stravinsky 0.2.67__py3-none-any.whl → 0.4.18__py3-none-any.whl

mcp_bridge/tools/semantic_search.py

@@ -0,0 +1,3042 @@
+"""
+Semantic Code Search - Vector-based code understanding
+
+Uses ChromaDB for persistent vector storage with multiple embedding providers:
+- Ollama (local, free) - nomic-embed-text (768 dims)
+- Mxbai (local, free) - mxbai-embed-large (1024 dims, better for code)
+- Gemini (cloud, OAuth) - gemini-embedding-001 (768-3072 dims)
+- OpenAI (cloud, OAuth) - text-embedding-3-small (1536 dims)
+- HuggingFace (cloud, token) - sentence-transformers/all-mpnet-base-v2 (768 dims)
+
+Enables natural language queries like "find authentication logic" without
+requiring exact pattern matching.
+
+Architecture:
+- Per-project ChromaDB storage at ~/.stravinsky/vectordb/<project_hash>/
+- Lazy initialization on first query
+- Provider abstraction for embedding generation
+- Chunking strategy: function/class level with context
+"""
+
+import atexit
+import hashlib
+import logging
+import sys
+import threading
+from abc import ABC, abstractmethod
+from pathlib import Path
+from typing import Literal
+
+import httpx
+from tenacity import retry, retry_if_exception_type, stop_after_attempt, wait_exponential
+
+from mcp_bridge.auth.token_store import TokenStore
+from mcp_bridge.tools.query_classifier import QueryCategory, classify_query
+
+logger = logging.getLogger(__name__)
+
+
+# Lazy imports for watchdog (avoid startup cost)
+_watchdog = None
+_watchdog_import_lock = threading.Lock()
+
+
+def get_watchdog():
+    """Lazy import of watchdog components for file watching."""
+    global _watchdog
+    if _watchdog is None:
+        with _watchdog_import_lock:
+            if _watchdog is None:
+                from watchdog.events import FileSystemEventHandler
+                from watchdog.observers import Observer
+
+                _watchdog = {"Observer": Observer, "FileSystemEventHandler": FileSystemEventHandler}
+    return _watchdog
+
+
+# Embedding provider type
+EmbeddingProvider = Literal["ollama", "mxbai", "gemini", "openai", "huggingface"]
+
+# Lazy imports to avoid startup cost
+_chromadb = None
+_ollama = None
+_httpx = None
+_filelock = None
+_import_lock = threading.Lock()
+
+
+def get_filelock():
+    global _filelock
+    if _filelock is None:
+        with _import_lock:
+            if _filelock is None:
+                import filelock
+
+                _filelock = filelock
+    return _filelock
+
+
+def get_chromadb():
+    global _chromadb
+    if _chromadb is None:
+        with _import_lock:
+            if _chromadb is None:
+                try:
+                    import chromadb
+
+                    _chromadb = chromadb
+                except ImportError as e:
+                    import sys
+
+                    if sys.version_info >= (3, 14):
+                        raise ImportError(
+                            "ChromaDB is not available on Python 3.14+. "
+                            "Semantic search is not supported on Python 3.14 yet. "
+                            "Use Python 3.11-3.13 for semantic search features."
+                        ) from e
+                    raise
+    return _chromadb
+
+
+def get_ollama():
+    global _ollama
+    if _ollama is None:
+        with _import_lock:
+            if _ollama is None:
+                import ollama
+
+                _ollama = ollama
+    return _ollama
+
+
+def get_httpx():
+    global _httpx
+    if _httpx is None:
+        with _import_lock:
+            if _httpx is None:
+                import httpx
+
+                _httpx = httpx
+    return _httpx
+
+
+# ========================
+# GITIGNORE MANAGER
+# ========================
+
+# Lazy import for pathspec
+_pathspec = None
+_pathspec_lock = threading.Lock()
+
+
+def get_pathspec():
+    """Lazy import of pathspec for gitignore pattern matching."""
+    global _pathspec
+    if _pathspec is None:
+        with _pathspec_lock:
+            if _pathspec is None:
+                import pathspec
+
+                _pathspec = pathspec
+    return _pathspec
+
+
+class GitIgnoreManager:
+    """Manages .gitignore and .stravignore pattern matching.
+
+    Loads and caches gitignore-style patterns from:
+    - .gitignore (standard git ignore patterns)
+    - .stravignore (Stravinsky-specific ignore patterns)
+
+    Patterns are combined and cached per project for efficient matching.
+    The manager automatically reloads patterns if the ignore files are modified.
+    """
+
+    # Cache of GitIgnoreManager instances per project path
+    _instances: dict[str, "GitIgnoreManager"] = {}
+    _instances_lock = threading.Lock()
+
+    @classmethod
+    def get_instance(cls, project_path: Path) -> "GitIgnoreManager":
+        """Get or create a GitIgnoreManager for a project.
+
+        Args:
+            project_path: Root path of the project
+
+        Returns:
+            Cached GitIgnoreManager instance for the project
+        """
+        path_str = str(project_path.resolve())
+        if path_str not in cls._instances:
+            with cls._instances_lock:
+                if path_str not in cls._instances:
+                    cls._instances[path_str] = cls(project_path)
+        return cls._instances[path_str]
+
+    @classmethod
+    def clear_cache(cls, project_path: Path | None = None) -> None:
+        """Clear cached GitIgnoreManager instances.
+
+        Args:
+            project_path: Clear only this project's cache, or all if None
+        """
+        with cls._instances_lock:
+            if project_path is None:
+                cls._instances.clear()
+            else:
+                path_str = str(project_path.resolve())
+                cls._instances.pop(path_str, None)
+
+    def __init__(self, project_path: Path):
+        """Initialize the GitIgnoreManager.
+
+        Args:
+            project_path: Root path of the project
+        """
+        self.project_path = project_path.resolve()
+        self._spec = None
+        self._gitignore_mtime: float | None = None
+        self._stravignore_mtime: float | None = None
+        self._lock = threading.Lock()
+
+    def _get_file_mtime(self, file_path: Path) -> float | None:
+        """Get modification time of a file, or None if it doesn't exist."""
+        try:
+            return file_path.stat().st_mtime
+        except (OSError, FileNotFoundError):
+            return None
+
+    def _needs_reload(self) -> bool:
+        """Check if ignore patterns need to be reloaded."""
+        gitignore_path = self.project_path / ".gitignore"
+        stravignore_path = self.project_path / ".stravignore"
+
+        current_gitignore_mtime = self._get_file_mtime(gitignore_path)
+        current_stravignore_mtime = self._get_file_mtime(stravignore_path)
+
+        # Check if either file has been modified or if we haven't loaded yet
+        if self._spec is None:
+            return True
+
+        if current_gitignore_mtime != self._gitignore_mtime:
+            return True
+
+        if current_stravignore_mtime != self._stravignore_mtime:
+            return True
+
+        return False
+
+    def _load_patterns(self) -> None:
+        """Load patterns from .gitignore and .stravignore files."""
+        pathspec = get_pathspec()
+
+        patterns = []
+        gitignore_path = self.project_path / ".gitignore"
+        stravignore_path = self.project_path / ".stravignore"
+
+        # Load .gitignore patterns
+        if gitignore_path.exists():
+            try:
+                with open(gitignore_path, encoding="utf-8") as f:
+                    patterns.extend(f.read().splitlines())
+                self._gitignore_mtime = self._get_file_mtime(gitignore_path)
+                logger.debug(f"Loaded .gitignore from {gitignore_path}")
+            except Exception as e:
+                logger.warning(f"Failed to load .gitignore: {e}")
+                self._gitignore_mtime = None
+        else:
+            self._gitignore_mtime = None
+
+        # Load .stravignore patterns
+        if stravignore_path.exists():
+            try:
+                with open(stravignore_path, encoding="utf-8") as f:
+                    patterns.extend(f.read().splitlines())
+                self._stravignore_mtime = self._get_file_mtime(stravignore_path)
+                logger.debug(f"Loaded .stravignore from {stravignore_path}")
+            except Exception as e:
+                logger.warning(f"Failed to load .stravignore: {e}")
+                self._stravignore_mtime = None
+        else:
+            self._stravignore_mtime = None
+
+        # Filter out empty lines and comments
+        patterns = [p for p in patterns if p.strip() and not p.strip().startswith("#")]
+
+        # Create pathspec matcher
+        self._spec = pathspec.PathSpec.from_lines("gitwildmatch", patterns)
+        logger.debug(f"Loaded {len(patterns)} ignore patterns for {self.project_path}")
+
+    @property
+    def spec(self):
+        """Get the PathSpec matcher, reloading if necessary."""
+        with self._lock:
+            if self._needs_reload():
+                self._load_patterns()
+            return self._spec
+
+    def is_ignored(self, file_path: Path) -> bool:
+        """Check if a file path should be ignored.
+
+        Args:
+            file_path: Absolute or relative path to check
+
+        Returns:
+            True if the file matches any ignore pattern, False otherwise
+        """
+        try:
+            # Convert to relative path from project root
+            if file_path.is_absolute():
+                rel_path = file_path.resolve().relative_to(self.project_path)
+            else:
+                rel_path = file_path
+
+            # pathspec expects forward slashes and string paths
+            rel_path_str = str(rel_path).replace("\\", "/")
+
+            # Check against patterns
+            return self.spec.match_file(rel_path_str)
+        except ValueError:
+            # Path is outside project - not ignored by gitignore (but may be ignored for other reasons)
+            return False
+        except Exception as e:
+            logger.warning(f"Error checking ignore status for {file_path}: {e}")
+            return False
+
+    def filter_paths(self, paths: list[Path]) -> list[Path]:
+        """Filter a list of paths, removing ignored ones.
+
+        Args:
+            paths: List of paths to filter
+
+        Returns:
+            List of paths that are not ignored
+        """
+        return [p for p in paths if not self.is_ignored(p)]
+
+
+# ========================
+# EMBEDDING PROVIDERS
+# ========================
+
+
+class BaseEmbeddingProvider(ABC):
+    """Abstract base class for embedding providers."""
+
+    @abstractmethod
+    async def get_embedding(self, text: str) -> list[float]:
+        """Get embedding vector for text."""
+        pass
+
+    @abstractmethod
+    async def check_available(self) -> bool:
+        """Check if the provider is available and ready."""
+        pass
+
+    @property
+    @abstractmethod
+    def dimension(self) -> int:
+        """Return the embedding dimension for this provider."""
+        pass
+
+    @property
+    @abstractmethod
+    def name(self) -> str:
+        """Return the provider name."""
+        pass
+
+
+class OllamaProvider(BaseEmbeddingProvider):
+    """Ollama local embedding provider using nomic-embed-text."""
+
+    MODEL = "nomic-embed-text"
+    DIMENSION = 768
+
+    def __init__(self):
+        self._available: bool | None = None
+
+    @property
+    def dimension(self) -> int:
+        return self.DIMENSION
+
+    @property
+    def name(self) -> str:
+        return "ollama"
+
+    async def check_available(self) -> bool:
+        if self._available is not None:
+            return self._available
+
+        try:
+            ollama = get_ollama()
+            models = ollama.list()
+            model_names = [m.model for m in models.models] if hasattr(models, "models") else []
+
+            if not any(name and self.MODEL in name for name in model_names):
+                print(
+                    f"⚠️  Embedding model '{self.MODEL}' not found. Run: ollama pull {self.MODEL}",
+                    file=sys.stderr,
+                )
+                self._available = False
+                return False
+
+            self._available = True
+            return True
+        except Exception as e:
+            print(f"⚠️  Ollama not available: {e}. Start with: ollama serve", file=sys.stderr)
+            self._available = False
+            return False
+
+    async def get_embedding(self, text: str) -> list[float]:
+        ollama = get_ollama()
+        # nomic-embed-text has 8192 token context. Code can be 1-2 chars/token.
+        # Truncate to 2000 chars (~1000-2000 tokens) for larger safety margin
+        truncated = text[:2000] if len(text) > 2000 else text
+        response = ollama.embeddings(model=self.MODEL, prompt=truncated)
+        return response["embedding"]
+
+
+class GeminiProvider(BaseEmbeddingProvider):
+    """Gemini embedding provider using OAuth authentication."""
