.8 # Likely relevant but not confident elif no_count > yes_count: return 0.2 # Likely irrelevant but not confident else: # Ambiguous response return 0.5 # ----------------------------- # Comment-aware reference search (ripgrep) # ----------------------------- def get_context_with_comments(file_path: Path, match_line: int, max_context_lines: int = 5) -> str: try: lines = file_path.read_text(encoding="utf-8").splitlines() except Exception: return "" idx = match_line - 1 context_lines = [] look = idx - 1 collected = 0 while look >= 0 and collected < max_context_lines: line = lines[look] stripped = line.strip() if stripped.startswith(("#", "//", "--")) or stripped == "": context_lines.insert(0, line) look -= 1 collected += 1 else: break context_lines.append(lines[idx] if 0 <= idx < len(lines) else "") return "\n".join(context_lines) def search_symbol(symbol: str, root_dir: Path, top_k: int = 20) -> List[Dict[str, Any]]: matches = [] rg_cmd = [ "rg", "-n", "-i", "-C", "3", "--no-ignore", "--no-heading", "--color", "never", "--glob", "!node_modules", "--glob", "!__pycache__", "--glob", "!.git", ] for ext in SUPPORTED_EXTENSIONS: rg_cmd.extend(["--glob", f"*{ext}"]) rg_cmd.extend(["--", symbol, "."]) try: proc = subprocess.run( rg_cmd, cwd=root_dir, capture_output=True, text=True, timeout=30 ) if proc.returncode not in (0, 1): logger.warning(f"ripgrep returned code {proc.returncode}: {proc.stderr}") return [] for line in proc.stdout.splitlines(): if not line.strip() or line.startswith("--"): continue parts = line.split(":", 2) if len(parts) < 3: continue file_rel, lineno_str, snippet = parts try: lineno = int(lineno_str) except ValueError: continue matches.append({ "file": file_rel, "line": lineno, "context": snippet.strip() }) except subprocess.TimeoutExpired: logger.warning("ripgrep timed out") except Exception as e: logger.warning(f"ripgrep failed: {e}") return matches def _extract_imports(source: str) -> list[str]: """Return a list of fully‑qualified imports found in `source`.""" return [m.group(1).strip() for m in re.finditer(r'^\s*use\s+([^;]+);', source, re.MULTILINE)] def _extract_calls(source: str) -> list[str]: """Return a list of called function names (bare name only).""" return [m.group(1).split("::")[-1] for m in re.finditer(r'([A-Za-z_][A-Za-z0-9_:]*)\s*\(', source, re.MULTILINE)] # ----------------------------- # Additional tool: read file lines with context # ----------------------------- def read_file_lines(path: str, start: int = 1, end: Optional[int] = None) -> str: """ Read file lines with intelligent context inclusion. - Includes requested lines - Expands to include full function/class definition if within one - Includes leading comments (docstrings, inline comments) - Shows function signature, return type, decorators """ # Handle both absolute and relative paths if path.startswith('/'): p = Path(path) else: p = CODEBASE_PATH / Path(path) # Security: Validate path is within codebase try: resolved = p.resolve() codebase_resolved = CODEBASE_PATH.resolve() if not resolved.is_relative_to(codebase_resolved): return f"Error: Path '{path}' is outside the codebase directory" except (ValueError, OSError) as e: return f"Error: Invalid path '{path}': {e}" try: content = p.read_text(encoding="utf-8") lines = content.splitlines() if end is None: end = start + 50 # reasonable default # Clamp to valid range start_i = max(1, start) end_i = min(len(lines), end) # Detect language language = detect_language(p) # Build context-aware line range context_start, context_end, context_info = find_context_boundaries( lines, start_i, end_i, language ) # Extract the context lines context_lines = lines[context_start - 1:context_end] # Build output with metadata output = [] output.append(f"File: {path}") output.append(f"Language: {language}") output.append(f"Requested: lines {start_i}-{end_i}") output.append(f"Showing: lines {context_start}-{context_end} (with context)") if context_info: output.append(f"Context: {context_info}") output.append("\n" + "="*60) # Add line numbers for i, line in enumerate(context_lines, start=context_start): # Highlight the originally requested range marker = ">>> " if start_i <= i <= end_i else " " output.append(f"{marker}{i:4d} | {line}") output.append("="*60) return "\n".join(output) except Exception as e: return f"Failed to read {path}: {e}" def find_context_boundaries(lines: List[str], start: int, end: int, language: str) -> Tuple[int, int, str]: """ Find intelligent context boundaries around the requested line range. Returns: (context_start_line, context_end_line, description) """ # Convert to 0-based indexing for processing start_idx = start - 1 end_idx = end - 1 context_info = [] # Expand upward to include leading comments and function/class