RFP AI — AI‑Assisted RFP Drafting (MVP)

End‑to‑end system to ingest firm Q&A knowledge, retrieve the most relevant content for an RFP question, and generate a concise draft answer.

• Stack: React (Vite) UI, Fastify + TypeScript API, Postgres + pgvector, Ollama for local embeddings + text generation, optional Hugging Face cross‑encoder reranker.
• Focus: Practical RAG with transparent retrieval, tight feedback loop, and local‑first privacy.

Features

• Retrieval‑Augmented Generation: Embed Q&A chunks, vector search, optional cross‑encoder rerank, then generation.
• Ingestion two ways:
  • UI: paste/upload .txt, auto‑extract Q&As, chunk, ingest.
  • CLI: .mjs scripts for documents and CSV.
• Controls: Filter by product, jurisdiction, audience; inject firm facts to keep numbers consistent.
• Local‑First: All inference runs locally via Ollama; Postgres via Docker.

Architecture

• Frontend (rfp-ui/): React + Vite app with two modes:
  • Draft: ask a question and view the generated draft.
  • Ingest: upload/paste .txt, auto‑extract Q&As, review, ingest all.
• Backend (server/): Fastify + TypeScript service exposing:
  • POST /ingest/chunks: upsert Q&A metadata and embeddings.
  • POST /rfp/answer/batch: retrieve → (optional rerank) → generate drafts for multiple questions in one request.
• Data: Postgres with pgvector for embeddings. Firm facts are read from firm_facts and injected at generation time.
• Models: Ollama provides nomic-embed-text for embeddings and llama3.1:8b-instruct-q4_K_M for drafting. Optional rerank via @huggingface/transformers using cross-encoder/ms-marco-MiniLM-L-6-v2 (can be toggled off).

Ingest Data

• UI:
  • In top command bar, type ingest.
  • Upload/paste .txt, click Parse, review, Ingest All.
• CLI:
  • TXT: node server/ingest_doc.mjs test-rfp.txt --product=Core+ --audience=pension --jurisdiction=US --no-confirm
  • CSV: node server/ingest_csv.mjs server/qas_ingest.csv

Ask a Question

• In Draft mode (default), type a question and click Suggest.
• Review the draft. Copy, edit, ship.

API Surface

• POST /ingest/chunks
  • Body: { qa: { id, question, answer, meta? }, chunks: string[] }
• POST /rfp/answer/batch
  • Body: { questions: string[], filters?: { product?, jurisdiction?, audience? } }
  • Returns: { items: [{ question, draft }] }

Key Files

• server/server.ts: API routes, embedding/generation helpers, optional reranker.
• server/ingest_doc.mjs: Parse .txt into Q&As and ingest (CLI).
• server/ingest_csv.mjs: Bulk ingest from CSV.
• rfp-ui/src/views/DraftView.tsx: Drafting UX + retrieved snippets.
• rfp-ui/src/views/IngestView.tsx: Ingest UX with parsing/preview.
• schema.sql: Reference schema and diagnostic queries.
• docker-compose.yml: Postgres (pgvector) and MinIO (future file storage).

Roadmap

• Inline source preview and jump‑to‑QA.
• Human‑in‑the‑loop editing workflow and approvals.
• Richer filters (tags), draft history, and audit trails.
• Document storage (MinIO/S3) and PDF extraction pipeline.