RAG with reranking

Hybrid lexical + vector retrieval, reranked with a cross-encoder. State-of-the-art recall.

What it is

The default retrieval stack in osFoundry combines BM25 lexical search with vector similarity, then reranks the top candidates with a cross-encoder. This out-of-the-box pipeline beats most "single technique" RAG setups for recall and relevance.

Every stage is configurable in osStudio. Swap rerank models. Add MMR for diversity. Plug your own retrieval stages. The default is good; the custom path is wide open.

Key capabilities

BM25 + vector hybrid retrieval
Cross-encoder reranker (voyage-rerank-2 default)
MMR diversity scoring on demand
Per-folder retrieval config in osStudio
Citations rendered inline in the chat

Use cases

Customer support agents searching big KBs
Research assistants pulling from many sources
Internal documentation chat for thousands of pages