FastAPI backend for a document question-answering application built around financial PDFs.

This repository contains the API, document-ingestion pipeline, storage layer, retrieval workflow, and model-provider integrations. The UI lives in StackRAG-Frontend.

The current codebase includes:

• PDF upload and background processing
• metadata extraction, sectioning, chunking, and embeddings
• Supabase Auth, Storage, Postgres, and pgvector integration
• user-scoped vector retrieval
• a chat endpoint that streams server-sent events
• a Python calculation tool for numeric questions
• evaluation scripts and a sample dataset under evaluation/

The evaluation files are tied to their dataset, model configuration, and test run. They are useful for comparing changes; they are not general accuracy or uptime guarantees.

flowchart LR
    Client[Web client or API client] --> API[FastAPI]
    API --> Auth[Supabase Auth]
    API --> Storage[Supabase Storage]
    API --> DB[(Postgres + pgvector)]
    API --> Models[OpenAI or Gemini]
    API --> Pipeline[Document pipeline]
    Pipeline --> Storage
    Pipeline --> DB

1. An authenticated client uploads a PDF.
2. The API creates a processing job and runs the ingestion pipeline in the background.
3. The pipeline parses the file, extracts metadata, creates sections and chunks, and stores embeddings.
4. The client can read processing status and the resulting document records.

1. The client sends conversation history to POST /api/v1/chat/stream.
2. The backend retrieves chunks for the authenticated user.
3. The RAG workflow can call retrieval and calculation tools while composing the response.
4. Response chunks are returned as server-sent events.

All endpoints below the /api prefix require a Supabase access token in Authorization: Bearer <token>. The health endpoint does not require authentication.

Interactive API documentation is available at http://localhost:8000/docs when the server is running.

Example upload:

curl -X POST "http://localhost:8000/api/v1/documents/process" \
  -H "Authorization: Bearer YOUR_SUPABASE_ACCESS_TOKEN" \
  -F "file=@financial_report.pdf"

Example chat request:

curl -N -X POST "http://localhost:8000/api/v1/chat/stream" \
  -H "Authorization: Bearer YOUR_SUPABASE_ACCESS_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"history": [{"kind": "request", "parts": [{"part_kind": "user-prompt", "content": "What was the revenue growth?"}]}]}'

• Python 3.12 or newer
• A Supabase project
• An OpenAI or Gemini API key
• Docker and Docker Compose, if using the container setup

git clone https://github.com/BryanTheLai/StackRAG-Backend.git
cd StackRAG-Backend
python -m venv .venv

Activate the environment, then install dependencies:

# macOS / Linux
source .venv/bin/activate

# Windows PowerShell
\.venv\Scripts\Activate.ps1

pip install -r requirements.txt

Create .env from .env.example and set the values required by your chosen provider and Supabase project:

GEMINI_API_KEY=
OPENAI_API_KEY=
SUPABASE_URL=https://your-project-ref.supabase.co
SUPABASE_ANON_KEY=your-anon-public-key

Start the API from the repository root:

uvicorn src.main:app --reload --host 127.0.0.1 --port 8000

Run the SQL files in scripts/ in numeric order in the Supabase SQL editor. They create the tables, vector search function, storage policies, and processing-job tracking used by the application.

The scripts contain development-oriented reset statements such as DROP TABLE. Review them before running them against an existing database.

docker compose up --build

The compose file exposes the API on port 8000 and loads variables from .env.

api/
  v1/endpoints/       FastAPI route handlers
src/
  llm/                Model clients, tools, and RAG workflow
  services/           Parsing, metadata, sectioning, chunking, embeddings
  storage/            Supabase persistence
  prompts/            Jinja prompt templates
  models/             Pydantic models
  pipeline.py         Ingestion orchestration
evaluation/            Evaluation scripts and sample data
scripts/               Database setup SQL
Dockerfile             Container setup for local use
Dockerfile.prod        Alternate multi-stage container build
docker-compose.yml     Local container orchestration

• CHAT_PROVIDER selects the chat provider; the workflow defaults to Gemini.
• GEMINI_CHAT_MODEL and OPENAI_CHAT_MODEL override provider model names.
• MAX_CONCURRENT_INGESTIONS limits concurrent document jobs.
• DOCUMENT_PROCESS_MAX_ATTEMPTS controls retry attempts for transient model errors.
• Keep .env out of version control. Client-side applications should use only the Supabase public/anon key, never a service-role key.

MIT. See LICENSE.