Skip to main content

RAG & Chat

Ground answers in your own content. Turing's retrieval-augmented generation pipeline pulls the most relevant passages, sharpens them with a pluggable reranker (LLM, cross-encoder or Cohere), and streams a cited answer token by token. Every citation traces back to the exact source passage — provider-agnostic provenance built on Turing's own retrieval metadata, not a vendor's citation API.

Answers stay honest by construction: a configurable relevance gate drops weak matches, an optional groundedness check appends a low-confidence caveat to anything it can't support, and a deterministic refusal fires when a question simply isn't covered by your data. Embeddings can use the built-in Lucene vector store or your engine's own vectors, and conversation memory keeps long chats coherent.