+
+    MODEL = "gemini-embedding-001"
+    DIMENSION = 768  # Using 768 for efficiency, can be up to 3072
+
+    def __init__(self):
+        self._available: bool | None = None
+        self._token_store = None
+
+    def _get_token_store(self):
+        if self._token_store is None:
+            from ..auth.token_store import TokenStore
+
+            self._token_store = TokenStore()
+        return self._token_store
+
+    @property
+    def dimension(self) -> int:
+        return self.DIMENSION
+
+    @property
+    def name(self) -> str:
+        return "gemini"
+
+    async def check_available(self) -> bool:
+        if self._available is not None:
+            return self._available
+
+        try:
+            token_store = self._get_token_store()
+            access_token = token_store.get_access_token("gemini")
+
+            if not access_token:
+                print(
+                    "⚠️  Gemini not authenticated. Run: stravinsky-auth login gemini",
+                    file=sys.stderr,
+                )
+                self._available = False
+                return False
+
+            self._available = True
+            return True
+        except Exception as e:
+            print(f"⚠️  Gemini not available: {e}", file=sys.stderr)
+            self._available = False
+            return False
+
+    async def get_embedding(self, text: str) -> list[float]:
+        import os
+
+        from ..auth.oauth import (
+            ANTIGRAVITY_DEFAULT_PROJECT_ID,
+            ANTIGRAVITY_ENDPOINTS,
+            ANTIGRAVITY_HEADERS,
+        )
+
+        token_store = self._get_token_store()
+        access_token = token_store.get_access_token("gemini")
+
+        if not access_token:
+            raise ValueError("Not authenticated with Gemini. Run: stravinsky-auth login gemini")
+
+        httpx = get_httpx()
+
+        # Use Antigravity endpoint for embeddings (same auth as invoke_gemini)
+        project_id = os.getenv("STRAVINSKY_ANTIGRAVITY_PROJECT_ID", ANTIGRAVITY_DEFAULT_PROJECT_ID)
+
+        headers = {
+            "Authorization": f"Bearer {access_token}",
+            "Content-Type": "application/json",
+            **ANTIGRAVITY_HEADERS,
+        }
+
+        # Wrap request for Antigravity API
+        import uuid
+
+        inner_payload = {
+            "model": f"models/{self.MODEL}",
+            "content": {"parts": [{"text": text}]},
+            "outputDimensionality": self.DIMENSION,
+        }
+
+        wrapped_payload = {
+            "project": project_id,
+            "model": self.MODEL,
+            "userAgent": "antigravity",
+            "requestId": f"embed-{uuid.uuid4()}",
+            "request": inner_payload,
+        }
+
+        # Try endpoints in order
+        last_error = None
+        async with httpx.AsyncClient(timeout=60.0) as client:
+            for endpoint in ANTIGRAVITY_ENDPOINTS:
+                api_url = f"{endpoint}/v1internal:embedContent"
+
+                try:
+                    response = await client.post(
+                        api_url,
+                        headers=headers,
+                        json=wrapped_payload,
+                    )
+
+                    if response.status_code in (401, 403):
+                        last_error = Exception(f"{response.status_code} from {endpoint}")
+                        continue
+
+                    response.raise_for_status()
+                    data = response.json()
+
+                    # Extract embedding from response
+                    inner_response = data.get("response", data)
+                    embedding = inner_response.get("embedding", {})
+                    values = embedding.get("values", [])
+
+                    if values:
+                        return values
+
+                    raise ValueError(f"No embedding values in response: {data}")
+
+                except Exception as e:
+                    last_error = e
+                    continue
+
+        raise ValueError(f"All Antigravity endpoints failed for embeddings: {last_error}")
+
+
+class OpenAIProvider(BaseEmbeddingProvider):
+    """OpenAI embedding provider using OAuth authentication."""
+
+    MODEL = "text-embedding-3-small"
+    DIMENSION = 1536
+
+    def __init__(self):
+        self._available: bool | None = None
+        self._token_store = None
+
+    def _get_token_store(self):
+        if self._token_store is None:
+            from ..auth.token_store import TokenStore
+
+            self._token_store = TokenStore()
+        return self._token_store
+
+    @property
+    def dimension(self) -> int:
+        return self.DIMENSION
+
+    @property
+    def name(self) -> str:
+        return "openai"
+
+    async def check_available(self) -> bool:
+        if self._available is not None:
+            return self._available
+
+        try:
+            token_store = self._get_token_store()
+            access_token = token_store.get_access_token("openai")
+
+            if not access_token:
+                print(
+                    "⚠️  OpenAI not authenticated. Run: stravinsky-auth login openai",
+                    file=sys.stderr,
+                )
+                self._available = False
+                return False
+
+            self._available = True
+            return True
+        except Exception as e:
+            print(f"⚠️  OpenAI not available: {e}", file=sys.stderr)
+            self._available = False
+            return False
+
+    async def get_embedding(self, text: str) -> list[float]:
+        token_store = self._get_token_store()
+        access_token = token_store.get_access_token("openai")
+
+        if not access_token:
+            raise ValueError("Not authenticated with OpenAI. Run: stravinsky-auth login openai")
+
+        httpx = get_httpx()
+
+        # Use standard OpenAI API for embeddings
+        api_url = "https://api.openai.com/v1/embeddings"
+
+        headers = {
+            "Authorization": f"Bearer {access_token}",
+            "Content-Type": "application/json",
+        }
+
+        payload = {
+            "model": self.MODEL,
+            "input": text,
+        }
+
+        async with httpx.AsyncClient(timeout=60.0) as client:
+            response = await client.post(api_url, headers=headers, json=payload)
+
+            if response.status_code == 401:
+                raise ValueError("OpenAI authentication failed. Run: stravinsky-auth login openai")
+
+            response.raise_for_status()
+            data = response.json()
+
+            # Extract embedding from response
+            embeddings = data.get("data", [])
+            if embeddings and "embedding" in embeddings[0]:
+                return embeddings[0]["embedding"]
+
+            raise ValueError(f"No embedding in response: {data}")
+
+
+class MxbaiProvider(BaseEmbeddingProvider):
+    """Ollama local embedding provider using mxbai-embed-large (better for code).
+
+    mxbai-embed-large is a 1024-dimensional model optimized for code understanding.
+    It generally outperforms nomic-embed-text on code-related retrieval tasks.
+    """
+
+    MODEL = "mxbai-embed-large"
+    DIMENSION = 1024
+
+    def __init__(self):
+        self._available: bool | None = None
+
+    @property
+    def dimension(self) -> int:
+        return self.DIMENSION
+
+    @property
+    def name(self) -> str:
+        return "mxbai"
+
+    async def check_available(self) -> bool:
+        if self._available is not None:
+            return self._available
+
+        try:
+            ollama = get_ollama()
+            models = ollama.list()
+            model_names = [m.model for m in models.models] if hasattr(models, "models") else []
+
+            if not any(name and self.MODEL in name for name in model_names):
+                print(
+                    f"⚠️  Embedding model '{self.MODEL}' not found. Run: ollama pull {self.MODEL}",
+                    file=sys.stderr,
+                )
+                self._available = False
+                return False
+
+            self._available = True
+            return True
+        except Exception as e:
+            print(f"⚠️  Ollama not available: {e}. Start with: ollama serve", file=sys.stderr)
+            self._available = False
+            return False
+
+    async def get_embedding(self, text: str) -> list[float]:
+        ollama = get_ollama()
+        # mxbai-embed-large has 512 token context. Code can be 1-2 chars/token.
+        # Truncate to 2000 chars (~1000-2000 tokens) for safety margin
+        truncated = text[:2000] if len(text) > 2000 else text
+        response = ollama.embeddings(model=self.MODEL, prompt=truncated)
+        return response["embedding"]
+
+
+class HuggingFaceProvider(BaseEmbeddingProvider):
+    """Hugging Face Inference API embedding provider.
+
+    Uses the Hugging Face Inference API for embeddings. Requires HF_TOKEN from:
+    1. Environment variable: HF_TOKEN or HUGGING_FACE_HUB_TOKEN
+    2. HF CLI config: ~/.cache/huggingface/token or ~/.huggingface/token
+
+    Default model: sentence-transformers/all-mpnet-base-v2 (768 dims, high quality)
+    """
+
+    DEFAULT_MODEL = "sentence-transformers/all-mpnet-base-v2"
+    DEFAULT_DIMENSION = 768
+
+    def __init__(self, model: str | None = None):
+        self._available: bool | None = None
+        self._model = model or self.DEFAULT_MODEL
+        # Dimension varies by model, but we'll use default for common models
+        self._dimension = self.DEFAULT_DIMENSION
+        self._token: str | None = None
+
+    @property
+    def dimension(self) -> int:
+        return self._dimension
+
+    @property
+    def name(self) -> str:
+        return "huggingface"
+
+    def _get_hf_token(self) -> str | None:
+        """Discover HF token from environment or CLI config."""
+        import os
+
+        # Check environment variables first
+        token = os.getenv("HF_TOKEN") or os.getenv("HUGGING_FACE_HUB_TOKEN")
+        if token:
+            return token
+
+        # Check HF CLI config locations
+        hf_token_paths = [
+            Path.home() / ".cache" / "huggingface" / "token",
+            Path.home() / ".huggingface" / "token",
+        ]
+
+        for token_path in hf_token_paths:
+            if token_path.exists():
+                try:
+                    return token_path.read_text().strip()
+                except Exception:
+                    continue
+
+        return None
+
+    async def check_available(self) -> bool:
+        if self._available is not None:
+            return self._available
+
+        try:
+            self._token = self._get_hf_token()
+            if not self._token:
+                print(
+                    "⚠️  Hugging Face token not found. Run: huggingface-cli login or set HF_TOKEN env var",
+                    file=sys.stderr,
+                )
+                self._available = False
+                return False
+
+            self._available = True
+            return True
+        except Exception as e:
+            print(f"⚠️  Hugging Face not available: {e}", file=sys.stderr)
+            self._available = False
+            return False
+
+    @retry(
+        stop=stop_after_attempt(3),
+        wait=wait_exponential(multiplier=1, min=2, max=10),
+        retry=retry_if_exception_type(httpx.HTTPStatusError),
+    )
+    async def get_embedding(self, text: str) -> list[float]:
+        """Get embedding from HF Inference API with retry logic."""
+        if not self._token:
+            self._token = self._get_hf_token()
+            if not self._token:
+                raise ValueError(
+                    "Hugging Face token not found. Run: huggingface-cli login or set HF_TOKEN"
+                )
+
+        httpx_client = get_httpx()
+
+        # HF Serverless Inference API endpoint
+        # Note: Free tier may have limited availability for some models
+        api_url = f"https://api-inference.huggingface.co/pipeline/feature-extraction/{self._model}"
+
+        headers = {
+            "Authorization": f"Bearer {self._token}",
+        }
+
+        # Truncate text to reasonable length (most models have 512 token limit)
+        # ~2000 chars ≈ 500 tokens for safety
+        truncated = text[:2000] if len(text) > 2000 else text
+
+        # HF Inference API accepts raw JSON with inputs field
+        payload = {"inputs": [truncated], "options": {"wait_for_model": True}}
+
+        async with httpx_client.AsyncClient(timeout=60.0) as client:
+            response = await client.post(api_url, headers=headers, json=payload)
+
+            # Handle specific error codes
+            if response.status_code == 401:
+                raise ValueError(
+                    "Hugging Face authentication failed. Run: huggingface-cli login or set HF_TOKEN"
+                )
+            elif response.status_code == 410:
+                # Model removed from free tier
+                raise ValueError(
+                    f"Model {self._model} is no longer available on HF free Inference API (410 Gone). "
+                    "Try a different model or use Ollama for local embeddings instead."
+                )
+            elif response.status_code == 503:
+                # Model loading - retry will handle this
+                logger.info(f"Model {self._model} is loading, retrying...")
+                response.raise_for_status()
+            elif response.status_code == 429:
+                # Rate limit - retry will handle with exponential backoff
+                logger.warning("HF API rate limit hit, retrying with backoff...")