headers context_start = start # Step 1: Include leading comments above start comment_start = find_leading_comments_start(lines, start_idx, language) if comment_start < start_idx: context_start = comment_start + 1 context_info.append("leading comments") # Step 2: Check if we're inside a function/class and include its signature func_start, func_end, func_name, func_type = find_enclosing_function_or_class( lines, start_idx, end_idx, language ) if func_start is not None: context_start = min(context_start, func_start + 1) if func_end is not None and func_end > end_idx: context_end = func_end + 1 else: context_end = end if func_name: context_info.append(f"{func_type} '{func_name}'") else: context_end = end # Step 3: Include decorators (Python) or annotations (Java) if language == "python": decorator_start = find_decorators_start(lines, context_start - 1) if decorator_start < context_start - 1: context_start = decorator_start + 1 context_info.append("decorators") info_str = " + ".join(context_info) if context_info else "no additional context" return context_start, context_end, info_str def find_leading_comments_start(lines: List[str], line_idx: int, language: str) -> int: """Find the start of leading comments above the given line.""" comment_patterns = { "python": ("#", '"""', "'''"), "javascript": ("//", "/*"), "typescript": ("//", "/*"), "go": ("//", "/*"), "java": ("//", "/*"), "sql": ("--", "/*"), "rust": ("//", "/*"), } patterns = comment_patterns.get(language, ("#", "//", "/*", "--")) start_idx = line_idx i = line_idx - 1 # Track if we're in a block comment in_block = False while i >= 0: line = lines[i].strip() # Empty lines are OK if we already have comments if not line: if i < line_idx - 1: # Allow one blank line between comments i -= 1 continue else: break # Check for block comment end (we're going backwards) if "*/" in line: in_block = True start_idx = i i -= 1 continue # Check for block comment start if in_block and ("/*" in line or "/**" in line): start_idx = i in_block = False i -= 1 continue # Check for single-line comments is_comment = any(line.startswith(p) for p in patterns) if is_comment or in_block: start_idx = i i -= 1 else: break return start_idx def find_decorators_start(lines: List[str], line_idx: int) -> int: """Find Python decorators above the given line.""" i = line_idx - 1 while i >= 0: line = lines[i].strip() if line.startswith("@"): i -= 1 elif not line: # Allow blank lines i -= 1 else: return i + 1 return 0 def find_enclosing_function_or_class(lines: List[str], start_idx: int, end_idx: int, language: str) -> Tuple[Optional[int], Optional[int], Optional[str], Optional[str]]: """ Find the function or class that encloses the given line range. Returns: (start_line_idx, end_line_idx, name, type) or (None, None, None, None) """ if language == "python": return find_python_function_or_class(lines, start_idx, end_idx) elif language in ("javascript", "typescript"): return find_js_function(lines, start_idx, end_idx) elif language == "go": return find_go_function(lines, start_idx, end_idx) elif language == "java": return find_java_method_or_class(lines, start_idx, end_idx) elif language == "sql": return find_sql_function_or_block(lines, start_idx, end_idx) return None, None, None, None def find_python_function_or_class(lines: List[str], start_idx: int, end_idx: int) -> Tuple[Optional[int], Optional[int], Optional[str], Optional[str]]: """Find enclosing Python function or class.""" # Look backwards for def/class with proper indentation target_indent = None for i in range(start_idx, -1, -1): line = lines[i] stripped = line.lstrip() if stripped.startswith(("def ", "class ", "async def ")): indent = len(line) - len(stripped) # Check if this could be our enclosing scope if target_indent is None or indent < target_indent: # Extract name match = re.match(r'(?:async\s+)?(?:def|class)\s+(\w+)', stripped) if match: name = match.group(1) func_type = "class" if stripped.startswith("class") else "function" # Find the end by looking for next same-or-lower indent non-empty line end_i = find_python_block_end(lines, i, indent) return i, end_i, name, func_type return None, None, None, None def find_python_block_end(lines: List[str], start_idx: int, base_indent: int) -> Optional[int]: """Find the end of a Python block based on indentation.""" for i in range(start_idx + 1, len(lines)): line = lines[i] stripped = line.lstrip() if not stripped: # Skip empty lines continue indent = len(line) - len(stripped) if indent <= base_indent and stripped and not stripped.startswith("#"): return i - 1 return len(lines) - 1 def find_js_function(lines: List[str], start_idx: int, end_idx: int) -> Tuple[Optional[int], Optional[int], Optional[str], Optional[str]]: """Find enclosing JavaScript/TypeScript function.""" for i in range(start_idx, -1, -1): line = lines[i].strip() # Match various JS function patterns patterns = [ r'function\s+(\w+)', r'(?:const|let|var)\s+(\w+)\s*=\s*(?:async\s+)?