Retrieval Advisor

Retrieval-augmented advisor for personal decisions that answers with one next move, its reasoning and the source passage it came from.

In progress

The need

General chatbots hedge. This one reasons over a curated, weighted corpus and commits to a concrete next step with a citation, and it learns from which answers the user accepted.

What was hard

  1. 1

    Weighted source registry

    A registry of sources, tagged by track and stage, drives which passages get retrieved.

  2. 2

    Citation grounding

    Every answer exposes the passages it relied on.

  3. 3

    Reply feedback loop

    Answers the user accepted feed additions back into the corpus.

  4. 4

    LLM-as-judge evaluation

    An evaluation harness scores both retrieval and answers using a model as judge.

  5. 5

    Atomic turn claiming

    Postgres RPCs claim each conversation turn atomically so concurrent conversations do not collide.

  6. 6

    Usage and cost tracking

    Usage and cost are recorded for every request.

Stack

Next.js 16React 19TypeScriptZodSupabase PostgrespgvectorVoyage embeddingsClaude OpusClerkPythonVercel