Response Assistant

A private evidence-grounded response workflow for preparing drafts from a curated public source library. The current build intentionally stops at evidence review before AI drafting.

Problem

Useful responses need to be fast, but speed makes it easy to lose source context, repeat stale information, or write a confident claim that the evidence cannot support.

My role

Product direction, workflow design, ingestion contracts, provenance rules, structure-aware chunking, persistence design, and the boundary between retrieval and generation.

What is built

A FastAPI backend with source contracts, a parser for repository-contained Markdown, deterministic sections and chunks, versioned evidence persistence, migrations, and a safe dry-run/apply catalogue runner.

Architecture

original source → searchable text → parsed document → structured chunks → versioned evidence records

The source pack currently contains four curated public climate-finance sources.

Evaluation & controls

Validate source and location consistency, preserve original/searchable separation, keep chunk identity document-scoped, and make catalogue writes explicit. Retrieval evaluation comes before vectors, tools, or drafting.

Trade-offs

Lexical retrieval and a deliberately narrow source pack are easier to inspect and test than an early “agent” abstraction. The trade-off is less coverage while the evidence contract is being proved.

Evidence of completion

Private repository with source-pack files, backend modules, migrations, focused tests, and a clean main branch observed on 26 August 2026. Not yet represented as a public production service.

Next step

Define and run the labelled retrieval cases, including paraphrases and no-evidence outcomes, then connect the verified retrieval endpoint to the review surface.