\(', r'(\w+)\s*\([^)]*\)\s*{', # Arrow functions r'async\s+function\s+(\w+)', ] for pattern in patterns: match = re.search(pattern, line) if match: name = match.group(1) end_i = find_brace_block_end(lines, i) return i, end_i, name, "function" return None, None, None, None def find_go_function(lines: List[str], start_idx: int, end_idx: int) -> Tuple[Optional[int], Optional[int], Optional[str], Optional[str]]: """Find enclosing Go function.""" for i in range(start_idx, -1, -1): line = lines[i].strip() # Match Go function: func (receiver) name(params) returnType { match = re.match(r'func\s+(?:\([^)]+\)\s+)?(\w+)', line) if match: name = match.group(1) end_i = find_brace_block_end(lines, i) return i, end_i, name, "function" return None, None, None, None def find_java_method_or_class(lines: List[str], start_idx: int, end_idx: int) -> Tuple[Optional[int], Optional[int], Optional[str], Optional[str]]: """Find enclosing Java method or class.""" for i in range(start_idx, -1, -1): line = lines[i].strip() # Match class class_match = re.match(r'(?:public|private|protected)?\s*(?:static)?\s*class\s+(\w+)', line) if class_match: name = class_match.group(1) end_i = find_brace_block_end(lines, i) return i, end_i, name, "class" # Match method method_match = re.match(r'(?:public|private|protected)?\s*(?:static)?\s*(?:\w+(?:<[^>]+>)?)\s+(\w+)\s*\(', line) if method_match: name = method_match.group(1) end_i = find_brace_block_end(lines, i) return i, end_i, name, "method" return None, None, None, None def find_sql_function_or_block(lines: List[str], start_idx: int, end_idx: int) -> Tuple[Optional[int], Optional[int], Optional[str], Optional[str]]: """Find enclosing SQL function or procedure.""" for i in range(start_idx, -1, -1): line = lines[i].strip().upper() if line.startswith(("CREATE FUNCTION", "CREATE OR REPLACE FUNCTION", "CREATE PROCEDURE")): # Extract name match = re.search(r'(?:FUNCTION|PROCEDURE)\s+(\w+)', line, re.IGNORECASE) if match: name = match.group(1) # SQL functions typically end with $$ or END end_i = find_sql_function_end(lines, i) func_type = "procedure" if "PROCEDURE" in line else "function" return i, end_i, name, func_type return None, None, None, None def find_sql_function_end(lines: List[str], start_idx: int) -> Optional[int]: """Find end of SQL function (looks for $$ or END;).""" for i in range(start_idx + 1, len(lines)): line = lines[i].strip().upper() if "$$" in line or line.startswith("END;") or line == "END": return i return len(lines) - 1 def find_brace_block_end(lines: List[str], start_idx: int) -> Optional[int]: """Find the end of a brace-delimited block {}.""" brace_count = 0 started = False for i in range(start_idx, len(lines)): line = lines[i] for char in line: if char == '{': brace_count += 1 started = True elif char == '}': brace_count -= 1 if started and brace_count == 0: return i return len(lines) - 1 # ----------------------------- # MCP Tools # ----------------------------- @mcp.tool() def health_check() -> str: """What it does - Quick JSON status of the RAG server (ready, indexed path, chunk count, age, graph stats). When to use - Before any other query, to confirm the index is current and see codebase structure. When not to use - After you already know the server is healthy; it adds no value. Example - health_check() β†’ { "status":"ready","total_chunks":11234,"graph_nodes":4567,"graph_edges":12345,"index_age_hours":4.7 }""" try: with _startup_lock: # Check if we have a git repo in working_repo repo_info = "No repository loaded" if (CODEBASE_PATH / ".git").exists(): try: branch = subprocess.check_output( ["git", "-C", str(CODEBASE_PATH), "rev-parse", "--abbrev-ref", "HEAD"], text=True ).strip() repo_info = f"Loaded: {branch}" except: repo_info = "Git repository (no branch info)" # Load graph to check if it exists - FIX: Convert Path to string graph = LocalGraph(root_path=str(CODEBASE_PATH)) # Convert Path to str graph_loaded = graph.load() status = { "status": "ready" if vectorstore is not None else "no_index", "working_repository": str(CODEBASE_PATH), "repo_status": repo_info, "ollama_url": OLLAMA_BASE_URL, "config": { "embedding_model": EMBEDDING_MODEL, "vector_weight": VECTOR_WEIGHT, "bm25_weight": BM25_WEIGHT, "rerank_enabled": ENABLE_RERANK, "batch_size": EMBEDDING_BATCH_SIZE, "graph_enabled": True }, "Tools": { "NOTE": "THESE TOOLS ARE RESTRICTED BY .gitignore AS WELL AS .ragignore", "semantic_rag_search": "Hybrid RAG search with graph context (cross-file deps, inheritance, calls)", "find_code_references": "Find exact symbol