feat: Support for vector search with Qdrant

mempalace/README.md

@@ -9,9 +9,9 @@ The Python package that powers MemPalace. All modules, all logic.
 | `cli.py` | CLI entry point — routes to mine, search, init, compress, wake-up |
 | `config.py` | Configuration loading — `~/.mempalace/config.json`, env vars, defaults |
 | `normalize.py` | Converts 5 chat formats (Claude Code JSONL, Claude.ai JSON, ChatGPT JSON, Slack JSON, plain text) to standard transcript format |
-| `miner.py` | Project file ingest — scans directories, chunks by paragraph, stores to ChromaDB |
+| `miner.py` | Project file ingest — scans directories, chunks by paragraph, stores to the vector store |
 | `convo_miner.py` | Conversation ingest — chunks by exchange pair (Q+A), detects rooms from content |
-| `searcher.py` | Semantic search via ChromaDB vectors — filters by wing/room, returns verbatim + scores |
+| `searcher.py` | Semantic search — filters by wing/room, returns verbatim + scores |
 | `layers.py` | 4-layer memory stack: L0 (identity), L1 (critical facts), L2 (room recall), L3 (deep search) |
 | `dialect.py` | AAAK compression — entity codes, emotion markers, 30x lossless ratio |
 | `knowledge_graph.py` | Temporal entity-relationship graph — SQLite, time-filtered queries, fact invalidation |
@@ -28,7 +28,7 @@ The Python package that powers MemPalace. All modules, all logic.
 ## Architecture
 
 ```
-User → CLI → miner/convo_miner → ChromaDB (palace)
+User → CLI → miner/convo_miner → vector store (palace: ChromaDB or Qdrant)
                                      ↕
                               knowledge_graph (SQLite)
                                      ↕
@@ -37,4 +37,4 @@ User → MCP Server → searcher → results
                   → diary    → agent journal
 ```
 
-The palace (ChromaDB) stores verbatim content. The knowledge graph (SQLite) stores structured relationships. The MCP server exposes both to any AI tool.
+The palace (ChromaDB or Qdrant) stores verbatim content. The knowledge graph (SQLite) stores structured relationships. The MCP server exposes both to any AI tool.

mempalace/backends/chroma_store.py

@@ -0,0 +1,53 @@
+"""ChromaDB vector store backend for MemPalace."""
+
+import os
+import types
+
+import chromadb
+
+_client_cache = None
+_collection_cache = None
+_config = None
+
+
+def _get_collection(create=False):
+    global _client_cache, _collection_cache
+    if _config is None:
+        return None
+    try:
+        if _client_cache is None:
+            _client_cache = chromadb.PersistentClient(path=_config.palace_path)
+        if create:
+            _collection_cache = _client_cache.get_or_create_collection(_config.collection_name)
+        elif _collection_cache is None:
+            _collection_cache = _client_cache.get_collection(_config.collection_name)
+        return _collection_cache
+    except Exception:
+        return None
+
+
+def get_chroma_collection(palace_path: str, collection_name: str, create: bool = False):
+    global _config, _client_cache, _collection_cache
+    if (
+        _config is None
+        or _config.palace_path != palace_path
+        or _config.collection_name != collection_name
+    ):
+        _config = types.SimpleNamespace(palace_path=palace_path, collection_name=collection_name)
+        _client_cache = None
+        _collection_cache = None
+    os.makedirs(palace_path, exist_ok=True)
+    return _get_collection(create=create)
+
+
+def reset_chroma_collection(palace_path: str, collection_name: str):
+    global _client_cache, _collection_cache
+
+    client = chromadb.PersistentClient(path=palace_path)
+    _client_cache = None
+    _collection_cache = None
+    try:
+        client.delete_collection(collection_name)
+    except Exception:
+        pass
+    return client.create_collection(collection_name)

