BCRA Mini-RAG
BCRA Mini-RAG
TL;DR — I built BCRA Mini-RAG: a small RAG assistant over Argentine FX regulation (BCRA CAMEX). Ask what a circular says and you get a cited clause — Comunicación “A” number plus punto — or honest silencio. It is not a general BCRA chatbot and not legal advice. I wrote the code spec-first with OpenSpec.
Stack — Python 3.11+, FastAPI, Gradio, Chroma, pydantic v2, OpenSpec,
uv
CAMEX is a living corpus: decades of Comunicaciones “A” plus a texto ordenado that is supposed to be the rule hoy. I wanted an assistant that would answer a clause question from that dump and then stop — cite the communication and the punto, or refuse — instead of chatting around the regulation.
The service ingests the official tipo-A catalog (Com. A 13 / 1981 → present) and the current Exterior y Cambios texto ordenado, refreshes when new A’s appear or the TO checksum changes, and retrieves from a Chroma index. Generation is structured — answer, deontic finding, citations. Named Com. A NNNN routes to that document, “vigente / hoy” prefers the TO plus later A’s, and guardrails log every turn. If the model cannot ground a clause, the finding is silencio and the UI abstains.
I built the whole thing spec-driven. Ingest, retrieval, query answering, guardrails, the observatory UI, and L1 evals each started as an OpenSpec change — proposal, requirements, design, and tasks — and only then the implementation. That is the same loop I used on check-unprotected-keys: the ports, the structured generation contract, and the silencio rule were cheap to change on paper, and the code is downstream of those artifacts.
The Gradio observatory has two layouts. Staff (IA) shows reasoning, a citation inspector, the guardrail log, Calidad L1, and dump dates; Usuario keeps question, answer, send, clear, and examples. L1 is an operator harness over a 30-question gold set with citation-id exact as the headline metric — the UI only reads the last static file; chat never scores itself. Answers are an unofficial extract of public bcra.gob.ar documents, dated as of the last dump refresh.