references across codebase", "read_file_lines_tool": "Read specific file lines with syntax highlighting", "find_path": "Find files by glob patterns", "grep": "Search file contents with regex", "list_directory": "List directory contents" } } if bm25_corpus: status["statistics"] = { "total_chunks": len(bm25_corpus), "by_language": count_by_language(chunks_metadata), "by_type": count_by_type(chunks_metadata), "avg_chunk_size_chars": round(calculate_avg_chunk_size(bm25_corpus), 1), "index_age_hours": round((time.time() - index_build_time) / 3600, 1) if index_build_time > 0 else None } # Add graph statistics - FIX: Use graph.graph.nodes instead of graph.nodes if graph_loaded: graph_stats = _calculate_graph_stats(graph) status["graph"] = graph_stats else: status["graph"] = { "status": "not_loaded", "nodes": 0, "edges": 0, "message": "Graph will be built on next index rebuild" } return json.dumps(status, indent=2) except Exception as e: return json.dumps({"status": "error", "message": str(e)}, indent=2) def _calculate_graph_stats(graph: LocalGraph) -> dict: """Calculate comprehensive graph statistics.""" # FIX: Use graph.graph.nodes and graph.graph.edges instead of graph.nodes/graph.edges if graph.graph.number_of_nodes() == 0: return {"status": "empty", "nodes": 0, "edges": 0} # Basic counts nodes = graph.graph.number_of_nodes() edges = graph.graph.number_of_edges() # Count by node type node_types = {} for _, data in graph.graph.nodes(data=True): # FIX: graph.graph.nodes node_type = data.get('type', 'unknown') node_types[node_type] = node_types.get(node_type, 0) + 1 # Count by edge type edge_types = {} for _, _, data in graph.graph.edges(data=True): # FIX: graph.graph.edges edge_type = data.get('type', 'unknown') edge_types[edge_type] = edge_types.get(edge_type, 0) + 1 # Language distribution languages = {} for _, data in graph.graph.nodes(data=True): # FIX: graph.graph.nodes lang = data.get('lang', 'unknown') # Changed from 'language' to 'lang' to match our new attribute languages[lang] = languages.get(lang, 0) + 1 # File statistics files = [n for n, d in graph.graph.nodes(data=True) if d.get('type') == 'File'] # FIX: graph.graph.nodes # Relationship density avg_edges_per_node = edges / nodes if nodes > 0 else 0 # Most connected nodes (hubs) degree_centrality = dict(graph.graph.degree()) # FIX: graph.graph.degree() top_hubs = sorted(degree_centrality.items(), key=lambda x: x[1], reverse=True)[:5] top_hubs_info = [] for node_id, degree in top_hubs: node_data = graph.graph.nodes[node_id] # FIX: graph.graph.nodes top_hubs_info.append({ "name": node_data.get('name', node_id), "type": node_data.get('type', 'unknown'), "file": node_data.get('file', ''), "connections": degree }) return { "status": "loaded", "nodes": nodes, "edges": edges, "node_types": node_types, "edge_types": edge_types, "languages": languages, "files": len(files), "avg_edges_per_node": round(avg_edges_per_node, 2), "top_connected_nodes": top_hubs_info, "relationship_density": "high" if avg_edges_per_node > 2.0 else "medium" if avg_edges_per_node > 1.0 else "low" } @mcp.tool() def semantic_rag_search(query: str, top_k: int = 5, rerank: bool = True) -> str: """Hybrid RAG search: returns ranked code snippets + file/line + docstrings + graph context (cross-file deps, inheritance, callers/callees, type refs). Use for: Open questions ("how does auth work?"), unknown patterns, architecture exploration. Natural language queries. Don't use for: Single symbol/exact phrase (use find_code_references or grep then read_file_lines_tool or read), exploring known entity relationships. rerank=False: fast, 65% accuracy | rerank=True: slower, 95% accuracy Output includes: code content, docs, cross_file:[calls/extends/uses@filepath], used_by:[entity@filepath], file:[same-file entities] Example: semantic_rag_search("user authentication", 5, True) β†’ 5 results with code + "cross_file:calls:ValidateToken@auth/jwt.go|used_by:setupAuth@server/routes.go" """ if vectorstore is None or bm25_corpus is None: return "❌ Index not built. Please call init_repo() or rebuild_index() first." graph = LocalGraph(root_path=str(CODEBASE_PATH)) graph_loaded = graph.load() if not graph_loaded: logger.warning("Graph not loaded - search will proceed without graph context") try: hy = hybrid_search(query, k=max(top_k, RERANK_TOP_N)) if ENABLE_RERANK and rerank: try: candidates = [t for t, m, s in hy] rr = rerank_with_ollama_enhanced(query, candidates[:RERANK_TOP_N], top_k=top_k) # Build enhanced results with graph context results = [] for i, (chunk_text, score) in enumerate(rr, 1): meta = next((m for t, m, s in hy if t == chunk_text), {}) results.append(_build_enhanced_result(chunk_text, meta, score, graph)) return _format_enhanced_results(results, query, "search") except