RoleFit

How it works

Hybrid retrieval. Dense embeddings match "owns the system end to end" against "full ownership of the stack" where no words overlap. They are bad at rare exact tokens, because LangGraph, pgvector and MCP sit near each other and near a hundred other tool names. Postgres full-text is the reverse. Both run, and the rankings fuse with Reciprocal Rank Fusion, which needs only each retriever's ordering and so avoids weighting cosine distance against ts_rank.

It grades its own retrieval. Retrieve-then-generate fails quietly: if retrieval misses, the model still writes a confident paragraph out of training data. Here a grader reads the chunks first and decides whether they can answer the question at all.

retrieve -> grade -+-- relevant ------> generate -> END
                   +-- not relevant --> rewrite -> retrieve   (max 2)
                   +-- out of attempts -> admit_gap -> END

The admit_gap branch is the point. Ask it something the corpus does not cover and it says so instead of inventing an answer.

FastAPI and LangGraph on Vercel, Postgres with pgvector on Supabase. Embeddings run on gte-small inside a Supabase Edge Function, because Groq has no embeddings endpoint and retrieval needs one anyway. Llama 3.3 70B on Groq generates and grades. API health