Research Ops Copilot (Local‑First RAG)

A production‑style, local‑first RAG MVP with SQL metadata + vector store, page‑level citations, observability, and a small evaluation harness.

Setup (macOS)

1. Start Postgres

docker compose up -d

2. Create tables (once)

psql postgresql://rocp:rocp_pw@localhost:5432/rocp_db -f db/schema.sql

3. Create a Python env + install deps

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Optional: set OpenAI key for embeddings + generation

export OPENAI_API_KEY=YOUR_KEY

Ingest documents

python -m ingestion.ingest --paths SampleData/pdfs SampleData/docs

Query (CLI)

python -m rag.query --query "What is this document about?"

Evaluation

python -m evaluation.run_eval --eval_file evaluation/examples/eval_demo.jsonl

Notes on Design

• SQL for metadata & traceability: documents, ingestion_runs, and chunks provide a durable audit trail and idempotency.
• Vector DB for retrieval: Chroma persists embeddings locally and stores chunk‑level metadata for filtering and citations.
• Idempotency: documents are de‑duplicated by uri + file_hash/text_hash; re‑ingest skips unchanged files.
• Grounded answers: output includes citations with doc title, uri, page number, and chunk id.
• Confidence threshold: if top score < threshold, the system refuses with “I don’t know.”