Exception as e: logger.warning(f"Rerank step failed: {e}") # Fall through to non-reranked results # Non-reranked results with graph context results = [] for i, (text, meta, score) in enumerate(hy[:top_k], 1): results.append(_build_enhanced_result(text, meta, score, graph)) return _format_enhanced_results(results, query, "search") except RuntimeError as e: return f"Error: {e}" except Exception as e: logger.exception("Search failed") return f"Search failed: {e}" def _build_enhanced_result(chunk_text: str, meta: dict, score: float, graph: Optional[LocalGraph] = None) -> dict: """Build enhanced result with graph context and important metadata.""" # Extract key metadata file_path = meta.get('file', '') line_num = meta.get('line', '') entity_type = meta.get('type', 'code') language = meta.get('language', '') entity_name = meta.get('name', '') # Extract clean content and docstring clean_content = _extract_clean_content(chunk_text) docstring = _extract_docstring(chunk_text) # Get graph context for this result graph_context = _get_result_graph_context(meta, graph) return { 'file': file_path, 'line': line_num, 'type': entity_type, 'language': language, 'name': entity_name, 'score': score, 'content': clean_content, 'docstring': docstring, 'graph_context': graph_context } def _get_result_graph_context(meta: dict, graph: Optional[LocalGraph] = None) -> dict: """Fixed graph context using enhanced graph methods.""" if graph is None or graph.graph.number_of_nodes() == 0: return {} context = {} file_path = meta.get('file', '') entity_name = meta.get('name', '') entity_type = meta.get('type', '') language = meta.get('language', '') # Try multiple node ID formats to find the right one node_candidates = [] # Format 1: Standard node ID if entity_name and file_path and language and entity_type: node_candidates.append(f"{language}::{entity_type}::{file_path}::{entity_name}") # Format 2: Method with class name if entity_type in ['method', 'function'] and meta.get('class'): class_name = meta.get('class') node_candidates.append(f"{language}::{entity_type}::{file_path}::{class_name}.{entity_name}") # Format 3: Svelte component if entity_type == 'component' and language == 'svelte': node_candidates.append(f"svelte::component::{file_path}::{entity_name}") # Try to find the node using graph's enhanced search target_node_id = None for candidate in node_candidates: if candidate in graph.graph.nodes: target_node_id = candidate break # If still not found, try attribute-based search if not target_node_id and entity_name and file_path: matches = list(graph.find_nodes_by_attributes( name=entity_name, file=file_path, type=entity_type if entity_type != 'code' else None )) if matches: target_node_id = matches[0][0] # Take first match if not target_node_id: return context # Use the enhanced graph methods to get relationships cross_file_out, cross_file_in = graph.get_cross_file_relationships(target_node_id, file_path) if cross_file_out: context['cross_file'] = cross_file_out if cross_file_in: context['used_by'] = cross_file_in # Get same-file relationships same_file_rels = [] for neighbor_id in graph.graph.successors(target_node_id): if neighbor_id in graph.graph.nodes: neighbor_file = graph.graph.nodes[neighbor_id].get('file', '') if neighbor_file == file_path: # Same file edge_data = graph.graph.edges[target_node_id, neighbor_id] rel_type = edge_data.get('type', 'related') neighbor_name = graph.graph.nodes[neighbor_id].get('name', '') neighbor_type = graph.graph.nodes[neighbor_id].get('type', '') same_file_rels.append(f"{rel_type}:{neighbor_type}:{neighbor_name}") if same_file_rels: context['same_file'] = same_file_rels # Get meaningful file entities (filter out noise) meaningful_types = {'function', 'method', 'class', 'struct', 'interface', 'enum', 'module', 'component', 'trait', 'impl'} file_entities = [] file_node_id = f"file::{file_path}" if file_node_id in graph.graph.nodes: for neighbor_id, neighbor_data in graph.neighbors(file_node_id): neighbor_type = neighbor_data.get('type', '') neighbor_name = neighbor_data.get('name', '') # Filter criteria if (neighbor_type in meaningful_types and neighbor_name and len(neighbor_name) > 1 and # No single letters neighbor_name != entity_name and not neighbor_name.startswith(('c:', 'b:', 'p:', 'r:', 't:', 'v:'))): # No import aliases file_entities.append(f"{neighbor_type}:{neighbor_name}") if file_entities: context['file_entities'] = file_entities return context def _format_enhanced_results(results: List[Dict], query: str, result_type: str) -> str: """Comprehensive format with ALL information preserved - optimized for LLM analysis.""" if not results: return f"# {result_type.title()}: {query}\nNo results found.