RFP Intelligence

A RAG-powered chat interface for querying your organisation's RFP documents, case studies, and pitch decks. Upload PDFs, ask questions in plain English, and get cited, synthesised answers grounded in your actual documents.

How it works

Upload PDF → extract + chunk → embed → store in Supabase (pgvector)
                                                    ↓
Query → HyDE expansion → decompose → multi-query retrieval → rerank → GPT-4o synthesis → answer + sources

Processing pipeline

• Text is extracted with pdfplumber; sparse pages (< 100 chars) fall back to GPT-4o Vision
• Repeated lines across 3+ pages (headers, footers, page numbers) are stripped
• Section headings are detected and prepended to the embedding text
• 150-token child chunks are embedded for retrieval; each carries a 500-token parent window used for synthesis

Search pipeline

1. HyDE — generates a hypothetical answer to the query, embeds that instead of the raw question for better semantic alignment
2. Multi-query — decomposes compound queries into up to 4 sub-questions and retrieves for each, then deduplicates
3. Hybrid search — combines vector similarity and full-text search via Supabase RPC with Reciprocal Rank Fusion
4. LLM reranking — GPT-4o-mini scores the merged candidate set; top 5 are kept for synthesis

Stack

Layer	Technology
Frontend	Next.js 15, React, Tailwind CSS
Backend	FastAPI (Python)
Database + vectors	Supabase (PostgreSQL + pgvector)
Embeddings	OpenAI text-embedding-3-small (1536 dims)
Synthesis	OpenAI gpt-4o
Reranking / HyDE	OpenAI gpt-4o-mini
PDF parsing	pdfplumber, PyMuPDF

Project structure

ai-rfp-intelligence/
├── backend/
│   ├── app/
│   │   ├── main.py               # FastAPI app, CORS config
│   │   ├── schemas.py            # Pydantic models
│   │   ├── routers/
│   │   │   ├── auth.py           # Login endpoint
│   │   │   ├── documents.py      # Upload, list, get, delete endpoints
│   │   │   └── search.py         # Query pipeline (HyDE, decompose, rerank, synthesise)
│   │   └── services/
│   │       └── processing.py     # PDF extraction, chunking, embedding
│   └── requirements.txt
└── frontend/
    └── src/
        ├── middleware.ts         # Auth cookie check, redirects to /login
        └── app/
            ├── page.tsx          # Chat UI and document manager
            ├── login/page.tsx    # Login form
            ├── components/       # Sidebar, Header, ChatInput, DocumentManager
            ├── services/api.ts   # API client
            └── types/types.ts    # Shared types

Local setup

Prerequisites

• Python 3.9+
• Node.js 18+
• A Supabase project with pgvector enabled
• An OpenAI API key

Backend

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

Create backend/.env:

OPENAI_API_KEY=sk-...
SUPABASE_URL=https://your-project.supabase.co
SUPABASE_KEY=your-service-role-key
LOGIN_USERNAME=your-username
LOGIN_PASSWORD=your-password
AUTH_TOKEN_SECRET=your-secret

Start the API:

uvicorn app.main:app --reload

The API runs at http://localhost:8000.

Frontend

cd frontend
npm install
npm run dev

The app runs at http://localhost:3000.

API reference

Method	Path	Description
POST	/api/auth/login	Authenticate. Body: { "username": "...", "password": "..." }
POST	/api/search	Query documents. Body: { "query": "..." }
GET	/api/documents	List all uploaded documents
GET	/api/documents/{id}	Get a single document by ID
POST	/api/documents	Upload a PDF. Form fields: file, document_type
DELETE	/api/documents/{id}	Delete a document and its chunks

Supported document types: RFP Document, Case Study, Pitch Deck, Client Presentation

Search response shape:

{
  "query": "What is our experience with federal contracts?",
  "confidence_level": "high",
  "synthesized_response": "...",
  "sources": [
    {
      "filename": "proposal-2024.pdf",
      "locator": "Page 12",
      "similarity_score": 0.021,
      "snippet": "..."
    }
  ]
}