RAG & Vector Stores

A RAG vector's store is chosen when the vector is created and cannot be changed afterwards. It decides where chunks live, what a query costs, and — for the graph and tree stores — what "retrieval" even means.

Tool id	Vendor	Credentials	Status	Divinci usage
cloudflare-v2 CloudFlare is a tool for vectorizing text.	Cloudflare	Divinci-managed	available	—
couchbase-byok Bring Your Own Couchbase Capella account for hyperscale vector search with billion-vector capability.	Couchbase	BYOK	available	—
divinci-pageindex Reasoning-based RAG using hierarchical tree indexes. No vector database or embeddings needed. Hosted by Divinci.	Divinci AI	Divinci-managed	available	—
lightrag Knowledge graph-enhanced RAG with entity extraction and community-based retrieval. Builds on Microsoft GraphRAG research with Cloudflare Vectorize.	Microsoft	Divinci-managed	available	—
mongodb-atlas Store and query vector embeddings directly in MongoDB Atlas with native vector search capabilities. Supports unlimited dimensions and integrates seamlessly with existing MongoDB data.	MongoDB	BYOK	available	—
neo4j-hybrid Hybrid graph + vector RAG using Neo4j. Combines vector similarity with knowledge graph traversal for multi-hop retrieval.	Neo4j	BYOK	available	—
pageindex Reasoning-based RAG using hierarchical tree indexes. No vector database or embeddings needed.	VectifyAI	BYOK	available	—
pinecone-serverless Bring your own Pinecone account. Each vector becomes a namespace inside a Divinci-managed serverless index sized to your embedding model.	Pinecone	BYOK	available	—
qdrant-cosine Uses Qdrant collections to store and query chunks.	Qdrant	BYOK	available	—
raptor Recursive Abstractive Processing for Tree-Organized Retrieval. Clusters chunks hierarchically, summarizes each level, and searches across all abstraction layers.	Divinci AI	Divinci-managed	available	—
redis-vector-search High-performance vector similarity search using Redis Query Engine. Supports HNSW and FLAT indexing with cosine, L2, and inner product distance metrics.	Redis	BYOK	available	—
turso-libsql Uses Turso (hosted libSQL/SQLite) tables with DiskANN vector indexes to store and query chunks.	Turso	BYOK	available	—
vertex-ai-vector-search Google Cloud's enterprise vector search built on ScaNN algorithm. Supports billions of vectors with low latency.	Google	Divinci-managed	available	—
vertex-ai-vector-search-v2 Google Cloud's next-gen vector search with Collections API. Fast setup (~5 min), self-tuning, supports billions of vectors. Beta.	Google	Divinci-managed	available	—

Usage figures for this category are not collected yet — the weekly job populates them.

Three families, not one list

The table above mixes three genuinely different things:

Plain vector stores — Cloudflare Vectorize, Qdrant, Pinecone, Redis, MongoDB Atlas, Vertex AI Vector Search, Turso, Couchbase. Chunks in, nearest neighbours out. Pick on where your data is allowed to live and whose bill it lands on.

Tree / hierarchical — RAPTOR builds a summary tree over your chunks and can retrieve at a level rather than a chunk, which helps on questions that span a whole document. PageIndex is vectorless: it reasons over a document tree instead of embedding, so there is no similarity threshold to tune.

Graph-enhanced — Neo4j Hybrid and LightRAG extract entities and relations and blend a graph traversal into retrieval. Worth it when the answer depends on how things relate rather than on which passage is most similar. Expect a slower, more expensive ingest in exchange.

A store choice is a one-way door for that vector

Changing vectorIndexTool on an existing vector does not move a single chunk. The supported migration is: create a new vector on the new store, re-ingest, and swap the release over — which is what a connector's draft-and-swap does for you.

Choosing

| You want | Reach for | | --- | --- | | the default that needs no account of your own | Cloudflare Vectorize | | your data to stay in infrastructure you own | Qdrant, Redis, MongoDB Atlas, Couchbase (all BYOK) | | questions that span whole documents | RAPTOR or PageIndex | | answers about relationships between entities | Neo4j Hybrid or LightRAG | | an existing vector database you already pay for | Pinecone, Qdrant, Vertex AI Vector Search |

Retrieval is not only the vector leg

Group retrieval runs a keyword (full-text) leg alongside the vector leg and fuses the results. That matters more than it sounds: on our own corpora lexical retrieval has measured markedly higher recall than the vector leg alone on questions containing exact terms, names or numbers. If a corpus answers well on paraphrases but badly on precise product names, the keyword leg is what you are missing.

See also

• Embeddings — the other half of a vector's identity.
• RAG Knowledge Base — creating and filling vectors.
• Knowledge (RAG) Settings — how a release resolves which vector to search.