+                response.raise_for_status()
+
+            response.raise_for_status()
+
+            # Response is a single embedding vector (list of floats)
+            embedding = response.json()
+
+            # Handle different response formats
+            if isinstance(embedding, list):
+                # Direct embedding or batch with single item
+                if isinstance(embedding[0], (int, float)):
+                    return embedding
+                elif isinstance(embedding[0], list):
+                    # Batch response with single embedding
+                    return embedding[0]
+
+            raise ValueError(f"Unexpected response format from HF API: {type(embedding)}")
+
+    async def embed_batch(self, texts: list[str]) -> list[list[float]]:
+        """Batch embedding support for HF API.
+
+        HF API supports batch requests, so we can send multiple texts at once.
+        """
+        if not texts:
+            return []
+
+        if not self._token:
+            self._token = self._get_hf_token()
+            if not self._token:
+                raise ValueError(
+                    "Hugging Face token not found. Run: huggingface-cli login or set HF_TOKEN"
+                )
+
+        httpx_client = get_httpx()
+
+        # HF Serverless Inference API endpoint
+        api_url = f"https://api-inference.huggingface.co/pipeline/feature-extraction/{self._model}"
+
+        headers = {
+            "Authorization": f"Bearer {self._token}",
+        }
+
+        # Truncate all texts
+        truncated_texts = [text[:2000] if len(text) > 2000 else text for text in texts]
+
+        payload = {"inputs": truncated_texts, "options": {"wait_for_model": True}}
+
+        async with httpx_client.AsyncClient(timeout=120.0) as client:
+            response = await client.post(api_url, headers=headers, json=payload)
+
+            if response.status_code == 401:
+                raise ValueError(
+                    "Hugging Face authentication failed. Run: huggingface-cli login or set HF_TOKEN"
+                )
+
+            response.raise_for_status()
+
+            embeddings = response.json()
+
+            # Response should be a list of embeddings
+            if isinstance(embeddings, list) and all(isinstance(e, list) for e in embeddings):
+                return embeddings
+
+            raise ValueError(f"Unexpected batch response format from HF API: {type(embeddings)}")
+
+
+# Embedding provider instance cache
+_embedding_provider_cache: dict[str, BaseEmbeddingProvider] = {}
+_embedding_provider_lock = threading.Lock()
+
+
+def get_embedding_provider(provider: EmbeddingProvider) -> BaseEmbeddingProvider:
+    """Factory function to get an embedding provider instance with caching."""
+    if provider not in _embedding_provider_cache:
+        with _embedding_provider_lock:
+            # Double-check pattern to avoid race condition
+            if provider not in _embedding_provider_cache:
+                providers = {
+                    "ollama": OllamaProvider,
+                    "mxbai": MxbaiProvider,
+                    "gemini": GeminiProvider,
+                    "openai": OpenAIProvider,
+                    "huggingface": HuggingFaceProvider,
+                }
+
+                if provider not in providers:
+                    raise ValueError(
+                        f"Unknown provider: {provider}. Available: {list(providers.keys())}"
+                    )
+
+                _embedding_provider_cache[provider] = providers[provider]()
+
+    return _embedding_provider_cache[provider]
+
+
+class CodebaseVectorStore:
+    """
+    Persistent vector store for a single codebase.
+
+    Storage: ~/.stravinsky/vectordb/<project_hash>_<provider>/
+    Embedding: Configurable via provider (ollama, gemini, openai)
+    """
+
+    CHUNK_SIZE = 50  # lines per chunk
+    CHUNK_OVERLAP = 10  # lines of overlap between chunks
+
+    # File patterns to index
+    CODE_EXTENSIONS = {
+        ".py",
+        ".js",
+        ".ts",
+        ".tsx",
+        ".jsx",
+        ".go",
+        ".rs",
+        ".rb",
+        ".java",
+        ".c",
+        ".cpp",
+        ".h",
+        ".hpp",
+        ".cs",
+        ".swift",
+        ".kt",
+        ".scala",
+        ".vue",
+        ".svelte",
+        ".md",
+        ".txt",
+        ".yaml",
+        ".yml",
+        ".json",
+        ".toml",
+    }
+
+    # Directories to skip
+    SKIP_DIRS = {
+        "node_modules",
+        ".git",
+        "__pycache__",
+        ".venv",
+        "venv",
+        "env",
+        "dist",
+        "build",
+        ".next",
+        ".nuxt",
+        "target",
+        ".tox",
+        ".pytest_cache",
+        ".mypy_cache",
+        ".ruff_cache",
+        "coverage",
+        ".stravinsky",
+        "scratches",
+        "consoles",
+        ".idea",
+    }
+
+    def __init__(self, project_path: str, provider: EmbeddingProvider = "ollama"):
+        self.project_path = Path(project_path).resolve()
+        self.project_hash = hashlib.md5(str(self.project_path).encode()).hexdigest()[:12]
+
+        # Initialize embedding provider
+        self.provider_name = provider
+        self.provider = get_embedding_provider(provider)
+
+        # Store in user's home directory, separate by provider to avoid dimension mismatch
+        self.db_path = Path.home() / ".stravinsky" / "vectordb" / f"{self.project_hash}_{provider}"
+        self.db_path.mkdir(parents=True, exist_ok=True)
+
+        # File lock for single-process access to ChromaDB (prevents corruption)
+        self._lock_path = self.db_path / ".chromadb.lock"
+        self._file_lock = None
+
+        self._client = None
+        self._collection = None
+
+        # File watcher attributes
+        self._watcher: CodebaseFileWatcher | None = None
+        self._watcher_lock = threading.Lock()
+
+        # Cancellation flag for indexing operations
+        self._cancel_indexing = False
+        self._cancel_lock = threading.Lock()
+
+    @property
+    def file_lock(self):
+        """Get or create the file lock for this database.
+
+        Uses filelock to ensure single-process access to ChromaDB,
+        preventing database corruption from concurrent writes.
+        """
+        if self._file_lock is None:
+            filelock = get_filelock()
+            # Timeout of 30 seconds - if lock can't be acquired, raise error
+            self._file_lock = filelock.FileLock(str(self._lock_path), timeout=30)
+        return self._file_lock
+
+    @property
+    def client(self):
+        if self._client is None:
+            chromadb = get_chromadb()
+
+            # Check for stale lock before attempting acquisition
+            # Prevents 30s timeout from dead processes causing MCP "Connection closed" errors
+            if self._lock_path.exists():
+                import time
+
+                lock_age = time.time() - self._lock_path.stat().st_mtime
+                # Lock older than 60 seconds is likely from a crashed process
+                # (Reduced from 300s to catch recently crashed processes)
+                if lock_age > 60:
+                    logger.warning(
+                        f"Removing stale ChromaDB lock (age: {lock_age:.0f}s, path: {self._lock_path})"
+                    )
+                    try:
+                        self._lock_path.unlink(missing_ok=True)
+                    except Exception as e:
+                        logger.warning(f"Could not remove stale lock: {e}")
+
+            # Acquire lock before creating client to prevent concurrent access
+            try:
+                with self.file_lock:  # Auto-releases on exit
+                    logger.debug(f"Acquired ChromaDB lock for {self.db_path}")
+                    self._client = chromadb.PersistentClient(path=str(self.db_path))
+            except Exception as e:
+                logger.warning(f"Could not acquire ChromaDB lock: {e}. Proceeding without lock.")
+                self._client = chromadb.PersistentClient(path=str(self.db_path))
+        return self._client
+
+    @property
+    def collection(self):
+        if self._collection is None:
+            self._collection = self.client.get_or_create_collection(
+                name="codebase", metadata={"hnsw:space": "cosine"}
+            )
+        return self._collection
+
+    async def check_embedding_service(self) -> bool:
+        """Check if the embedding provider is available."""
+        return await self.provider.check_available()
+
+    async def get_embedding(self, text: str) -> list[float]:
+        """Get embedding vector for text using the configured provider."""
+        return await self.provider.get_embedding(text)
+
+    async def get_embeddings_batch(
+        self, texts: list[str], max_concurrent: int = 10
+    ) -> list[list[float]]:
+        """Get embeddings for multiple texts with parallel execution.
+
+        Uses asyncio.gather with semaphore-based concurrency control to avoid
+        overwhelming the embedding service while maximizing throughput.
+
+        Args:
+            texts: List of text strings to embed
+            max_concurrent: Maximum concurrent embedding requests (default: 10)
+
+        Returns:
+            List of embedding vectors in the same order as input texts.
+        """
+        import asyncio
+
+        if not texts:
+            return []
+
+        # Use semaphore to limit concurrent requests
+        semaphore = asyncio.Semaphore(max_concurrent)
+
+        async def get_with_semaphore(text: str, index: int) -> tuple[int, list[float]]:
+            async with semaphore:
+                emb = await self.get_embedding(text)
+                return (index, emb)
+
+        # Launch all embedding requests concurrently (respecting semaphore)
+        tasks = [get_with_semaphore(text, i) for i, text in enumerate(texts)]
+        results = await asyncio.gather(*tasks)
+
+        # Sort by original index to maintain order
+        sorted_results = sorted(results, key=lambda x: x[0])
+        return [emb for _, emb in sorted_results]
+
+    def _chunk_file(self, file_path: Path) -> list[dict]:
+        """Split a file into chunks with metadata.
+
+        Uses AST-aware chunking for Python files to respect function/class
+        boundaries. Falls back to line-based chunking for other languages.
+        """
+        try:
+            content = file_path.read_text(encoding="utf-8", errors="ignore")
+        except Exception:
+            return []
+
+        lines = content.split("\n")
+        if len(lines) < 5:  # Skip very small files
+            return []
+
+        rel_path = str(file_path.resolve().relative_to(self.project_path.resolve()))
+        language = file_path.suffix.lstrip(".")
+
+        # Use AST-aware chunking for Python files
+        if language == "py":
+            chunks = self._chunk_python_ast(content, rel_path, language)
+            if chunks:  # If AST parsing succeeded
+                return chunks
+
+        # Fallback: line-based chunking for other languages or if AST fails
+        return self._chunk_by_lines(lines, rel_path, language)
+
+    def _chunk_python_ast(self, content: str, rel_path: str, language: str) -> list[dict]:
+        """Parse Python file and create chunks based on function/class boundaries.
+
+        Each function, method, and class becomes its own chunk, preserving
+        semantic boundaries for better embedding quality.
+        """
+        import ast
+
+        try:
+            tree = ast.parse(content)
+        except SyntaxError:
+            return []  # Fall back to line-based chunking
+
+        lines = content.split("\n")
+        chunks = []
+
+        def get_docstring(node: ast.AST) -> str:
+            """Extract docstring from a node if present."""
+            if (
+                isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef, ast.ClassDef))
+                and node.body
+            ):
+                first = node.body[0]
+                if isinstance(first, ast.Expr) and isinstance(first.value, ast.Constant):
+                    if isinstance(first.value.value, str):
+                        return first.value.value
+            return ""
+
+        def get_decorators(
+            node: ast.FunctionDef | ast.AsyncFunctionDef | ast.ClassDef,
+        ) -> list[str]:
+            """Extract decorator names from a node."""
+            decorators = []
+            for dec in node.decorator_list:
+                if isinstance(dec, ast.Name):
+                    decorators.append(f"@{dec.id}")
+                elif isinstance(dec, ast.Attribute):
+                    decorators.append(f"@{ast.unparse(dec)}")
+                elif isinstance(dec, ast.Call):
+                    if isinstance(dec.func, ast.Name):
+                        decorators.append(f"@{dec.func.id}")
+                    elif isinstance(dec.func, ast.Attribute):
+                        decorators.append(f"@{ast.unparse(dec.func)}")
+            return decorators
+
+        def get_base_classes(node: ast.ClassDef) -> list[str]:
+            """Extract base class names from a class definition."""
+            bases = []
+            for base in node.bases:
+                if isinstance(base, ast.Name):
+                    bases.append(base.id)
+                elif isinstance(base, ast.Attribute):
+                    bases.append(ast.unparse(base))
+                else:
+                    bases.append(ast.unparse(base))
+            return bases
+
+        def get_return_type(node: ast.FunctionDef | ast.AsyncFunctionDef) -> str:
+            """Extract return type annotation from a function."""