mempalace/backends/qdrant_store.py

@@ -0,0 +1,225 @@
+import uuid
+from typing import Optional
+
+from . import VectorCollection
+
+_KEY_DOCUMENT = "document"
+_KEY_DRAWER_ID = "_drawer_id"
+_KEY_IDS = "ids"
+_KEY_DOCUMENTS = "documents"
+_KEY_METADATAS = "metadatas"
+_KEY_DISTANCES = "distances"
+
+_DEFAULT_EMBEDDING_MODEL = "sentence-transformers/all-MiniLM-L6-v2"
+
+_embedders: dict = {}
+
+
+def _get_embedder(model: str):
+    if model not in _embedders:
+        from fastembed import TextEmbedding
+
+        _embedders[model] = TextEmbedding(model)
+    return _embedders[model]
+
+
+def _embed(texts: list[str], model: str) -> list[list[float]]:
+    return [list(v) for v in _get_embedder(model).embed(texts)]
+
+
+def _vector_size_for_model(model: str) -> int:
+    return _get_embedder(model).embedding_size
+
+
+def _str_to_uuid(s: str) -> str:
+    return str(uuid.uuid5(uuid.NAMESPACE_DNS, s))
+
+
+def _unpack_payload(point_or_record) -> tuple[str, str, dict]:
+    payload = dict(point_or_record.payload or {})
+    doc = payload.pop(_KEY_DOCUMENT, "")
+    orig_id = payload.pop(_KEY_DRAWER_ID, str(point_or_record.id))
+    return orig_id, doc, payload
+
+
+def _where_to_filter(where: Optional[dict]):
+    if not where:
+        return None
+
+    from qdrant_client.models import FieldCondition, Filter, MatchValue
+
+    pairs = []
+    if "$and" in where:
+        for clause in where["$and"]:
+            pairs.extend(clause.items())
+    else:
+        pairs = list(where.items())
+
+    conditions = [FieldCondition(key=k, match=MatchValue(value=v)) for k, v in pairs]
+    return Filter(must=conditions) if conditions else None  # type: ignore[arg-type]
+
+
+def _ensure_collection(client, collection_name: str, vector_size: int) -> None:
+    from qdrant_client.models import Distance, VectorParams
+
+    if not client.collection_exists(collection_name):
+        client.create_collection(
+            collection_name=collection_name,
+            vectors_config=VectorParams(size=vector_size, distance=Distance.COSINE),
+        )
+
+
+class QdrantCollection(VectorCollection):
+    _SCROLL_END = object()
+
+    def __init__(self, client, collection_name: str, model: str, vector_size: int):
+        self._client = client
+        self._name = collection_name
+        self._model = model
+        self._vector_size = vector_size
+        self._scroll_cursor = None
+
+    def add(self, documents, ids, metadatas):
+        from qdrant_client.models import PointStruct
+
+        docs = list(documents)
+        _ensure_collection(self._client, self._name, self._vector_size)
+
+        vectors = _embed(docs, self._model)
+        points = []
+        for doc, meta, orig_id, vec in zip(docs, metadatas, ids, vectors):
+            payload = dict(meta)
+            payload[_KEY_DRAWER_ID] = orig_id
+            payload[_KEY_DOCUMENT] = doc
+            points.append(PointStruct(id=_str_to_uuid(orig_id), vector=vec, payload=payload))
+
+        self._client.upsert(collection_name=self._name, points=points)
+
+    def upsert(self, ids, documents, metadatas):
+        self.add(documents=documents, ids=ids, metadatas=metadatas)
+
+    def delete(self, ids):
+        qdrant_ids = [_str_to_uuid(i) for i in ids]
+        self._client.delete(
+            collection_name=self._name,
+            points_selector=qdrant_ids,
+        )
+
+    def query(self, query_texts, n_results, where=None, include=None):
+        query_filter = _where_to_filter(where)
+        query_vec = _embed([query_texts[0]], self._model)[0]
+
+        results = self._client.query_points(
+            collection_name=self._name,
+            query=query_vec,
+            query_filter=query_filter,
+            limit=n_results,
+            with_payload=True,
+        ).points
+
+        docs, metas, dists, ids = [], [], [], []
+        for point in results:
+            orig_id, doc, meta = _unpack_payload(point)
+            docs.append(doc)
+            metas.append(meta)
+            dists.append(min(1.0, max(0.0, 1.0 - point.score)))
+            ids.append(orig_id)
+
+        return {
+            _KEY_DOCUMENTS: [docs],
+            _KEY_METADATAS: [metas],
+            _KEY_DISTANCES: [dists],
+            _KEY_IDS: [ids],
+        }
+
+    def get(self, where=None, limit=None, offset=None, include=None, ids=None):
+        if ids is not None:
+            qdrant_ids = [_str_to_uuid(i) for i in ids]
+            records = self._client.retrieve(
+                collection_name=self._name, ids=qdrant_ids, with_payload=True
+            )
+            return self._records_to_result(records, include)
+
+        scroll_filter = _where_to_filter(where)
+
+        if not offset:
+            self._scroll_cursor = None
+
+        if self._scroll_cursor is self._SCROLL_END:
+            return self._records_to_result([], include)
+
+        records, next_cursor = self._client.scroll(
+            collection_name=self._name,
+            scroll_filter=scroll_filter,
+            limit=limit or 10,
+            offset=self._scroll_cursor,
+            with_payload=True,
+            with_vectors=False,
+        )
+        self._scroll_cursor = next_cursor if next_cursor is not None else self._SCROLL_END
+
+        return self._records_to_result(records, include)
+
+    def count(self):
+        return self._client.count(collection_name=self._name).count
+
+    def _records_to_result(self, records, include=None):
+        result_ids, result_docs, result_metas = [], [], []
+
+        for record in records:
+            orig_id, doc, meta = _unpack_payload(record)
+            result_ids.append(orig_id)
+            if not include or _KEY_DOCUMENTS in include:
+                result_docs.append(doc)
+            if not include or _KEY_METADATAS in include:
+                result_metas.append(meta)
+
+        result: dict = {_KEY_IDS: result_ids}
+        if not include or _KEY_DOCUMENTS in include:
+            result[_KEY_DOCUMENTS] = result_docs
+        if not include or _KEY_METADATAS in include:
+            result[_KEY_METADATAS] = result_metas
+        return result
+
+
+def get_qdrant_client(config):
+    try:
+        from qdrant_client import QdrantClient
+    except ImportError as exc:
+        raise ImportError(
+            "qdrant-client is not installed. "
+            "Install with: pip install 'mempalace[qdrant]' "
+            "or: pip install 'qdrant-client[fastembed]'"
+        ) from exc
+
+    qdrant_url = getattr(config, "qdrant_url", None)
+    qdrant_api_key = getattr(config, "qdrant_api_key", None)
+
+    return QdrantClient(url=qdrant_url, api_key=qdrant_api_key or None)
+
+
+def get_qdrant_collection(config, collection_name: str, create: bool = False) -> QdrantCollection:
+    client = get_qdrant_client(config)
+    model = getattr(config, "qdrant_embedding_model", _DEFAULT_EMBEDDING_MODEL)
+    vector_size = _vector_size_for_model(model)
+
+    if not create and not client.collection_exists(collection_name):
+        raise ValueError(
+            f"Qdrant collection '{collection_name}' not found. Run: mempalace mine <dir>"
+        )
+
+    if create:
+        _ensure_collection(client, collection_name, vector_size)
+
+    return QdrantCollection(client, collection_name, model=model, vector_size=vector_size)
+
+
+def reset_qdrant_collection(config, collection_name: str) -> QdrantCollection:
+    client = get_qdrant_client(config)
+    model = getattr(config, "qdrant_embedding_model", _DEFAULT_EMBEDDING_MODEL)
+    vector_size = _vector_size_for_model(model)
+
+    client.delete_collection(collection_name)
+
+    _ensure_collection(client, collection_name, vector_size)
+    return QdrantCollection(client, collection_name, model=model, vector_size=vector_size)

