Integrations

Each integration is an optional extra, so the core stays free of dependencies. The OPA, OpenFGA, SpiceDB, Cerbos and Keto authorizers need no extra at all: they use the standard library’s HTTP client.

Integration	Kind	Install	Pre-filters
PostgreSQL + pgvector	Backend	wardowl[pgvector]	Yes, inside the SQL query
Chroma	Backend	wardowl[chroma]	Yes, as a where clause
Elasticsearch	Backend	wardowl[elasticsearch]	Yes, inside the kNN search
OpenSearch	Backend	wardowl[opensearch]	Yes, inside the k-NN search
Qdrant	Backend	wardowl[qdrant]	Yes, as a query_filter
Milvus	Backend	wardowl[milvus]	Yes, inside the vector search, with parameterized filters
MongoDB Atlas Vector Search	Backend	wardowl[mongodb]	Yes, as the $vectorSearch pre-filter
Weaviate	Backend	wardowl[weaviate]	Yes, inside the vector search
Redis	Backend	wardowl[redis]	Yes, in the same KNN query
LanceDB	Backend	wardowl[lancedb]	Yes, before the vector search ranks
FAISS	Backend, in process	wardowl[faiss]	Yes, an ID selector inside the search
Vespa	Backend	wardowl[vespa]	Yes, in the YQL query, inside the nearest-neighbor search
Apache Cassandra	Backend	wardowl[cassandra]	Yes, in each ANN query’s WHERE
Pinecone	Backend	wardowl[pinecone]	Yes, as the query’s metadata filter
Open Policy Agent	Authorizer	none	When the policy returns a filter
OpenFGA	Authorizer	none	Yes, through ListObjects
SpiceDB	Authorizer	none	Yes, through LookupResources
Cerbos	Authorizer	none	Yes, from query plans
Cedar	Authorizer, in process	wardowl[cedar]	No: post-filters, every candidate decided by Cedar
Casbin	Authorizer, in process	wardowl[casbin]	No: post-filters, every candidate decided by Casbin
Ory Keto	Authorizer	none	No: post-filters, Keto has no reverse lookup
Google Drive	Permission sync onto the chunks	none (bring Google’s client)	Through MetadataACLAuthorizer
LangChain	Framework, and any LangChain vector store as a backend	wardowl[langchain]	Wrapped stores post-filter
LlamaIndex	Framework, and any LlamaIndex index as a backend	wardowl[llamaindex]	Wrapped indexes post-filter
Haystack	Framework	wardowl[haystack]	Through any wardowl backend
MCP	Agent tool server	wardowl[mcp]	Through any wardowl backend
Agent frameworks	OpenAI Agents SDK, Pydantic AI, CrewAI, AutoGen, Semantic Kernel, DSPy, through MCP	none (the framework’s MCP client)	Through any wardowl backend
HTTP service	Retrieval for any language, OIDC tokens	wardowl[server]	Through any wardowl backend
FastAPI	Reference RAG service: verified tokens, tenants, safe errors	none (your app’s FastAPI)	Through any wardowl backend
OpenTelemetry	Observer	wardowl[otel]

Choosing a store

Every store applies the permission filter inside its own search, passes the conformance kit in CI against the real store, and keeps permission values out of query text. They differ in what is worth knowing in production:

Store	A write is searchable	Native aretrieve	Permission values travel as	Keyword search	CI tests against
PostgreSQL + pgvector	When add commits	AsyncConnectionPool	Bound SQL parameters (jsonb containment)	Full-text, mode="keyword"	PostgreSQL 17 with pgvector
Chroma	When add returns	No (a thread)	A where dictionary	No	Chroma’s in-process client
Elasticsearch	When add returns (refresh="wait_for")	AsyncElasticsearch	A terms query on type-tagged terms	BM25, mode="keyword"	9.5.5
OpenSearch	When add returns (refresh="wait_for")	AsyncOpenSearch	A terms query on type-tagged terms	BM25, mode="keyword"	3.9.0
Qdrant	When add returns (wait=True)	AsyncQdrantClient	Filter objects (MatchAny)	BM25 (IDF modifier), mode="keyword"	1.19.2
Milvus	At once (strong consistency, the default)	AsyncMilvusClient	filter_params, never in the expression	BM25, mode="keyword"	3.0.2
MongoDB Atlas Vector Search	When add returns: it waits for the search index	AsyncMongoClient	A BSON filter ($in)	BM25 (Atlas Search), mode="keyword"	Atlas Local 8.0.30
Weaviate	When add returns	WeaviateAsyncClient	Filter objects (contains_any)	BM25, mode="keyword"	1.39.12
Redis	When add returns	redis.asyncio	Hex-encoded TAG values	BM25, mode="keyword"	8.10.2
LanceDB	When add returns	No (a thread)	Hex-encoded list values (array_has_any)	BM25, mode="keyword"	LanceDB 0.40, embedded
FAISS	At once	No (a thread)	matches over the stored metadata, then an ID selector	None	faiss-cpu 1.15, in process
Vespa	When add returns	No (a thread)	Hex-encoded terms in YQL (in)	BM25, mode="keyword"	8.763.13
Apache Cassandra	When add returns, at the session’s consistency level	No (a thread)	Bound CONTAINS parameters, one ANN query per AND	No	5.0.9
Pinecone	Within seconds; at once with wait=	No (a thread)	$in on hex-encoded terms, $and, $or	No	Pinecone serverless, API 2026-07

Keyword and vector search combine into hybrid search with RankFusion, across stores as well as within one. Measured recall, latency and throughput at a million chunks are in Benchmarks for pgvector, Qdrant, Elasticsearch, OpenSearch and Weaviate.

Backends and authorizers combine freely: pgvector with OpenFGA, Chroma with OPA, a LangChain store with permissions on the chunks. A store not listed here needs only a search method; see Extending wardowl.