+            if node.returns:
+                return ast.unparse(node.returns)
+            return ""
+
+        def get_parameters(node: ast.FunctionDef | ast.AsyncFunctionDef) -> list[str]:
+            """Extract parameter signatures from a function."""
+            params = []
+            for arg in node.args.args:
+                param = arg.arg
+                if arg.annotation:
+                    param += f": {ast.unparse(arg.annotation)}"
+                params.append(param)
+            return params
+
+        def add_chunk(
+            node: ast.FunctionDef | ast.AsyncFunctionDef | ast.ClassDef,
+            node_type: str,
+            name: str,
+            parent_class: str | None = None,
+        ) -> None:
+            """Add a chunk for a function/class node."""
+            start_line = node.lineno
+            end_line = node.end_lineno or start_line
+
+            # Extract the source code for this node
+            chunk_lines = lines[start_line - 1 : end_line]
+            chunk_text = "\n".join(chunk_lines)
+            content_hash = hashlib.md5(chunk_text.encode("utf-8")).hexdigest()[:12]
+
+            # Skip very small chunks
+            if len(chunk_lines) < 3:
+                return
+
+            # Build descriptive header
+            docstring = get_docstring(node)
+            if parent_class:
+                header = f"File: {rel_path}\n{node_type}: {parent_class}.{name}\nLines: {start_line}-{end_line}"
+            else:
+                header = f"File: {rel_path}\n{node_type}: {name}\nLines: {start_line}-{end_line}"
+
+            if docstring:
+                header += f"\nDocstring: {docstring[:200]}..."
+
+            document = f"{header}\n\n{chunk_text}"
+
+            chunks.append(
+                {
+                    "id": f"{rel_path}:{start_line}-{end_line}:{content_hash}",
+                    "document": document,
+                    "metadata": {
+                        "file_path": rel_path,
+                        "start_line": start_line,
+                        "end_line": end_line,
+                        "language": language,
+                        "node_type": node_type.lower(),
+                        "name": f"{parent_class}.{name}" if parent_class else name,
+                        # Structural metadata for filtering
+                        "decorators": ",".join(get_decorators(node)),
+                        "is_async": isinstance(node, ast.AsyncFunctionDef),
+                        # Class-specific metadata
+                        "base_classes": ",".join(get_base_classes(node))
+                        if isinstance(node, ast.ClassDef)
+                        else "",
+                        # Function-specific metadata
+                        "return_type": get_return_type(node)
+                        if isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef))
+                        else "",
+                        "parameters": ",".join(get_parameters(node))
+                        if isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef))
+                        else "",
+                    },
+                }
+            )
+
+        # Walk the AST and extract functions/classes
+        for node in ast.walk(tree):
+            if isinstance(node, ast.ClassDef):
+                add_chunk(node, "Class", node.name)
+                # Also add methods as separate chunks for granular search
+                for item in node.body:
+                    if isinstance(item, (ast.FunctionDef, ast.AsyncFunctionDef)):
+                        add_chunk(item, "Method", item.name, parent_class=node.name)
+            elif isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef)):
+                # Only top-level functions (not methods)
+                # Check if this function is inside a class body
+                is_method = False
+                for parent in ast.walk(tree):
+                    if isinstance(parent, ast.ClassDef):
+                        body = getattr(parent, "body", None)
+                        if isinstance(body, list) and node in body:
+                            is_method = True
+                            break
+                if not is_method:
+                    add_chunk(node, "Function", node.name)
+
+        # If we found no functions/classes, chunk module-level code
+        if not chunks and len(lines) >= 5:
+            # Add module-level chunk for imports and constants
+            module_chunk = "\n".join(lines[: min(50, len(lines))])
+            chunks.append(
+                {
+                    "id": f"{rel_path}:1-{min(50, len(lines))}",
+                    "document": f"File: {rel_path}\nModule-level code\nLines: 1-{min(50, len(lines))}\n\n{module_chunk}",
+                    "metadata": {
+                        "file_path": rel_path,
+                        "start_line": 1,
+                        "end_line": min(50, len(lines)),
+                        "language": language,
+                        "node_type": "module",
+                        "name": rel_path,
+                    },
+                }
+            )
+
+        return chunks
+
+    def _chunk_by_lines(self, lines: list[str], rel_path: str, language: str) -> list[dict]:
+        """Fallback line-based chunking with overlap."""
+        chunks = []
+
+        for i in range(0, len(lines), self.CHUNK_SIZE - self.CHUNK_OVERLAP):
+            chunk_lines = lines[i : i + self.CHUNK_SIZE]
+            if len(chunk_lines) < 5:  # Skip tiny trailing chunks
+                continue
+
+            chunk_text = "\n".join(chunk_lines)
+            content_hash = hashlib.md5(chunk_text.encode("utf-8")).hexdigest()[:12]
+            start_line = i + 1
+            end_line = i + len(chunk_lines)
+
+            # Create a searchable document with context
+            document = f"File: {rel_path}\nLines: {start_line}-{end_line}\n\n{chunk_text}"
+
+            chunks.append(
+                {
+                    "id": f"{rel_path}:{start_line}-{end_line}:{content_hash}",
+                    "document": document,
+                    "metadata": {
+                        "file_path": rel_path,
+                        "start_line": start_line,
+                        "end_line": end_line,
+                        "language": language,
+                    },
+                }
+            )
+
+        return chunks
+
+    def _get_files_to_index(self) -> list[Path]:
+        """Get all indexable files in the project."""
+        files = []
+        for file_path in self.project_path.rglob("*"):
+            if file_path.is_file():
+                # Skip files outside project boundaries (symlink traversal protection)
+                try:
+                    resolved_file = file_path.resolve()
+                    resolved_project = self.project_path.resolve()
+
+                    # Check if file is under project using parent chain with samefile()
+                    # This handles macOS /var → /private/var aliasing and symlinks
+                    found = False
+                    current = resolved_file.parent
+                    while current != current.parent:  # Stop at filesystem root
+                        try:
+                            if current.samefile(resolved_project):
+                                found = True
+                                break
+                        except OSError:
+                            # samefile can fail on some filesystems; try string comparison
+                            if current == resolved_project:
+                                found = True
+                                break
+                        current = current.parent
+
+                    if not found:
+                        continue  # Outside project
+                except (ValueError, OSError):
+                    continue  # Outside project boundaries
+
+                # Skip hidden files and directories
+                if any(
+                    part.startswith(".") for part in file_path.parts[len(self.project_path.parts) :]
+                ) and file_path.suffix not in {".md", ".txt"}:  # Allow .github docs
+                    continue
+
+                # Skip excluded directories
+                if any(skip_dir in file_path.parts for skip_dir in self.SKIP_DIRS):
+                    continue
+
+                # Only include code files
+                if file_path.suffix.lower() in self.CODE_EXTENSIONS:
+                    files.append(file_path)
+
+        return files
+
+    def request_cancel_indexing(self) -> None:
+        """Request cancellation of ongoing indexing operation.
+
+        Sets a flag that will be checked between batches. The operation will
+        stop gracefully after completing the current batch.
+        """
+        with self._cancel_lock:
+            self._cancel_indexing = True
+            logger.info(f"Cancellation requested for {self.project_path}")
+
+    def clear_cancel_flag(self) -> None:
+        """Clear the cancellation flag."""
+        with self._cancel_lock:
+            self._cancel_indexing = False
+
+    def is_cancellation_requested(self) -> bool:
+        """Check if cancellation has been requested."""
+        with self._cancel_lock:
+            return self._cancel_indexing
+
+    async def index_codebase(self, force: bool = False) -> dict:
+        """
+        Index the entire codebase into the vector store.
+
+        This operation can be cancelled by calling request_cancel_indexing().
+        Cancellation happens between batches, so the current batch will complete.
+
+        Args:
+            force: If True, reindex everything. Otherwise, only index new/changed files.
+
+        Returns:
+            Statistics about the indexing operation.
+        """
+        import time
+
+        # Clear any previous cancellation requests
+        self.clear_cancel_flag()
+
+        # Start timing
+        start_time = time.time()
+
+        print(f"🔍 SEMANTIC-INDEX: {self.project_path}", file=sys.stderr)
+
+        # Notify reindex start (non-blocking)
+        notifier = None  # Initialize to avoid NameError in error handlers
+        try:
+            from mcp_bridge.notifications import get_notification_manager
+
+            notifier = get_notification_manager()
+            await notifier.notify_reindex_start(str(self.project_path))
+        except Exception as e:
+            logger.warning(f"Failed to send reindex start notification: {e}")
+
+        try:
+            if not await self.check_embedding_service():
+                error_msg = "Embedding service not available"
+                # Notify error
+                try:
+                    if notifier:
+                        await notifier.notify_reindex_error(error_msg)
+                except Exception as e:
+                    logger.warning(f"Failed to send reindex error notification: {e}")
+                return {"error": error_msg, "indexed": 0}
+
+            # Get existing document IDs
+            existing_ids = set()
+            try:
+                # Only fetch IDs to minimize overhead
+                existing = self.collection.get(include=[])
+                existing_ids = set(existing["ids"]) if existing["ids"] else set()
+            except Exception:
+                pass
+
+            if force:
+                # Clear existing collection
+                try:
+                    self.client.delete_collection("codebase")
+                    self._collection = None
+                    existing_ids = set()
+                except Exception:
+                    pass
+
+            files = self._get_files_to_index()
+            all_chunks = []
+            current_chunk_ids = set()
+
+            # Mark: Generate all chunks for current codebase
+            for file_path in files:
+                chunks = self._chunk_file(file_path)
+                all_chunks.extend(chunks)
+                for c in chunks:
+                    current_chunk_ids.add(c["id"])
+
+            # Sweep: Identify stale chunks to remove
+            to_delete = existing_ids - current_chunk_ids
+
+            # Identify new chunks to add
+            to_add_ids = current_chunk_ids - existing_ids
+            chunks_to_add = [c for c in all_chunks if c["id"] in to_add_ids]
+
+            # Prune stale chunks
+            if to_delete:
+                print(f"  Pruning {len(to_delete)} stale chunks...", file=sys.stderr)
+                self.collection.delete(ids=list(to_delete))
+
+            if not chunks_to_add:
+                stats = {
+                    "indexed": 0,
+                    "pruned": len(to_delete),
+                    "total_files": len(files),
+                    "message": "No new chunks to index",
+                    "time_taken": round(time.time() - start_time, 1),
+                }
+                # Notify completion
+                try:
+                    if notifier:
+                        await notifier.notify_reindex_complete(stats)
+                except Exception as e:
+                    logger.warning(f"Failed to send reindex complete notification: {e}")
+                return stats
+
+            # Batch embed and store
+            batch_size = 50
+            total_indexed = 0
+
+            for i in range(0, len(chunks_to_add), batch_size):
+                # Check for cancellation between batches
+                if self.is_cancellation_requested():
+                    print(f"  ⚠️  Indexing cancelled after {total_indexed} chunks", file=sys.stderr)
+                    stats = {
+                        "indexed": total_indexed,
+                        "pruned": len(to_delete),
+                        "total_files": len(files),
+                        "db_path": str(self.db_path),
+                        "time_taken": round(time.time() - start_time, 1),
+                        "cancelled": True,
+                        "message": f"Cancelled after {total_indexed}/{len(chunks_to_add)} chunks",
+                    }
+                    # Notify cancellation
+                    try:
+                        if notifier:
+                            await notifier.notify_reindex_error(
+                                f"Indexing cancelled by user after {total_indexed} chunks"
+                            )
+                    except Exception as e:
+                        logger.warning(f"Failed to send cancellation notification: {e}")
+                    return stats
+
+                batch = chunks_to_add[i : i + batch_size]
+
+                documents = [c["document"] for c in batch]
+                embeddings = await self.get_embeddings_batch(documents)
+
+                self.collection.add(
+                    ids=[c["id"] for c in batch],
+                    documents=documents,
+                    embeddings=embeddings,  # type: ignore[arg-type]
+                    metadatas=[c["metadata"] for c in batch],
+                )
+                total_indexed += len(batch)
+                print(f"  Indexed {total_indexed}/{len(chunks_to_add)} chunks...", file=sys.stderr)
+
+            stats = {
+                "indexed": total_indexed,
+                "pruned": len(to_delete),
+                "total_files": len(files),
+                "db_path": str(self.db_path),
+                "time_taken": round(time.time() - start_time, 1),
+            }
+
+            # Notify completion
+            try:
+                if notifier:
+                    await notifier.notify_reindex_complete(stats)
+            except Exception as e:
+                logger.warning(f"Failed to send reindex complete notification: {e}")
+
+            return stats
+
+        except Exception as e:
+            error_msg = str(e)
+            logger.error(f"Reindexing failed: {error_msg}")
+
+            # Notify error
+            try:
+                if notifier:
+                    await notifier.notify_reindex_error(error_msg)
+            except Exception as notify_error:
+                logger.warning(f"Failed to send reindex error notification: {notify_error}")
+
+            raise
+
+    async def search(
+        self,
+        query: str,
+        n_results: int = 10,
+        language: str | None = None,
+        node_type: str | None = None,
+        decorator: str | None = None,
+        is_async: bool | None = None,
+        base_class: str | None = None,
+    ) -> list[dict]:
+        """
+        Search the codebase with a natural language query.