mempalace/cli.py

@@ -983,7 +983,8 @@ def cmd_repair(args):
         sys.exit(1)
 
     print(f"\n  Repair complete. {filed} drawers rebuilt.")
-    print(f"  Backup saved at {backup_path}")
+    if backup_path:
+        print(f"  Backup saved at {backup_path}")
     print(f"\n{'=' * 55}\n")
 
 
@@ -1029,7 +1030,8 @@ def cmd_compress(args):
     from .dialect import Dialect
     from .palace import get_closets_collection
 
-    palace_path = os.path.expanduser(args.palace) if args.palace else MempalaceConfig().palace_path
+    cfg = MempalaceConfig()
+    palace_path = os.path.expanduser(args.palace) if args.palace else cfg.palace_path
 
     # Load dialect (with optional entity config)
     config_path = args.config

mempalace/config.py

@@ -322,9 +322,43 @@ def tunnel_file(self):
 
     @property
     def collection_name(self):
-        """ChromaDB collection name."""
+        """Vector store collection name."""
         return self._file_config.get("collection_name", DEFAULT_COLLECTION_NAME)
 
+    @property
+    def vector_backend(self):
+        """Vector store backend: 'chroma' (default) or 'qdrant'."""
+        env_val = os.environ.get("MEMPALACE_VECTOR_BACKEND")
+        if env_val:
+            return env_val.lower()
+        return self._file_config.get("vector_backend", "chroma")
+
+    @property
+    def qdrant_url(self):
+        """Qdrant server URL."""
+        env_val = os.environ.get("MEMPALACE_QDRANT_URL")
+        if env_val:
+            return env_val
+        return self._file_config.get("qdrant_url", None)
+
+    @property
+    def qdrant_api_key(self):
+        """Qdrant API key."""
+        env_val = os.environ.get("MEMPALACE_QDRANT_API_KEY")
+        if env_val:
+            return env_val
+        return self._file_config.get("qdrant_api_key", None)
+
+    @property
+    def qdrant_embedding_model(self):
+        """fastembed model name for Qdrant embeddings."""
+        env_val = os.environ.get("MEMPALACE_QDRANT_EMBEDDING_MODEL")
+        if env_val:
+            return env_val
+        return self._file_config.get(
+            "qdrant_embedding_model", "sentence-transformers/all-MiniLM-L6-v2"
+        )
+
     @property
     def people_map(self):
         """Mapping of name variants to canonical names."""

mempalace/convo_miner.py

@@ -18,13 +18,13 @@
 from typing import Optional
 
 from .normalize import normalize
+from .vector_store import get_collection as _get_vector_store_collection
 from .palace import (
     NORMALIZE_VERSION,
     SKIP_DIRS,
     _metadata_matches_extract_mode,
     _validate_palace_fts5_after_mine,
     file_already_mined,
-    get_collection,
     mine_lock,
     mine_palace_lock,
     prefetch_mined_set,
@@ -603,7 +603,7 @@ def _mine_convos_impl(
         print("  DRY RUN — nothing will be filed")
     print(f"{'-' * 55}\n")
 
-    collection = get_collection(palace_path) if not dry_run else None
+    collection = _get_vector_store_collection(palace_path, create=True) if not dry_run else None
 
     # Bulk pre-fetch already-mined set in one paginated pass instead of
     # `len(files)` separate WHERE-source_file queries. On a 150k-drawer