\n" lines = [ f"# {result_type.title()}: {query}", f"results_count={len(results)}" ] for i, result in enumerate(results, 1): # Extract ALL metadata file_path = result.get('file', '') line_num = result.get('line', '') entity_type = result.get('type', 'code') entity_name = result.get('name', '') language = result.get('language', '') score = result.get('score', 0.0) content = result.get('content', '') # NO TRUNCATION - let LLM handle it docstring = result.get('docstring', '') graph_context = result.get('graph_context', {}) # Build comprehensive result block lines.append(f"\n--- RESULT {i} ---") lines.append(f"file:{file_path}:{line_num}") lines.append(f"entity:{entity_type}:{entity_name}") lines.append(f"language:{language}") lines.append(f"score:{score:.3f}") # Content (FULL, no truncation) if content: lines.append(f"content: {content}") # Docstring if available if docstring: lines.append(f"doc: {docstring}") # Graph context - PRESERVE EVERYTHING if graph_context: lines.append("graph_context:") # Cross-file dependencies (highest value) if graph_context.get('cross_file'): lines.append(f" cross_file: {' | '.join(graph_context['cross_file'])}") # Reverse dependencies if graph_context.get('used_by'): lines.append(f" used_by: {' | '.join(graph_context['used_by'])}") # Same-file relationships if graph_context.get('same_file'): lines.append(f" same_file_rels: {' | '.join(graph_context['same_file'])}") # All relationships (comprehensive) if graph_context.get('all_rels'): lines.append(f" all_relationships: {' | '.join(graph_context['all_rels'][:15])}") # Reasonable limit # File contents if graph_context.get('file_contents'): lines.append(f" file_entities: {' | '.join(graph_context['file_contents'])}") return "\n".join(lines) def _extract_clean_content(chunk_text: str) -> str: """Extract clean content WITHOUT truncation - preserve full context.""" lines = chunk_text.split('\n') content_lines = [] # Skip only the pure metadata header lines skip_metadata = True for line in lines: stripped = line.strip() if skip_metadata: # Only skip actual metadata labels, not content if any(stripped.startswith(prefix) for prefix in ['File:', 'Type:', 'Name:', 'Language:', 'Doc:']): continue # Stop skipping when we hit actual code/content if stripped and not stripped.startswith(('File:', 'Type:', 'Name:', 'Language:', 'Doc:')): skip_metadata = False content_lines.append(stripped) else: content_lines.append(stripped) # Join with spaces but preserve ALL content content = ' '.join(content_lines) content = re.sub(r'\s+', ' ', content) # Only normalize whitespace return content.strip() def _build_enhanced_result(chunk_text: str, meta: dict, score: float, graph: Optional[LocalGraph] = None) -> dict: """Build comprehensive result with ALL context preserved.""" # Extract key metadata file_path = meta.get('file', '') line_num = meta.get('line', '') entity_type = meta.get('type', 'code') language = meta.get('language', '') entity_name = meta.get('name', '') # Extract FULL content and docstring clean_content = _extract_clean_content(chunk_text) docstring = _extract_docstring(chunk_text) # Get COMPREHENSIVE graph context graph_context = _get_result_graph_context(meta, graph) return { 'file': file_path, 'line': line_num, 'type': entity_type, 'language': language, 'name': entity_name, 'score': score, 'content': clean_content, # NO TRUNCATION 'docstring': docstring, 'graph_context': graph_context } def _extract_docstring(chunk_text: str) -> str: """Extract docstring or comments from chunk text.""" lines = chunk_text.split('\n') doc_lines = [] # Look for docstring patterns in_docstring = False for line in lines: stripped = line.strip() # Java/Go style comments if stripped.startswith('//') or stripped.startswith('/*') or stripped.startswith('*'): doc_lines.append(stripped) # Python style docstrings elif '"""' in line or "'''" in line: if not in_docstring: in_docstring = True else: in_docstring = False break elif in_docstring: doc_lines.append(stripped) # Specific doc patterns in the metadata section elif stripped.startswith('Doc:') and len(stripped) > 4: doc_content = stripped[4:].strip() if doc_content and doc_content not in ['None', '""']: doc_lines.append(doc_content) # Clean up docstring if doc_lines: docstring = ' '.join(doc_lines) docstring = re.sub(r'\s+', ' ', docstring) return docstring.strip()[:300] # Reasonable limit return "" def _format_enhanced_results(results: List[Dict], query: str, result_type: str) -> str: """Comprehensive format with ALL information preserved - optimized for LLM analysis.""" if not results: return f"# {result_type.title()}: {query}\nNo results found.