+
+        Args:
+            query: Natural language search query
+            n_results: Maximum number of results to return
+            language: Filter by language (e.g., "py", "ts", "js")
+            node_type: Filter by node type (e.g., "function", "class", "method")
+            decorator: Filter by decorator (e.g., "@property", "@staticmethod")
+            is_async: Filter by async status (True = async only, False = sync only)
+            base_class: Filter by base class (e.g., "BaseClass")
+
+        Returns:
+            List of matching code chunks with metadata.
+        """
+        filters = []
+        if language:
+            filters.append(f"language={language}")
+        if node_type:
+            filters.append(f"node_type={node_type}")
+        if decorator:
+            filters.append(f"decorator={decorator}")
+        if is_async is not None:
+            filters.append(f"is_async={is_async}")
+        if base_class:
+            filters.append(f"base_class={base_class}")
+        filter_str = f" [{', '.join(filters)}]" if filters else ""
+        print(f"🔎 SEMANTIC-SEARCH: '{query[:50]}...'{filter_str}", file=sys.stderr)
+
+        if not await self.check_embedding_service():
+            return [{"error": "Embedding service not available"}]
+
+        # Check if collection has documents
+        try:
+            count = self.collection.count()
+            if count == 0:
+                return [{"error": "No documents indexed", "hint": "Run index_codebase first"}]
+        except Exception as e:
+            return [{"error": f"Collection error: {e}"}]
+
+        # Get query embedding
+        query_embedding = await self.get_embedding(query)
+
+        # Build where clause for metadata filtering
+        where_filters = []
+        if language:
+            where_filters.append({"language": language})
+        if node_type:
+            where_filters.append({"node_type": node_type.lower()})
+        if decorator:
+            # ChromaDB $like for substring match in comma-separated field
+            # Use % wildcards for pattern matching
+            where_filters.append({"decorators": {"$like": f"%{decorator}%"}})
+        if is_async is not None:
+            where_filters.append({"is_async": is_async})
+        if base_class:
+            # Use $like for substring match
+            where_filters.append({"base_classes": {"$like": f"%{base_class}%"}})
+
+        where_clause = None
+        if len(where_filters) == 1:
+            where_clause = where_filters[0]
+        elif len(where_filters) > 1:
+            where_clause = {"$and": where_filters}
+
+        # Search with optional filtering
+        query_kwargs: dict = {
+            "query_embeddings": [query_embedding],
+            "n_results": n_results,
+            "include": ["documents", "metadatas", "distances"],
+        }
+        if where_clause:
+            query_kwargs["where"] = where_clause
+
+        results = self.collection.query(**query_kwargs)
+
+        # Format results
+        formatted = []
+        if results["ids"] and results["ids"][0]:
+            for i, _doc_id in enumerate(results["ids"][0]):
+                metadata = results["metadatas"][0][i] if results["metadatas"] else {}
+                distance = results["distances"][0][i] if results["distances"] else 0
+                document = results["documents"][0][i] if results["documents"] else ""
+
+                # Extract just the code part (skip file/line header)
+                code_lines = document.split("\n\n", 1)
+                code = code_lines[1] if len(code_lines) > 1 else document
+
+                formatted.append(
+                    {
+                        "file": metadata.get("file_path", "unknown"),
+                        "lines": f"{metadata.get('start_line', '?')}-{metadata.get('end_line', '?')}",
+                        "language": metadata.get("language", ""),
+                        "relevance": round(1 - distance, 3),  # Convert distance to similarity
+                        "code_preview": code[:500] + "..." if len(code) > 500 else code,
+                    }
+                )
+
+        return formatted
+
+    def get_stats(self) -> dict:
+        """Get statistics about the vector store."""
+        try:
+            count = self.collection.count()
+            return {
+                "project_path": str(self.project_path),
+                "db_path": str(self.db_path),
+                "chunks_indexed": count,
+                "embedding_provider": self.provider.name,
+                "embedding_dimension": self.provider.dimension,
+            }
+        except Exception as e:
+            return {"error": str(e)}
+
+    def start_watching(self, debounce_seconds: float = 2.0) -> "CodebaseFileWatcher":
+        """Start watching the project directory for file changes.
+
+        Args:
+            debounce_seconds: Time to wait before reindexing after changes (default: 2.0s)
+
+        Returns:
+            The CodebaseFileWatcher instance
+        """
+        with self._watcher_lock:
+            if self._watcher is None:
+                # Avoid circular import by importing here
+                self._watcher = CodebaseFileWatcher(
+                    project_path=self.project_path,
+                    store=self,
+                    debounce_seconds=debounce_seconds,
+                )
+                self._watcher.start()
+            else:
+                if not self._watcher.is_running():
+                    self._watcher.start()
+                else:
+                    logger.warning(f"Watcher for {self.project_path} is already running")
+            return self._watcher
+
+    def stop_watching(self) -> bool:
+        """Stop watching the project directory.
+
+        Returns:
+            True if watcher was stopped, False if no watcher was active
+        """
+        with self._watcher_lock:
+            if self._watcher is not None:
+                self._watcher.stop()
+                self._watcher = None
+                return True
+            return False
+
+    def is_watching(self) -> bool:
+        """Check if the project directory is being watched.
+
+        Returns:
+            True if watcher is active and running, False otherwise
+        """
+        with self._watcher_lock:
+            if self._watcher is not None:
+                return self._watcher.is_running()
+            return False
+
+
+# --- Module-level API for MCP tools ---
+
+_stores: dict[str, CodebaseVectorStore] = {}
+_stores_lock = threading.Lock()
+
+# Module-level watcher management
+_watchers: dict[str, "CodebaseFileWatcher"] = {}
+_watchers_lock = threading.Lock()
+
+
+def _cleanup_watchers():
+    """Cleanup function to stop all watchers on exit.
+
+    Registered with atexit to ensure graceful shutdown when Python exits normally.
+    Note: This won't be called if the process is killed (SIGKILL) or crashes.
+    """
+    with _watchers_lock:
+        for path, watcher in list(_watchers.items()):
+            try:
+                logger.debug(f"Stopping watcher for {path} on exit")
+                watcher.stop()
+            except Exception as e:
+                logger.warning(f"Error stopping watcher for {path}: {e}")
+
+
+# Register cleanup handler for graceful shutdown
+atexit.register(_cleanup_watchers)
+
+
+def get_store(project_path: str, provider: EmbeddingProvider = "ollama") -> CodebaseVectorStore:
+    """Get or create a vector store for a project.
+
+    Note: Cache key includes provider to prevent cross-provider conflicts
+    (different providers have different embedding dimensions).
+    """
+    path = str(Path(project_path).resolve())
+    cache_key = f"{path}:{provider}"
+    if cache_key not in _stores:
+        with _stores_lock:
+            # Double-check pattern to avoid race condition
+            if cache_key not in _stores:
+                _stores[cache_key] = CodebaseVectorStore(path, provider)
+    return _stores[cache_key]
+
+
+async def semantic_search(
+    query: str,
+    project_path: str = ".",
+    n_results: int = 10,
+    language: str | None = None,
+    node_type: str | None = None,
+    decorator: str | None = None,
+    is_async: bool | None = None,
+    base_class: str | None = None,
+    provider: EmbeddingProvider = "ollama",
+) -> str:
+    """
+    Search codebase with natural language query.
+
+    Args:
+        query: Natural language search query (e.g., "find authentication logic")
+        project_path: Path to the project root
+        n_results: Maximum number of results to return
+        language: Filter by language (e.g., "py", "ts", "js")
+        node_type: Filter by node type (e.g., "function", "class", "method")
+        decorator: Filter by decorator (e.g., "@property", "@staticmethod")
+        is_async: Filter by async status (True = async only, False = sync only)
+        base_class: Filter by base class (e.g., "BaseClass")
+        provider: Embedding provider (ollama, mxbai, gemini, openai, huggingface)
+
+    Returns:
+        Formatted search results with file paths and code snippets.
+    """
+    store = get_store(project_path, provider)
+    results = await store.search(
+        query,
+        n_results,
+        language,
+        node_type,
+        decorator=decorator,
+        is_async=is_async,
+        base_class=base_class,
+    )
+
+    if not results:
+        return "No results found"
+
+    if "error" in results[0]:
+        return f"Error: {results[0]['error']}\nHint: {results[0].get('hint', 'Check Ollama is running')}"
+
+    lines = [f"Found {len(results)} results for: '{query}'\n"]
+    for i, r in enumerate(results, 1):
+        lines.append(f"{i}. {r['file']}:{r['lines']} (relevance: {r['relevance']})")
+        lines.append(f"```{r['language']}")
+        lines.append(r["code_preview"])
+        lines.append("```\n")
+
+    return "\n".join(lines)
+
+
+async def hybrid_search(
+    query: str,
+    pattern: str | None = None,
+    project_path: str = ".",
+    n_results: int = 10,
+    language: str | None = None,
+    node_type: str | None = None,
+    decorator: str | None = None,
+    is_async: bool | None = None,
+    base_class: str | None = None,
+    provider: EmbeddingProvider = "ollama",
+) -> str:
+    """
+    Hybrid search combining semantic similarity with structural AST matching.
+
+    Performs semantic search first, then optionally filters/boosts results
+    that also match an ast-grep structural pattern.
+
+    Args:
+        query: Natural language search query (e.g., "find authentication logic")
+        pattern: Optional ast-grep pattern for structural matching (e.g., "def $FUNC($$$):")
+        project_path: Path to the project root
+        n_results: Maximum number of results to return
+        language: Filter by language (e.g., "py", "ts", "js")
+        node_type: Filter by node type (e.g., "function", "class", "method")
+        decorator: Filter by decorator (e.g., "@property", "@staticmethod")
+        is_async: Filter by async status (True = async only, False = sync only)
+        base_class: Filter by base class (e.g., "BaseClass")
+        provider: Embedding provider (ollama, gemini, openai)
+
+    Returns:
+        Formatted search results with relevance scores and structural match indicators.
