Tender Intelligence Console

A production-style tender QA + bid/no-bid copilot for European procurement PDFs.

Built for an anonymized industrial client (codename Aurora Works) who wanted faster, evidence‑based tender reviews without losing traceability.

Demo

What it does

• Ingest tender PDFs (drag & drop into chat or batch import)
• Parse tenders into sections (buyer, scope, lots, deadlines, requirements, submissions) with collapsible UI
• Assess readiness (can we qualify/deliver) using tender requirements plus your company profile, compliance, and delivery history (Postgres)
• Find similar tenders with evidence-backed matches and snippets
• Answer with traceable citations (file#page) for every factual claim

Why it matters

Bid teams waste hours manually scanning PDFs and still miss constraints. This console turns a tender into a decision‑ready brief with explicit evidence and next steps.

Highlights (portfolio‑worthy)

• LangGraph tool-calling agent: model routes to the right tool (breakdown, readiness, similarity, chat history)
• Evidence-first answers: every factual claim is grounded in retrieved excerpts with citations
• Semantic chunking + query rewrite for higher retrieval precision on long documents
• Server-side chat memory (threaded sessions) + retrieval controls (depth, source filter)
• Full product surface: API + web UI + ingestion pipeline + DB seed + Docker Compose

Tech stack

• Backend: FastAPI, LangChain, LangGraph
• RAG: Pinecone (vector DB), OpenAI embeddings
• LLM: OpenAI Chat Completions (configurable)
• DB: Postgres (company profile + delivery history)
• Frontend: Next.js
• Ops: Docker Compose, .env configuration, structured logging

Quickstart (Docker)

1. Create agent/src/.env with your secrets:

OPENAI_API_TOKEN=...
PINECONE_TOKEN=...
PINECONE_INDEX=...
PINECONE_NAMESPACE=tenders

2. Start the stack:

cd agent
docker compose up --build

3. Open:

• Web UI: http://localhost:3000
• API: http://localhost:8000

Typical workflow

1. Put tender PDFs into agent/src/data/tenders/raw/ (optional if you attach via chat)
2. Chunk + embed + upsert to Pinecone (CLI)
3. Ask questions like:

• “Break down the tender by scope, deadlines, requirements, and submission details.”
• “What are the qualification thresholds and what evidence do we need?”
• “Find similar tenders and explain the match.”
• “Can we qualify? Give a decision-ready assessment with gaps and next steps.”

Architecture (high level)

Next.js UI → FastAPI → LangGraph Agent (tools) → {Pinecone, Postgres, PDF parser} → response (with sources)

Repo layout

• agent/ — the full application (API, UI, ingestion scripts, DB seed)

For detailed CLI commands (chunking, embeddings, Pinecone upload), see agent/README.md.

Data & privacy notes

• Secrets are intentionally excluded from git (.env is ignored).
• Intermediate artifacts (chunk/embedding dumps) should not be committed.
• The company profile and deliveries included here are synthetic demo data to showcase readiness checks.

License

Apache‑2.0 (see LICENSE).