vector-mcp 2.1.0__tar.gz → 2.1.2__tar.gz

{vector_mcp-2.1.0 → vector_mcp-2.1.2}/AGENTS.md

@@ -179,13 +179,24 @@ async def my_tool(param: str) -> str:
 ## Vector Database Backends
 
 ### Supported Backends
-- **ChromaDB**: Local filesystem-based vector database (no container required)
+- **Epistemic-Graph** (**DEFAULT**): the native local AI-native engine — durable
+  (redb-authoritative) with a native ANN/HNSW index, so semantic search is the engine's
+  O(log N) vector search. Zero external infra (autostarts). Selected when `db_type` is
+  unspecified (override with `VECTOR_DB_TYPE`).
+- **ChromaDB**: Local filesystem-based vector database (no container required). Opt-in via the `[chromadb]` extra.
 - **PostgreSQL/PGVector**: PostgreSQL with pgvector extension (container required)
 - **MongoDB**: MongoDB with manual cosine similarity calculation (container required)
 - **Qdrant**: Qdrant vector database (container required)
 - **Couchbase**: Couchbase with REST API fallback and manual cosine similarity (container required, partially functional)
 
 ### Implementation Notes
+- **Epistemic-Graph**: stores each chunk as an engine node (text in the indexed
+  `description` property, plus metadata) with its embedding in the engine's native ANN
+  index via `SyncEpistemicGraphClient` (`.nodes.add` + `.graph.add_embedding`). Semantic
+  search is `.graph.semantic_search` (native ANN); lexical/keyword search is the engine's
+  one-round-trip `.graph.discover` (keyword overlap + semantic, scalable — a client-side
+  term scan is only a fallback for an engine too old for `discover`). A collection maps to
+  an engine graph (`graph_name`); embeddings come from the shared `create_embedding_model`.
 - **MongoDB**: Uses raw MongoClient instead of MongoDBAtlasVectorSearch to avoid authentication issues with local test containers. Implements manual cosine similarity calculation for semantic search.
 - **Couchbase**: Uses simple client approach with REST API fallback to bypass SDK authentication issues. Implements manual cosine similarity calculation for semantic search. Core search functionality working (10/14 tests passing), CRUD operations limited by N1QL service configuration.
 - **PostgreSQL**: Uses native PGVector with proper JSONB querying for get_documents_by_ids.

{vector_mcp-2.1.0 → vector_mcp-2.1.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: vector-mcp
-Version: 2.1.0
+Version: 2.1.2
 Summary: Integrate RAG into AI Agents via MCP Server. Supports multiple Vector database technologies.
 Author-email: Audel Rouhi <knucklessg1@gmail.com>
 License: MIT
@@ -83,7 +83,7 @@ Dynamic: license-file
 ![PyPI - Wheel](https://img.shields.io/pypi/wheel/vector-mcp)
 ![PyPI - Implementation](https://img.shields.io/pypi/implementation/vector-mcp)
 
-*Version: 2.1.0*
+*Version: 2.1.2*
 
 ---
 

{vector_mcp-2.1.0 → vector_mcp-2.1.2}/README.md

@@ -20,7 +20,7 @@
 ![PyPI - Wheel](https://img.shields.io/pypi/wheel/vector-mcp)
 ![PyPI - Implementation](https://img.shields.io/pypi/implementation/vector-mcp)
 
-*Version: 2.1.0*
+*Version: 2.1.2*
 
 ---
 

{vector_mcp-2.1.0 → vector_mcp-2.1.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "vector-mcp"
-version = "2.1.0"
+version = "2.1.2"
 description = "Integrate RAG into AI Agents via MCP Server. Supports multiple Vector database technologies."
 readme = "README.md"
 classifiers = [ "Development Status :: 5 - Production/Stable", "License :: Public Domain", "Environment :: Console", "Operating System :: POSIX :: Linux", "Programming Language :: Python :: 3",]

{vector_mcp-2.1.0 → vector_mcp-2.1.2}/vector_mcp/agent_server.py

@@ -4,7 +4,7 @@ import os
 import sys
 import warnings
 
-__version__ = "2.1.0"
+__version__ = "2.1.2"
 
 logging.basicConfig(
     level=logging.INFO,

{vector_mcp-2.1.0 → vector_mcp-2.1.2}/vector_mcp/mcp_server.py

@@ -37,7 +37,7 @@ from starlette.responses import JSONResponse
 from vector_mcp.auth import get_client
 from vector_mcp.vector_api import Api
 
-__version__ = "2.1.0"
+__version__ = "2.1.2"
 
 logger = get_logger(name="vector-mcp")
 logger.setLevel(logging.INFO)

{vector_mcp-2.1.0 → vector_mcp-2.1.2}/vector_mcp/vectordb/epistemic_graph.py

@@ -30,9 +30,12 @@ with optional_import_block():
 
 logger = get_logger(__name__)
 