+    """
+    from mcp_bridge.tools.code_search import ast_grep_search
+
+    # Get semantic results (fetch more if we're going to filter)
+    fetch_count = n_results * 2 if pattern else n_results
+    semantic_result = await semantic_search(
+        query=query,
+        project_path=project_path,
+        n_results=fetch_count,
+        language=language,
+        node_type=node_type,
+        decorator=decorator,
+        is_async=is_async,
+        base_class=base_class,
+        provider=provider,
+    )
+
+    if not pattern:
+        return semantic_result
+
+    if semantic_result.startswith("Error:") or semantic_result == "No results found":
+        return semantic_result
+
+    # Get structural matches from ast-grep
+    ast_result = await ast_grep_search(
+        pattern=pattern,
+        directory=project_path,
+        language=language or "",
+    )
+
+    # Extract file paths from ast-grep results
+    ast_files: set[str] = set()
+    if ast_result and not ast_result.startswith("Error:") and ast_result != "No matches found":
+        for line in ast_result.split("\n"):
+            if line.startswith("- "):
+                # Format: "- file.py:123"
+                file_part = line[2:].split(":")[0]
+                ast_files.add(file_part)
+
+    if not ast_files:
+        # No structural matches, return semantic results with note
+        return f"{semantic_result}\n\n[Note: No structural matches for pattern '{pattern}']"
+
+    # Parse semantic results and boost/annotate files that appear in both
+    lines = []
+    result_lines = semantic_result.split("\n")
+    header = result_lines[0] if result_lines else ""
+    lines.append(header.replace("results for:", "hybrid results for:"))
+    lines.append(f"[Structural pattern: {pattern}]\n")
+
+    i = 1
+    boosted_count = 0
+    while i < len(result_lines):
+        line = result_lines[i]
+        if line and (line[0].isdigit() or line.startswith("```") or line.strip()):
+            # Check if this is a result header line (e.g., "1. file.py:10-20")
+            if line and line[0].isdigit() and "." in line:
+                file_part = line.split()[1].split(":")[0] if len(line.split()) > 1 else ""
+                if file_part in ast_files:
+                    lines.append(f"{line} 🎯 [structural match]")
+                    boosted_count += 1
+                else:
+                    lines.append(line)
+            else:
+                lines.append(line)
+        else:
+            lines.append(line)
+        i += 1
+
+    lines.append(
+        f"\n[{boosted_count}/{len(ast_files)} semantic results also match structural pattern]"
+    )
+
+    return "\n".join(lines)
+
+
+async def index_codebase(
+    project_path: str = ".",
+    force: bool = False,
+    provider: EmbeddingProvider = "ollama",
+) -> str:
+    """
+    Index a codebase for semantic search.
+
+    Args:
+        project_path: Path to the project root
+        force: If True, reindex everything. Otherwise, only new/changed files.
+        provider: Embedding provider - ollama (local/free), mxbai (local/free),
+                  gemini (cloud/OAuth), openai (cloud/OAuth), huggingface (cloud/token)
+
+    Returns:
+        Indexing statistics.
+    """
+    store = get_store(project_path, provider)
+    stats = await store.index_codebase(force=force)
+
+    if "error" in stats:
+        return f"Error: {stats['error']}"
+
+    if stats.get("cancelled"):
+        return (
+            f"⚠️ Indexing cancelled\n"
+            f"Indexed {stats['indexed']} chunks from {stats['total_files']} files before cancellation\n"
+            f"{stats.get('message', '')}"
+        )
+
+    return (
+        f"Indexed {stats['indexed']} chunks from {stats['total_files']} files\n"
+        f"Database: {stats.get('db_path', 'unknown')}\n"
+        f"{stats.get('message', '')}"
+    )
+
+
+def cancel_indexing(
+    project_path: str = ".",
+    provider: EmbeddingProvider = "ollama",
+) -> str:
+    """
+    Cancel an ongoing indexing operation.
+
+    The cancellation happens gracefully between batches - the current batch
+    will complete before the operation stops.
+
+    Args:
+        project_path: Path to the project root
+        provider: Embedding provider (must match the one used for indexing)
+
+    Returns:
+        Confirmation message.
+    """
+    try:
+        store = get_store(project_path, provider)
+        store.request_cancel_indexing()
+        return f"✅ Cancellation requested for {project_path}\nIndexing will stop after current batch completes."
+    except Exception as e:
+        return f"❌ Error requesting cancellation: {e}"
+
+
+async def semantic_stats(
+    project_path: str = ".",
+    provider: EmbeddingProvider = "ollama",
+) -> str:
+    """
+    Get statistics about the semantic search index.
+
+    Args:
+        project_path: Path to the project root
+        provider: Embedding provider - ollama (local/free), mxbai (local/free),
+                  gemini (cloud/OAuth), openai (cloud/OAuth), huggingface (cloud/token)
+
+    Returns:
+        Index statistics.
+    """
+    store = get_store(project_path, provider)
+    stats = store.get_stats()
+
+    if "error" in stats:
+        return f"Error: {stats['error']}"
+
+    return (
+        f"Project: {stats['project_path']}\n"
+        f"Database: {stats['db_path']}\n"
+        f"Chunks indexed: {stats['chunks_indexed']}\n"
+        f"Embedding provider: {stats['embedding_provider']} ({stats['embedding_dimension']} dims)"
+    )
+
+
+def delete_index(
+    project_path: str = ".",
+    provider: EmbeddingProvider | None = None,
+    delete_all: bool = False,
+) -> str:
+    """
+    Delete semantic search index for a project.
+
+    Args:
+        project_path: Path to the project root
+        provider: Embedding provider (if None and delete_all=False, deletes all providers for this project)
+        delete_all: If True, delete ALL indexes for ALL projects (ignores project_path and provider)
+
+    Returns:
+        Confirmation message with deleted paths.
+    """
+    import shutil
+
+    vectordb_base = Path.home() / ".stravinsky" / "vectordb"
+
+    if not vectordb_base.exists():
+        return "✅ No semantic search indexes found (vectordb directory doesn't exist)"
+
+    if delete_all:
+        # Delete entire vectordb directory
+        try:
+            shutil.rmtree(vectordb_base)
+            return "✅ Deleted all semantic search indexes for all projects"
+        except Exception as e:
+            return f"❌ Error deleting all indexes: {e}"
+
+    # Generate project hash
+    project_path_resolved = Path(project_path).resolve()
+    project_hash = hashlib.md5(str(project_path_resolved).encode()).hexdigest()[:12]
+
+    deleted = []
+    errors = []
+
+    if provider:
+        # Delete specific provider index for this project
+        index_path = vectordb_base / f"{project_hash}_{provider}"
+        if index_path.exists():
+            try:
+                shutil.rmtree(index_path)
+                deleted.append(str(index_path))
+            except Exception as e:
+                errors.append(f"{provider}: {e}")
+        else:
+            errors.append(f"{provider}: Index not found")
+    else:
+        # Delete all provider indexes for this project
+        providers: list[EmbeddingProvider] = ["ollama", "mxbai", "gemini", "openai", "huggingface"]
+        for prov in providers:
+            index_path = vectordb_base / f"{project_hash}_{prov}"
+            if index_path.exists():
+                try:
+                    shutil.rmtree(index_path)
+                    deleted.append(str(index_path))
+                except Exception as e:
+                    errors.append(f"{prov}: {e}")
+
+    if not deleted and not errors:
+        return f"⚠️  No indexes found for project: {project_path_resolved}\nProject hash: {project_hash}"
+
+    result = []
+    if deleted:
+        result.append(f"✅ Deleted {len(deleted)} index(es):")
+        for path in deleted:
+            result.append(f"  - {path}")
+    if errors:
+        result.append(f"\n❌ Errors ({len(errors)}):")
+        for error in errors:
+            result.append(f"  - {error}")
+
+    return "\n".join(result)
+
+
+async def semantic_health(project_path: str = ".", provider: EmbeddingProvider = "ollama") -> str:
+    """Check health of semantic search system."""
+    store = get_store(project_path, provider)
+
+    status = []
+
+    # Check Provider
+    try:
+        is_avail = await store.check_embedding_service()
+        status.append(
+            f"Provider ({store.provider.name}): {'✅ Online' if is_avail else '❌ Offline'}"
+        )
+    except Exception as e:
+        status.append(f"Provider ({store.provider.name}): ❌ Error - {e}")
+
+    # Check DB
+    try:
+        count = store.collection.count()
+        status.append(f"Vector DB: ✅ Online ({count} documents)")
+    except Exception as e:
+        status.append(f"Vector DB: ❌ Error - {e}")
+
+    return "\n".join(status)
+
+
+# ========================
+# FILE WATCHER MANAGEMENT
+# ========================
+
+
+async def start_file_watcher(
+    project_path: str,
+    provider: EmbeddingProvider = "ollama",
+    debounce_seconds: float = 2.0,
+) -> "CodebaseFileWatcher":
+    """Start watching a project directory for file changes.
+
+    If an index exists, automatically performs an incremental reindex to catch up
+    on any changes that happened while the watcher was not running.
+
+    Args:
+        project_path: Path to the project root
+        provider: Embedding provider to use for reindexing
+        debounce_seconds: Time to wait before reindexing after changes
+
+    Returns:
+        The started CodebaseFileWatcher instance
+    """
+    path = str(Path(project_path).resolve())
+    with _watchers_lock:
+        if path not in _watchers:
+            store = get_store(project_path, provider)
+
+            # Check if index exists - CRITICAL: Must have index before watching
+            try:
+                stats = store.get_stats()
+                chunks_indexed = stats.get("chunks_indexed", 0)
+                if chunks_indexed == 0:
+                    raise ValueError(
+                        f"No semantic index found for '{path}'. "
+                        f"Run semantic_index(project_path='{path}', provider='{provider}') "
+                        f"before starting the file watcher."
+                    )
+
+                # Index exists - catch up on any missed changes
+                print(f"📋 Catching up on changes since last index...", file=sys.stderr)
+                await store.index_codebase(force=False)
+                print(f"✅ Index updated, starting file watcher", file=sys.stderr)
+
+            except ValueError:
+                # Re-raise ValueError (our intentional error)
+                raise
+            except Exception as e:
+                # Collection doesn't exist or other error
+                raise ValueError(
+                    f"No semantic index found for '{path}'. "
+                    f"Run semantic_index(project_path='{path}', provider='{provider}') "
+                    f"before starting the file watcher."
+                ) from e
+
+            watcher = store.start_watching(debounce_seconds=debounce_seconds)
+            _watchers[path] = watcher
+        else:
+            watcher = _watchers[path]
+            if not watcher.is_running():
+                watcher.start()
+        return _watchers[path]
+
+
+def stop_file_watcher(project_path: str) -> bool:
+    """Stop watching a project directory.
+
+    Args:
+        project_path: Path to the project root
+
+    Returns:
+        True if watcher was stopped, False if no watcher was active
+    """
+    path = str(Path(project_path).resolve())
+    with _watchers_lock:
+        if path in _watchers:
+            watcher = _watchers[path]
+            watcher.stop()
+            del _watchers[path]
+            return True
+        return False
+
+
+def get_file_watcher(project_path: str) -> "CodebaseFileWatcher | None":
+    """Get an active file watcher for a project.
+
+    Args:
+        project_path: Path to the project root
+
+    Returns:
+        The CodebaseFileWatcher if active, None otherwise
+    """
+    path = str(Path(project_path).resolve())
+    with _watchers_lock:
+        watcher = _watchers.get(path)
+        if watcher is not None and watcher.is_running():
+            return watcher
+        return None
+
+
+def list_file_watchers() -> list[dict]:
+    """List all active file watchers.
+
+    Returns:
+        List of dicts with watcher info (project_path, debounce_seconds, provider, status)
+    """
+    with _watchers_lock:
+        watchers_info = []
+        for path, watcher in _watchers.items():
+            watchers_info.append(
+                {
+                    "project_path": path,
+                    "debounce_seconds": watcher.debounce_seconds,
+                    "provider": watcher.store.provider_name,
+                    "status": "running" if watcher.is_running() else "stopped",
+                }
+            )
+        return watchers_info
+
+
+# ========================
+# MULTI-QUERY EXPANSION & DECOMPOSITION
+# ========================
+
+
+async def _expand_query_with_llm(query: str, num_variations: int = 3) -> list[str]:
+    """
+    Use LLM to rephrase a query into multiple semantic variations.
+
+    For example: "database connection" -> ["SQLAlchemy engine setup",
+    "connect to postgres", "db session management"]
+
+    Args:
+        query: Original search query
+        num_variations: Number of variations to generate (default: 3)
+
+    Returns:
+        List of query variations including the original
+    """
+    from mcp_bridge.tools.model_invoke import invoke_gemini
+
+    prompt = f"""You are a code search query expander. Given a search query, generate {num_variations} alternative phrasings that would help find relevant code.
