Docuery AI

Docuery AI is a full-stack, multi-document RAG assistant that lets users upload PDFs, ask natural-language questions, receive citation-backed answers, and keep private chat/document history per signed-in user.

Live demo: muthapriyanka27-pdf-qa-assistant.hf.space

Highlights

• Multi-PDF upload with automatic parsing, chunking, embedding, and indexing.
• Retrieval-augmented answers with filename and page citations.
• General-question fallback when a question is not supported by uploaded PDFs.
• Hugging Face and Google OAuth login for private, user-scoped document workspaces.
• Persistent SQLite chat sessions, messages, document collections, and auth sessions.
• ChromaDB vector search with per-user metadata filtering.
• Clear chat, rename chat, delete chat, and delete document flows.
• Dark/light responsive UI with a public Hugging Face Space deployment.

Tech Stack

Layer	Tools
Frontend	React, HTML, CSS
Backend	FastAPI, Pydantic, Uvicorn
RAG	LangChain, ChromaDB, Hugging Face embeddings
PDF Processing	PyPDF, LangChain text splitters
LLM Inference	Ollama
Persistence	SQLite, Chroma persistent storage
Auth	Hugging Face OAuth, Google OAuth
Deployment	Docker, Hugging Face Spaces

Architecture

User
  -> React UI
  -> FastAPI routes
  -> PDF parser + text splitter
  -> Hugging Face embeddings
  -> ChromaDB vector index
  -> Ollama LLM
  -> citation-backed answer

SQLite stores users, OAuth sessions, chat sessions, messages, and document
metadata. ChromaDB stores embedded PDF chunks with user/document metadata.

Key Backend Features

• POST /api/upload uploads one or more PDFs and indexes their chunks.
• POST /api/ask routes questions through document retrieval or general LLM mode.
• GET /api/state returns only the signed-in user's documents and chats.
• POST /api/sessions creates a chat session.
• PATCH /api/sessions/{session_id} renames a chat or changes its document context.
• DELETE /api/sessions/{session_id}/messages clears a chat.
• DELETE /api/sessions/{session_id} deletes a chat.
• DELETE /api/collections/{collection_id} deletes a document collection and its vector chunks.
• GET /api/auth/login starts Hugging Face OAuth login.
• GET /api/auth/google/login starts Google OAuth login.
• POST /api/auth/logout clears the auth session.

Privacy Model

The deployed app supports Hugging Face OAuth and Google OAuth. Each collection, document, chat session, and retrieval query is scoped by user_id, so one signed-in user cannot see another user's uploaded PDFs or chat history.

For local development, auth is relaxed by default and uses a local development user. Set REQUIRE_AUTH=true to force the same auth behavior locally.

Local Development

1. Backend

cd backend
python3 -m venv venv
./venv/bin/pip install -r requirements.txt
./venv/bin/uvicorn app.main:app --reload

Backend URL:

http://127.0.0.1:8000

2. Frontend

The frontend is a lightweight React app served as static files.

cd frontend-simple
python3 -m http.server 5173

Frontend URL:

http://127.0.0.1:5173

3. Ollama

Run Ollama locally and pull a model:

ollama pull llama3
ollama serve

Optional backend environment variables:

OLLAMA_URL=http://127.0.0.1:11434/api/generate
OLLAMA_MODEL=llama3
REQUIRE_AUTH=false
GOOGLE_CLIENT_ID=your-google-client-id
GOOGLE_CLIENT_SECRET=your-google-client-secret

For Google login, add this authorized redirect URI in Google Cloud Console:

http://127.0.0.1:8000/api/auth/google/callback

For the deployed Hugging Face Space, use:

https://muthapriyanka27-pdf-qa-assistant.hf.space/api/auth/google/callback

Deployment

The project is Docker-ready for Hugging Face Spaces.

cd backend
./venv/bin/hf auth login

cd ..
backend/venv/bin/hf upload muthapriyanka27/pdf-qa-assistant . . \
  --repo-type space \
  --include "README.md" \
  --include "Dockerfile" \
  --include ".dockerignore" \
  --include "DEPLOYMENT.md" \
  --include "docker/**" \
  --include "backend/app/**" \
  --include "backend/requirements.txt" \
  --include "frontend-simple/**" \
  --exclude "backend/venv/**" \
  --exclude "backend/chroma_db/**" \
  --exclude "backend/*.sqlite3" \
  --exclude "frontend-simple/node_modules/**" \
  --exclude "**/__pycache__/**" \
  --exclude "**/*.pyc"

The Hugging Face Space metadata at the top of this README enables Docker and OAuth:

sdk: docker
app_port: 7860
hf_oauth: true

Project Structure

backend/
  app/
    routes/        FastAPI API routes
    services/      PDF parsing, chunking, embeddings, vector search, LLM calls
    storage/       SQLite persistence
frontend-simple/   React UI and CSS
docker/            Space startup script
Dockerfile         Hugging Face Spaces container
DEPLOYMENT.md      Deployment notes

Resume Summary

Built a full-stack RAG document assistant with FastAPI, ChromaDB, LangChain, Hugging Face embeddings, Ollama inference, multi-provider OAuth, SQLite-backed chat persistence, citation-backed answers, multi-PDF upload, and a responsive React UI deployed on Hugging Face Spaces.