-# Node property key the chunk text is stored under (metadata takes the rest).
-_CONTENT_KEY = "content"
-_RESERVED_KEYS = {_CONTENT_KEY, "label", "id", "_similarity"}
+# The chunk text is stored under the engine's standard ``description`` property: it is
+# the field the engine keyword-indexes and that ``graph.discover`` hydrates, so lexical
+# search runs engine-side (scalable) instead of an O(N) client scan. Everything not in
+# the reserved set is treated as user metadata.
+_TEXT_KEY = "description"
+_RESERVED_KEYS = {_TEXT_KEY, "label", "id", "_similarity", "name", "type", "score"}
 
 
 @require_optional_import(["epistemic_graph", "agent_utilities"], "epistemic-graph")
@@ -107,7 +110,7 @@ class EpistemicGraphVectorDB(VectorDB):
     @staticmethod
     def _props_to_doc(node_id: str, props: dict | None) -> Document:
         props = dict(props or {})
-        content = props.pop(_CONTENT_KEY, "") or ""
+        content = props.get(_TEXT_KEY, "") or ""
         meta = {k: v for k, v in props.items() if k not in _RESERVED_KEYS}
         return Document(id=node_id, content=content, metadata=meta, embedding=None)
 
@@ -173,7 +176,7 @@ class EpistemicGraphVectorDB(VectorDB):
             embedding = doc.get("embedding")
             if embedding is None:
                 embedding = self._embed(text)
-            client.nodes.add(doc_id, {_CONTENT_KEY: text, "label": "Document", **meta})
+            client.nodes.add(doc_id, {_TEXT_KEY: text, "label": "Document", **meta})
             client.graph.add_embedding(doc_id, [float(x) for x in embedding])
 
     def update_documents(
@@ -271,27 +274,66 @@ class EpistemicGraphVectorDB(VectorDB):
         n_results: int = 10,
         **kwargs: Any,
     ) -> QueryResults:
+        """Keyword search over the engine's scalable index.
+
+        Routes through the engine's one-round-trip ``discover`` op — it ranks nodes by
+        keyword overlap over the indexed text (the chunk text lives in ``description``)
+        AND semantic similarity, returning a hydrated top-k. Falls back to a client-side
+        term scan only if the engine is too old to support ``discover``.
+        """
         client = self._client_for(collection_name)
         results: QueryResults = []
         for query in queries:
+            terms = list(dict.fromkeys(t for t in query.lower().split() if t))
+            try:
+                q_emb = self._embed(query, is_query=True)
+            except Exception:
+                q_emb = []  # discover degrades to a bounded keyword-only scan
             try:
-                hits = client.graph.match_ontology_terms(query) or []
+                hits = client.graph.discover(terms, q_emb, n_results) or []
             except Exception as e:
-                logger.info(f"lexical search: {e}")
-                hits = []
+                logger.info(f"discover unavailable, using client scan: {e}")
+                results.append(self._lexical_scan(client, terms, n_results))
+                continue
             query_result: list[tuple[Document, float]] = []
             for item in hits[:n_results]:
                 if isinstance(item, dict):
                     node_id = str(item.get("id", ""))
-                    score = float(item.get("score", 1.0) or 1.0)
+                    score = float(item.get("score", 0.0) or 0.0)
+                    # discover hydrates text in-band; fetch full props for metadata.
+                    try:
+                        props = client.nodes.properties(node_id) or item
+                    except Exception:
+                        props = item
                 else:
-                    node_id, score = str(item), 1.0
+                    node_id, score, props = str(item), 0.0, None
                 if not node_id:
                     continue
-                try:
-                    props = client.nodes.properties(node_id)
-                except Exception:
-                    props = None
                 query_result.append((self._props_to_doc(node_id, props), score))
             results.append(query_result)
         return results
+
+    def _lexical_scan(
+        self, client: Any, terms: list[str], n_results: int
+    ) -> list[tuple[Document, float]]:
+        """Fallback term-frequency scan (only when the engine lacks ``discover``)."""
+        try:
+            listing = client.nodes.list() or []
+        except Exception as e:
+            logger.info(f"lexical scan (list nodes): {e}")
+            return []
+        scored: list[tuple[Document, float]] = []
+        for entry in listing:
+            node_id = str(entry[0]) if isinstance(entry, list | tuple) else str(entry)
+            try:
+                props = client.nodes.properties(node_id) or {}
+            except Exception:
+                continue
+            content = str(props.get(_TEXT_KEY, "") or "").lower()
+            if not content:
+                continue
+            freq = sum(content.count(t) for t in terms)
+            if freq > 0:
+                scored.append((self._props_to_doc(node_id, props), float(freq)))
+        scored.sort(key=lambda x: x[1], reverse=True)
+        return scored[:n_results]

{vector_mcp-2.1.0 → vector_mcp-2.1.2}/vector_mcp.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: vector-mcp
-Version: 2.1.0
+Version: 2.1.2
 Summary: Integrate RAG into AI Agents via MCP Server. Supports multiple Vector database technologies.
 Author-email: Audel Rouhi <knucklessg1@gmail.com>
 License: MIT
@@ -83,7 +83,7 @@ Dynamic: license-file
 ![PyPI - Wheel](https://img.shields.io/pypi/wheel/vector-mcp)
 ![PyPI - Implementation](https://img.shields.io/pypi/implementation/vector-mcp)
 
-*Version: 2.1.0*
+*Version: 2.1.2*
 
 ---