Use cases / RAG retrieval
Retrieval for RAG that finds the right chunk, not just a similar one.
Vectors miss exact terms: error codes, part numbers, names. Keyword search misses paraphrases. WarpSearch runs both in one query against one index and fuses the lists, so your model gets the passages that actually answer the question.
What a RAG backend needs, in one place
Hybrid retrieval
BM25 and vector search in one request, merged with reciprocal rank fusion or weighted linear fusion. Elasticsearch rrf retrievers and OpenSearch hybrid queries work too.
Filters inside the search
Tenant, date, source or language filters apply during the vector search, not after it, so you still get k results.
Permissions from your data
A lookup query joins a user's groups to document ACLs in the same request. See Graph search.
Smaller vectors
Matryoshka truncation, float16, int8 and bit encodings cut storage without re-embedding. See Dimensions and storage.
Built for agents
An MCP server lets agents create collections, ingest and search. See the guide for AI agents.
Cheap when idle
A knowledge base queried a few times an hour pays per request, not for a cluster. Idle collections cost only their storage.
What a small RAG backend costs
1M chunks with 768-dim embeddings and 1M queries a month, at list prices in US East.
| Platform | Small RAG |
|---|---|
| WarpSearch | $10 |
| turbopuffer | $161.6x cheaper |
| Weaviate | $48 to $1434.8 to 14x cheaper |
| Elastic Serverless | About $100 to $21010 to 21x cheaper |
| Pinecone | $118 to $13712 to 14x cheaper |
| OpenSearch Service | $132 to $39713 to 40x cheaper |
| OpenSearch Serverless | $357 to $53036 to 53x cheaper |
| Aiven | $365 to $73037 to 73x cheaper |
| Small RAG | 1M text chunks with 768-dim embeddings, 7 GB of JSON. 1M queries a month spread through the day. 1 GB written a month. |
|---|
List prices in US dollars, checked October 20, 2026. Ranges cover regions, plan choices, or how the vendor sizes compute. All comparisons and sources.
Good to know
You compute embeddings yourself: WarpSearch does not embed text or rerank results for you. Send vectors with your documents and the query, and rerank the top results in your application if you use a cross-encoder.
Questions
Why use hybrid search for RAG?
Vector search finds passages with similar meaning but often misses exact identifiers, names and rare terms. BM25 finds those but misses paraphrases. Fusing both lists with RRF gives the model better context than either alone.
Do I need a separate vector database?
No. Vectors are fields on the same documents as your text and metadata, so one write keeps both in sync and one query searches both.
Does WarpSearch generate embeddings?
No. Compute embeddings with the model you choose and send them with your documents and queries.
Read next
Free up to 1 GB, no card. Pay per GB and per request after that.