RAG Retrieval
Three search providers share one SearchProvider interface: vector, fulltext,
and hybrid. All return chunks with a Source for provenance.
Providers
- Vector: embed the query, search a vector store, return nearest neighbors
- Fulltext: keyword search over a fulltext index
- Hybrid: run vector and fulltext in parallel, fuse the results
Hybrid fusion
| Strategy | Behavior |
|---|---|
rrf | Reciprocal Rank Fusion (default, k=60) |
weighted | w_v * norm(vec) + w_f * norm(ft) |
cascade | Fulltext recall, then vector rerank |
Embeddings and rerankers
Embeddings: fake, OpenAI text-embedding-3-small, bge-m3 (local HTTP),
Ollama nomic-embed-text (local). Rerankers: fake, Cohere, bge-reranker-v2-m3.
Indexing
The indexing pipeline splits documents with a ChunkStrategy (recursive,
token, sentence, markdown, none) and writes with an index_sync mode
(dual, vector_only, fulltext_only).
hanflow index my-store ./docsThe Knowledge node consumes retrieval through ctx.retrieve(store, query, top_k, rerank, filter).