Anchor

Retrieval that says no source instead of guessing.

  • TypeScript
  • Next.js
  • PostgreSQL
  • pgvector
  • Prisma
LivePlaygroundCode

The problem

A search always returns its closest match, even when nothing in the data is a good answer, and a model given a weak match will answer with confidence. Most RAG demos improvise when nothing in the index matches, and in real estate an improvised answer is a legal problem.

How it works

Anchor runs a cosine search on pgvector and returns a null answer when the best hit falls under a floor you set per index.

  • Listings are stored as vectors in PostgreSQL with pgvector and queried through Prisma.
  • The retrieval step, in TypeScript, embeds the query, scores the best match and compares it with the floor. Above it, you get the chunks. Below it, you get { answer: null, reason: "no_match" }.
  • Ingestion upserts on a hash of each chunk's content, so re-running the same documents does not duplicate rows.
  • A Next.js and React page lets you type a question and see the answer next to the listing it came from, or see the refusal.

The floor is a config value, not something the model can talk itself out of. The trade-off is that a floor set too high makes Anchor refuse questions it could half answer, and tuning that number is on you.

What it does not do

  • One threshold per index. There is no per-query re-ranking or keyword pass.
  • The embedding model is fixed per index. Switching models means re-embedding everything.
  • The floor cannot tell "no source" apart from "a source exists but embeds poorly".

Links