\n" lines = [ f"# {result_type.title()}: {query}", f"results_count={len(results)}" ] for i, result in enumerate(results, 1): # Extract ALL metadata file_path = result.get('file', '') line_num = result.get('line', '') entity_type = result.get('type', 'code') entity_name = result.get('name', '') language = result.get('language', '') score = result.get('score', 0.0) content = result.get('content', '') # NO TRUNCATION - let LLM handle it docstring = result.get('docstring', '') graph_context = result.get('graph_context', {}) # Build comprehensive result block lines.append(f"\n--- RESULT {i} ---") lines.append(f"file:{file_path}:{line_num}") lines.append(f"entity:{entity_type}:{entity_name}") lines.append(f"language:{language}") lines.append(f"score:{score:.3f}") # Content (FULL, no truncation) if content: lines.append(f"content: {content}") # Docstring if available if docstring: lines.append(f"doc: {docstring}") # Graph context - PRESERVE EVERYTHING if graph_context: lines.append("graph_context:") # Cross-file dependencies (highest value) if graph_context.get('cross_file'): lines.append(f" cross_file: {' | '.join(graph_context['cross_file'])}") # Reverse dependencies if graph_context.get('used_by'): lines.append(f" used_by: {' | '.join(graph_context['used_by'])}") # Same-file relationships if graph_context.get('same_file'): lines.append(f" same_file_rels: {' | '.join(graph_context['same_file'])}") # All relationships (comprehensive) if graph_context.get('all_rels'): lines.append(f" all_relationships: {' | '.join(graph_context['all_rels'][:15])}") # Reasonable limit # File contents if graph_context.get('file_contents'): lines.append(f" file_entities: {' | '.join(graph_context['file_contents'])}") return "\n".join(lines) def _extract_clean_content(chunk_text: str) -> str: """Extract clean content WITHOUT truncation - preserve full context.""" lines = chunk_text.split('\n') content_lines = [] # Skip only the pure metadata header lines skip_metadata = True for line in lines: stripped = line.strip() if skip_metadata: # Only skip actual metadata labels, not content if any(stripped.startswith(prefix) for prefix in ['File:', 'Type:', 'Name:', 'Language:', 'Doc:']): continue # Stop skipping when we hit actual code/content if stripped and not stripped.startswith(('File:', 'Type:', 'Name:', 'Language:', 'Doc:')): skip_metadata = False content_lines.append(stripped) else: content_lines.append(stripped) # Join with spaces but preserve ALL content content = ' '.join(content_lines) content = re.sub(r'\s+', ' ', content) # Only normalize whitespace return content.strip() @mcp.tool() def find_code_references(symbol: str, top_k: int = 20) -> str: """What it does - Fast, unranked lookup of all file/line occurrences of a symbol. When to use - After identifying a symbol in a search hit, or when you need every use of a class/method. When not to use - When you need context-rich code; use read_file_lines_tool after getting the locations. Example - find_code_references("AuthService") β†’ list of references.""" matches = search_symbol(symbol, CODEBASE_PATH) if not matches: return f"# References: {symbol}\nNo references found.\n" # Build clean TOON format directly results = [] for i, m in enumerate(matches[:top_k], 1): results.append({ 'file': m['file'], 'line': m['line'], 'context': m['context'] }) return _format_reference_results(results, symbol) def _format_reference_results(results: List[Dict], symbol: str) -> str: """Format reference results in clean TOON format.""" if not results: return f"# References: {symbol}\nNo references found.\n" lines = [f"# References: {symbol}", f"references[{len(results)}]{{file,line,context}}:"] for result in results: row = [ result.get('file', ''), result.get('line', ''), result.get('context', '')[:300] # Reasonable limit ] # Escape fields if needed escaped_row = [] for field in row: field_str = str(field) if any(char in field_str for char in [',', '"', '\n', '\r']): escaped = field_str.replace('"', '\\"') escaped_row.append(f'"{escaped}"') else: escaped_row.append(field_str) lines.append(" " + ",".join(escaped_row)) return "\n".join(lines) @mcp.tool() def read_file_lines_tool(path: str, start: int = 1, end: Optional[int] = None) -> str: """What it does - Reads a requested line range and auto-expands to include surrounding context (function body, docstring, comments). When to use - To inspect a specific hit from find_code_references or a snippet from semantic_rag_search. When not to use - For browsing the entire file or unrelated sections; use semantic_rag_search first to pinpoint relevant parts. Example - read_file_lines_tool("src/auth/login.py", start=45, end=60) β†’ full function with context. Note: This tool treats / as '/home/popertots/Crussell/' so adjust paths accordingly""" return EnhancedToon.file_content_results(read_file_lines(path, start, end), path) @mcp.tool() def init_repo(git_url: str) -> str: """What it does - Initialise the RAG stack with a brand-new Git repository.""" import shutil from pathlib import Path import subprocess # ----- 1. Always wipe and recreate working_repo --------------------- if CODEBASE_PATH.exists(): shutil.rmtree(CODEBASE_PATH) logger.info(f"Wiped existing working repo: {CODEBASE_PATH}") CODEBASE_PATH.mkdir(exist_ok=True) # ----- 2. Clone fresh ----------------------------------------------- try: subprocess.run( ["git", "clone", git_url, str(CODEBASE_PATH)], check=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, ) logger.info(f"Cloned {git_url} to {CODEBASE_PATH}") except subprocess.CalledProcessError as e: return f"❌ Failed to clone {git_url}: {e.stderr.strip()}" # ----- 3. Reset ALL global state ------------------------------------ global vectorstore, bm25, bm25_corpus, chunks_metadata, index_build_time vectorstore = None bm25 = None bm25_corpus = None chunks_metadata = None index_build_time = 0 # ----- 4. Remove any stale indices ---------------------------------- for artefact in ("chroma_db", "bm25_index.json", "embeddings", ".mcp_cache"): artefact_path = CODEBASE_PATH / artefact if artefact_path.exists(): try: if artefact_path.is_dir(): shutil.rmtree(artefact_path) else: artefact_path.unlink() logger.info(f"Removed stale artefact: {artefact_path}") except Exception as e: logger.warning(f"Could not delete {artefact_path}: {e}") # ----- 5. Re-build the index ---------------------------------------- try: build_indexes() logger.info("Index rebuilt successfully") except Exception as e: logger.exception("Index rebuild failed") return f"❌ Index rebuild failed: {e}" # ----- 6. Report success -------------------------------------------- try: branch_name = subprocess.check_output( ["git", "-C", str(CODEBASE_PATH), "rev-parse", "--abbrev-ref", "HEAD"], text=True, ).strip() git_hash = subprocess.check_output( ["git", "-C", str(CODEBASE_PATH), "rev-parse", "HEAD"], text=True, ).strip() # Verify the index was actually built if vectorstore is None or bm25_corpus is None: return f"⚠️ Repository cloned but index failed to build properly" return f"βœ… Initialized repository at {CODEBASE_PATH}\n Branch: {branch_name}\n Commit: {git_hash[:8]}\n Indexed chunks: {len(bm25_corpus)}" except Exception as e: return f"⚠️ Repository cloned but status check failed: {e}" @mcp.tool() def rebuild_index() -> str: """What it does - Re-creates all embeddings, BM25, and metadata indexes after major code changes. When to use - When the codebase has been pulled or refactored and you suspect stale search results, and only after explicit instruction to do so. When not to use - On every query; it's expensive and unnecessary if the index is already up-to-date. Example - rebuild_index() β†’ 'βœ… Index rebuilt (13.2 s); 11,234 chunks'""" with _startup_lock: try: # Clear LRU caches parse_python_file_cached.cache_clear() parse_sql_file_cached.cache_clear() parse_go_file_cached.cache_clear() parse_java_file_cached.cache_clear() parse_svelte_file_cached.cache_clear() parse_shell_file_cached.cache_clear() LocalGraph(root_path=CODEBASE_PATH).clear() t0 = time.time() build_indexes() dt = time.time() - t0 return f"βœ… Index rebuilt successfully! (took {dt:.1f}s)\n\nStatistics:\n" + json.dumps({ "total_chunks": len(bm25_corpus), "by_language": count_by_language(chunks_metadata), "by_type": count_by_type(chunks_metadata) }, indent=2) except Exception as e: logger.exception("Index rebuild failed") return f"❌ Rebuild failed: {e}" # ----------------------------- # Startup # ----------------------------- def signal_handler(sig, frame): """Handle graceful shutdown on Ctrl+C.""" logger.info("\n=== Shutdown signal received ===") # Clear any LRU caches to release memory try: parse_python_file_cached.cache_clear() parse_sql_file_cached.cache_clear() parse_go_file_cached.cache_clear() parse_java_file_cached.cache_clear() parse_rust_file_cached.cache_clear() parse_svelte_file_cached.cache_clear() parse_shell_file_cached.cache_clear() logger.info("βœ“ Caches cleared") except Exception as e: logger.warning(f"Cache clearing error: {e}") # Note: Chroma auto-persists, no explicit close needed logger.info("βœ“ Indexes are already persisted to disk") logger.info("Goodbye!\n") # Force exit immediately to avoid thread hang os._exit(0) if __name__ == "__main__": # Register signal handlers signal.signal(signal.SIGINT, signal_handler) signal.signal(signal.SIGTERM, signal_handler) try: logger.info("="*60) logger.info("πŸš€ MCP RAG Server is ready!") logger.info("Press Ctrl+C to stop") # Just run in stdio mode by default mcp.run(transport='stdio') except KeyboardInterrupt: signal_handler(signal.SIGINT, None) except Exception as e: logger.exception("Fatal error during startup") os._exit(1)