+
+Original query: "{query}"
+
+Generate {num_variations} alternative queries that:
+1. Use different technical terminology (e.g., "database" -> "SQLAlchemy", "ORM", "connection pool")
+2. Reference specific implementations or patterns
+3. Include related concepts that might appear in code
+
+Return ONLY the alternative queries, one per line. No numbering, no explanations.
+Example output for "database connection":
+SQLAlchemy engine configuration
+postgres connection setup
+db session factory pattern"""
+
+    try:
+        result = await invoke_gemini(
+            token_store=TokenStore(),
+            prompt=prompt,
+            model="gemini-2.0-flash",
+            temperature=0.7,
+            max_tokens=200,
+        )
+
+        # Parse variations from response
+        variations = [line.strip() for line in result.strip().split("\n") if line.strip()]
+        # Always include original query first
+        all_queries = [query] + variations[:num_variations]
+        return all_queries
+
+    except Exception as e:
+        logger.warning(f"Query expansion failed: {e}, using original query only")
+        return [query]
+
+
+async def _decompose_query_with_llm(query: str) -> list[str]:
+    """
+    Break a complex query into smaller, focused sub-questions.
+
+    For example: "Initialize the DB and then create a user model" ->
+    ["database initialization", "user model definition"]
+
+    Args:
+        query: Complex search query
+
+    Returns:
+        List of sub-queries, or [query] if decomposition not needed
+    """
+    from mcp_bridge.tools.model_invoke import invoke_gemini
+
+    prompt = f"""You are a code search query analyzer. Determine if this query should be broken into sub-queries.
+
+Query: "{query}"
+
+If the query contains multiple distinct concepts (connected by "and", "then", "also", etc.),
+break it into separate focused sub-queries.
+
+If the query is already focused on a single concept, return just that query.
+
+Return ONLY the sub-queries, one per line. No numbering, no explanations.
+
+Examples:
+- "Initialize the DB and then create a user model" -> 
+database initialization
+user model definition
+
+- "authentication logic" ->
+authentication logic"""
+
+    try:
+        result = await invoke_gemini(
+            token_store=TokenStore(),
+            prompt=prompt,
+            model="gemini-2.0-flash",
+            temperature=0.3,  # Lower temperature for more consistent decomposition
+            max_tokens=150,
+        )
+
+        # Parse sub-queries from response
+        sub_queries = [line.strip() for line in result.strip().split("\n") if line.strip()]
+        return sub_queries if sub_queries else [query]
+
+    except Exception as e:
+        logger.warning(f"Query decomposition failed: {e}, using original query")
+        return [query]
+
+
+def _aggregate_results(
+    all_results: list[list[dict]],
+    n_results: int = 10,
+) -> list[dict]:
+    """
+    Aggregate and deduplicate results from multiple queries.
+
+    Uses reciprocal rank fusion to combine relevance scores from different queries.
+
+    Args:
+        all_results: List of result lists from different queries
+        n_results: Maximum number of results to return
+
+    Returns:
+        Deduplicated and re-ranked results
+    """
+    # Track seen files to avoid duplicates
+    seen_files: dict[str, dict] = {}  # file:lines -> result with best score
+    file_scores: dict[str, float] = {}  # file:lines -> aggregated score
+
+    # Reciprocal Rank Fusion constant
+    k = 60
+
+    for _query_idx, results in enumerate(all_results):
+        for rank, result in enumerate(results):
+            file_key = f"{result.get('file', '')}:{result.get('lines', '')}"
+
+            # RRF score contribution
+            rrf_score = 1 / (k + rank + 1)
+
+            if file_key not in seen_files:
+                seen_files[file_key] = result.copy()
+                file_scores[file_key] = rrf_score
+            else:
+                # Aggregate scores
+                file_scores[file_key] += rrf_score
+                # Keep higher original relevance if available
+                if result.get("relevance", 0) > seen_files[file_key].get("relevance", 0):
+                    seen_files[file_key] = result.copy()
+
+    # Sort by aggregated score and return top N
+    sorted_keys = sorted(file_scores.keys(), key=lambda k: file_scores[k], reverse=True)
+
+    aggregated = []
+    for key in sorted_keys[:n_results]:
+        result = seen_files[key]
+        # Update relevance to reflect aggregated score (normalized)
+        max_score = max(file_scores.values()) if file_scores else 1
+        result["relevance"] = round(file_scores[key] / max_score, 3)
+        aggregated.append(result)
+
+    return aggregated
+
+
+async def multi_query_search(
+    query: str,
+    project_path: str = ".",
+    n_results: int = 10,
+    num_expansions: int = 3,
+    language: str | None = None,
+    node_type: str | None = None,
+    provider: EmbeddingProvider = "ollama",
+) -> str:
+    """
+    Search with LLM-expanded query variations for better recall.
+
+    Rephrases the query into multiple semantic variations, searches for each,
+    and aggregates results using reciprocal rank fusion.
+
+    Args:
+        query: Natural language search query
+        project_path: Path to the project root
+        n_results: Maximum number of results to return
+        num_expansions: Number of query variations to generate (default: 3)
+        language: Filter by language (e.g., "py", "ts")
+        node_type: Filter by node type (e.g., "function", "class")
+        provider: Embedding provider
+
+    Returns:
+        Formatted search results with relevance scores.
+    """
+    import asyncio
+
+    print(f"🔍 MULTI-QUERY: Expanding '{query[:50]}...'", file=sys.stderr)
+
+    # Get query expansions
+    expanded_queries = await _expand_query_with_llm(query, num_expansions)
+    print(f"  Generated {len(expanded_queries)} query variations", file=sys.stderr)
+
+    # Get store once
+    store = get_store(project_path, provider)
+
+    # Search with all queries in parallel
+    async def search_single(q: str) -> list[dict]:
+        return await store.search(
+            q,
+            n_results=n_results,  # Get full results for each query
+            language=language,
+            node_type=node_type,
+        )
+
+    all_results = await asyncio.gather(*[search_single(q) for q in expanded_queries])
+
+    # Filter out error results
+    valid_results = [r for r in all_results if r and "error" not in r[0]]
+
+    if not valid_results:
+        if all_results and all_results[0] and "error" in all_results[0][0]:
+            return f"Error: {all_results[0][0]['error']}"
+        return "No results found"
+
+    # Aggregate results
+    aggregated = _aggregate_results(valid_results, n_results)
+
+    if not aggregated:
+        return "No results found"
+
+    # Format output
+    lines = [f"Found {len(aggregated)} results for multi-query expansion of: '{query}'"]
+    lines.append(
+        f"[Expanded to: {', '.join(q[:30] + '...' if len(q) > 30 else q for q in expanded_queries)}]\n"
+    )
+
+    for i, r in enumerate(aggregated, 1):
+        lines.append(f"{i}. {r['file']}:{r['lines']} (relevance: {r['relevance']})")
+        lines.append(f"```{r.get('language', '')}")
+        lines.append(r.get("code_preview", ""))
+        lines.append("```\n")
+
+    return "\n".join(lines)
+
+
+async def decomposed_search(
+    query: str,
+    project_path: str = ".",
+    n_results: int = 10,
+    language: str | None = None,
+    node_type: str | None = None,
+    provider: EmbeddingProvider = "ollama",
+) -> str:
+    """
+    Search by decomposing complex queries into focused sub-questions.
+
+    Breaks multi-part queries like "Initialize the DB and create a user model"
+    into separate searches, returning organized results for each part.
+
+    Args:
+        query: Complex search query (may contain multiple concepts)
+        project_path: Path to the project root
+        n_results: Maximum results per sub-query
+        language: Filter by language
+        node_type: Filter by node type
+        provider: Embedding provider
+
+    Returns:
+        Formatted results organized by sub-question.
+    """
+    import asyncio
+
+    print(f"🔍 DECOMPOSED-SEARCH: Analyzing '{query[:50]}...'", file=sys.stderr)
+
+    # Decompose query
+    sub_queries = await _decompose_query_with_llm(query)
+    print(f"  Decomposed into {len(sub_queries)} sub-queries", file=sys.stderr)
+
+    if len(sub_queries) == 1 and sub_queries[0] == query:
+        # No decomposition needed, use regular search
+        return await semantic_search(
+            query=query,
+            project_path=project_path,
+            n_results=n_results,
+            language=language,
+            node_type=node_type,
+            provider=provider,
+        )
+
+    # Get store once
+    store = get_store(project_path, provider)
+
+    # Search each sub-query in parallel
+    async def search_sub(q: str) -> tuple[str, list[dict]]:
+        results = await store.search(
+            q,
+            n_results=n_results // len(sub_queries) + 2,  # Distribute results
+            language=language,
+            node_type=node_type,
+        )
+        return (q, results)
+
+    sub_results = await asyncio.gather(*[search_sub(q) for q in sub_queries])
+
+    # Format output with sections for each sub-query
+    lines = [f"Decomposed search for: '{query}'"]
+    lines.append(f"[Split into {len(sub_queries)} sub-queries]\n")
+
+    total_results = 0
+    for sub_query, results in sub_results:
+        lines.append(f"### {sub_query}")
+
+        if not results or (results and "error" in results[0]):
+            lines.append("  No results found\n")
+            continue
+
+        for i, r in enumerate(results[:5], 1):  # Limit per sub-query
+            lines.append(f"  {i}. {r['file']}:{r['lines']} (relevance: {r['relevance']})")
+            # Shorter preview for decomposed results
+            preview = r.get("code_preview", "")[:200]
+            if len(r.get("code_preview", "")) > 200:
+                preview += "..."
+            lines.append(f"     ```{r.get('language', '')}")
+            lines.append(f"     {preview}")
+            lines.append("     ```")
+            total_results += 1
+        lines.append("")
+
+    lines.append(f"[Total: {total_results} results across {len(sub_queries)} sub-queries]")
+
+    return "\n".join(lines)
+
+
+async def enhanced_search(
+    query: str,
+    project_path: str = ".",
+    n_results: int = 10,
+    mode: str = "auto",
+    language: str | None = None,
+    node_type: str | None = None,
+    provider: EmbeddingProvider = "ollama",
+) -> str:
+    """
+    Unified enhanced search combining expansion and decomposition.
+
+    Automatically selects the best strategy based on query complexity:
+    - Simple queries: Multi-query expansion for better recall
+    - Complex queries: Decomposition + expansion for comprehensive coverage
+
+    Args:
+        query: Search query (simple or complex)
+        project_path: Path to the project root
+        n_results: Maximum number of results
+        mode: Search mode - "auto", "expand", "decompose", or "both"
+        language: Filter by language
+        node_type: Filter by node type
+        provider: Embedding provider
+
+    Returns:
+        Formatted search results.
+    """
+    # Use classifier for intelligent mode selection
+    classification = classify_query(query)
+    logger.debug(
+        f"Query classified as {classification.category.value} "
+        f"(confidence: {classification.confidence:.2f}, suggested: {classification.suggested_tool})"
+    )
+
+    # Determine mode based on classification
+    if mode == "auto":
+        # HYBRID → decompose (complex multi-part queries)
+        # SEMANTIC → expand (conceptual queries benefit from variations)
+        # PATTERN/STRUCTURAL → expand (simple queries, quick path)
+        mode = "decompose" if classification.category == QueryCategory.HYBRID else "expand"
+
+    if mode == "decompose":
+        return await decomposed_search(
+            query=query,
+            project_path=project_path,
+            n_results=n_results,
+            language=language,
+            node_type=node_type,
+            provider=provider,
+        )
+    elif mode == "expand":
+        return await multi_query_search(
+            query=query,
+            project_path=project_path,
+            n_results=n_results,
+            language=language,
+            node_type=node_type,
+            provider=provider,
+        )
+    elif mode == "both":
+        # Decompose first, then expand each sub-query
+        sub_queries = await _decompose_query_with_llm(query)
+
+        all_results: list[list[dict]] = []
+        store = get_store(project_path, provider)
+
+        for sub_q in sub_queries:
+            # Expand each sub-query
+            expanded = await _expand_query_with_llm(sub_q, num_variations=2)
+            for exp_q in expanded:
+                results = await store.search(
+                    exp_q,
+                    n_results=5,
+                    language=language,
+                    node_type=node_type,
+                )
+                if results and "error" not in results[0]:
+                    all_results.append(results)
+
+        aggregated = _aggregate_results(all_results, n_results)
+
+        if not aggregated:
+            return "No results found"
+
+        lines = [f"Enhanced search (decompose+expand) for: '{query}'"]
+        lines.append(f"[{len(sub_queries)} sub-queries × expansions]\n")
+
+        for i, r in enumerate(aggregated, 1):
+            lines.append(f"{i}. {r['file']}:{r['lines']} (relevance: {r['relevance']})")
+            lines.append(f"```{r.get('language', '')}")
+            lines.append(r.get("code_preview", ""))
+            lines.append("```\n")
+
+        return "\n".join(lines)
+
+    else:
+        return f"Unknown mode: {mode}. Use 'auto', 'expand', 'decompose', or 'both'"
+
+
+# ========================
+# FILE WATCHER IMPLEMENTATION
+# ========================
+
+
+class CodebaseFileWatcher:
+    """Watch a project directory for file changes and trigger reindexing.
+
+    Features:
+    - Watches for file create, modify, delete, move events
+    - Filters to .py files only
+    - Skips hidden files and directories (., .git, __pycache__, venv, etc.)
+    - Debounces rapid changes to batch them into a single reindex
+    - Thread-safe with daemon threads for clean shutdown
+    - Integrates with CodebaseVectorStore for incremental indexing
+    """
+
+    # Default debounce time in seconds
+    DEFAULT_DEBOUNCE_SECONDS = 2.0
+
+    def __init__(
+        self,
+        project_path: Path | str,
+        store: CodebaseVectorStore,
+        debounce_seconds: float = DEFAULT_DEBOUNCE_SECONDS,
+    ):
+        """Initialize the file watcher.
+
+        Args:
+            project_path: Path to the project root to watch
+            store: CodebaseVectorStore instance for reindexing
+            debounce_seconds: Time to wait before reindexing after changes (default: 2.0s)
+        """
+        self.project_path = Path(project_path).resolve()
+        self.store = store
+        self.debounce_seconds = debounce_seconds
+
+        # Observer and handler for watchdog
+        self._observer = None
+        self._event_handler = None
+
+        # Thread safety
+        self._lock = threading.Lock()
+        self._running = False
+
+        # Debouncing
+        self._pending_reindex_timer: threading.Timer | None = None
+        self._pending_files: set[Path] = set()
+        self._pending_lock = threading.Lock()
+
+    def start(self) -> None:
+        """Start watching the project directory.
+
+        Creates and starts a watchdog observer in a daemon thread.
+        """
+        with self._lock:
+            if self._running:
+                logger.warning(f"Watcher for {self.project_path} is already running")
+                return
+
+            try:
+                watchdog = get_watchdog()
+                Observer = watchdog["Observer"]
+
+                # Create event handler class and instantiate
+                FileChangeHandler = _create_file_change_handler_class()
+                self._event_handler = FileChangeHandler(
+                    project_path=self.project_path,
+                    watcher=self,
+                )
+
+                # Create and start observer (daemon mode for clean shutdown)
+                self._observer = Observer()
+                self._observer.daemon = True
+                self._observer.schedule(
+                    self._event_handler,
+                    str(self.project_path),
+                    recursive=True,
+                )
+                self._observer.start()
+                self._running = True
+                logger.info(f"File watcher started for {self.project_path}")
+
+            except Exception as e:
+                logger.error(f"Failed to start file watcher: {e}")
+                self._running = False
+                raise
+
+    def stop(self) -> None:
+        """Stop watching the project directory.
+
+        Cancels any pending reindex timers and stops the observer.
+        """
+        with self._lock:
+            # Cancel pending reindex
+            if self._pending_reindex_timer is not None:
+                self._pending_reindex_timer.cancel()
+                self._pending_reindex_timer = None
+
+            # Stop observer
+            if self._observer is not None:
+                self._observer.stop()
+                self._observer.join(timeout=5)  # Wait up to 5 seconds for shutdown
+                self._observer = None
+
+            self._event_handler = None
+            self._running = False
+            logger.info(f"File watcher stopped for {self.project_path}")
+
+    def is_running(self) -> bool:
+        """Check if the watcher is currently running.
+
+        Returns:
+            True if watcher is active, False otherwise
+        """
+        with self._lock:
+            return self._running and self._observer is not None and self._observer.is_alive()
+
+    def _on_file_changed(self, file_path: Path) -> None:
+        """Called when a file changes (internal use by _FileChangeHandler).
+
+        Accumulates files and triggers debounced reindex.
+
+        Args:
+            file_path: Path to the changed file
+        """
+        with self._pending_lock:
+            self._pending_files.add(file_path)
+
+            # Cancel previous timer
+            if self._pending_reindex_timer is not None:
+                self._pending_reindex_timer.cancel()
+
+            # Start new timer
+            self._pending_reindex_timer = self._create_debounce_timer()
+            self._pending_reindex_timer.start()
+
+    def _create_debounce_timer(self) -> threading.Timer:
+        """Create a new debounce timer for reindexing.
+
+        Returns:
+            A threading.Timer configured for debounce reindexing
+        """
+        return threading.Timer(
+            self.debounce_seconds,
+            self._trigger_reindex,
+        )
+
+    def _trigger_reindex(self) -> None:
+        """Trigger reindexing of accumulated changed files.
+
+        This is called after the debounce period expires. It performs an
+        incremental reindex focusing on the changed files.
+        """
+        import asyncio
+
+        with self._pending_lock:
+            if not self._pending_files:
+                self._pending_reindex_timer = None
+                return
+
+            files_to_index = list(self._pending_files)
+            self._pending_files.clear()
+            self._pending_reindex_timer = None
+
+        # Run async reindex in a new event loop
+        try:
+            loop = asyncio.new_event_loop()
+            asyncio.set_event_loop(loop)
+            try:
+                loop.run_until_complete(self.store.index_codebase(force=False))
+                logger.debug(f"Reindexed {len(files_to_index)} changed files")
+            finally:
+                loop.close()
+        except Exception as e:
+            logger.error(f"Error during file watcher reindex: {e}")
+
+
+def _create_file_change_handler_class():
+    """Create FileChangeHandler class that inherits from FileSystemEventHandler.
+
+    This is a factory function that creates the handler class dynamically
+    after watchdog is imported, allowing for lazy loading.
+    """
+    watchdog = get_watchdog()
+    FileSystemEventHandler = watchdog["FileSystemEventHandler"]
+
+    class _FileChangeHandler(FileSystemEventHandler):
+        """Watchdog event handler for file system changes.
+
+        Detects file create, modify, delete, and move events, filters them,
+        and notifies the watcher of relevant changes.
+        """
+
+        def __init__(self, project_path: Path, watcher: CodebaseFileWatcher):
+            """Initialize the event handler.
+
+            Args:
+                project_path: Root path of the project being watched
+                watcher: CodebaseFileWatcher instance to notify
+            """
+            super().__init__()
+            self.project_path = project_path
+            self.watcher = watcher
+
+        def on_created(self, event) -> None:
+            """Called when a file is created."""
+            if not event.is_directory and self._should_index_file(event.src_path):
+                logger.debug(f"File created: {event.src_path}")
+                self.watcher._on_file_changed(Path(event.src_path))
+
+        def on_modified(self, event) -> None:
+            """Called when a file is modified."""
+            if not event.is_directory and self._should_index_file(event.src_path):
+                logger.debug(f"File modified: {event.src_path}")
+                self.watcher._on_file_changed(Path(event.src_path))
+
+        def on_deleted(self, event) -> None:
+            """Called when a file is deleted."""
+            if not event.is_directory and self._should_index_file(event.src_path):
+                logger.debug(f"File deleted: {event.src_path}")
+                self.watcher._on_file_changed(Path(event.src_path))
+
+        def on_moved(self, event) -> None:
+            """Called when a file is moved."""
+            if not event.is_directory:
+                # Check destination path
+                if self._should_index_file(event.dest_path):
+                    logger.debug(f"File moved: {event.src_path} -> {event.dest_path}")
+                    self.watcher._on_file_changed(Path(event.dest_path))
+                # Also check source path (for deletion case)
+                elif self._should_index_file(event.src_path):
+                    logger.debug(f"File moved out: {event.src_path}")
+                    self.watcher._on_file_changed(Path(event.src_path))
+
+        def _should_index_file(self, file_path: str) -> bool:
+            """Check if a file should trigger reindexing.
+
+            Filters based on:
+            - File extension (.py only)
+            - Hidden files and directories (starting with .)
+            - Skip directories (venv, __pycache__, .git, node_modules, etc.)
+
+            Args:
+                file_path: Path to the file to check
+
+            Returns:
+                True if file should trigger reindexing, False otherwise
+            """
+            path = Path(file_path)
+
+            # Only .py files
+            if path.suffix != ".py":
+                return False
+
+            # Skip hidden files
+            if path.name.startswith("."):
+                return False
+
+            # Check for skip directories in the path
+            for part in path.parts:
+                if part.startswith("."):  # Hidden directories like .git, .venv
+                    return False
+                if part in {"__pycache__", "venv", "env", "node_modules"}:
+                    return False
+
+            # File is within project (resolve both paths to handle symlinks)
+            try:
+                path.resolve().relative_to(self.project_path)
+                return True
+            except ValueError:
+                # File is outside project
+                return False
+
+    return _FileChangeHandler
+
+
+# ========================
+# CHROMADB LOCK CLEANUP
+# ========================
+
+
+def _is_process_alive(pid: int) -> bool:
+    """Check if a process with given PID is currently running.
+
+    Cross-platform process existence check.
+
+    Args:
+        pid: Process ID to check
+
+    Returns:
+        True if process exists, False otherwise
+    """
+    import os
+    import sys
+
+    if sys.platform == "win32":
+        # Windows: Use tasklist command
+        import subprocess
+
+        try:
+            result = subprocess.run(
+                ["tasklist", "/FI", f"PID eq {pid}"], capture_output=True, text=True, timeout=2
+            )
+            return str(pid) in result.stdout
+        except Exception:
+            return False
+    else:
+        # Unix/Linux/macOS: Use os.kill(pid, 0)
+        try:
+            os.kill(pid, 0)
+            return True
+        except OSError:
+            return False
+        except Exception:
+            return False
+
+
+def cleanup_stale_chromadb_locks() -> int:
+    """Remove stale ChromaDB lock files on MCP server startup.
+
+    Scans all vectordb directories and removes lock files that:
+    1. Are older than 60 seconds (short grace period for active operations)
+    2. Don't have an owning process running (if PID can be determined)
+
+    This prevents 'Connection closed' errors from dead process locks.
+
+    Returns:
+        Number of stale locks removed
+    """
+    vectordb_base = Path.home() / ".stravinsky" / "vectordb"
+    if not vectordb_base.exists():
+        return 0  # No vectordb yet, nothing to cleanup
+
+    import time
+
+    removed_count = 0
+
+    for project_dir in vectordb_base.iterdir():
+        if not project_dir.is_dir():
+            continue
+
+        lock_path = project_dir / ".chromadb.lock"
+        if not lock_path.exists():
+            continue
+
+        # Check lock age
+        try:
+            lock_age = time.time() - lock_path.stat().st_mtime
+        except Exception:
+            continue
+
+        # Aggressive cleanup: remove locks older than 60 seconds
+        # This catches recently crashed processes (old 300s was too conservative)
+        is_stale = lock_age > 60
+
+        # TODO: If lock file contains PID, check if process is alive
+        # filelock doesn't write PID by default, but we could enhance this
+
+        if is_stale:
+            try:
+                lock_path.unlink(missing_ok=True)
+                removed_count += 1
+                logger.info(f"Removed stale lock: {lock_path} (age: {lock_age:.0f}s)")
+            except Exception as e:
+                logger.warning(f"Could not remove stale lock {lock_path}: {e}")
+
+    if removed_count > 0:
+        logger.info(f"Startup cleanup: removed {removed_count} stale ChromaDB lock(s)")
+
+    return removed_count