How lillianassefa/contract-ai-chatbot works
Project type and purpose
• This is a full-stack RAG (Retrieval-Augmented Generation) system for Contract Q&A.
• Front-end: React single-page application (SPA) for user interaction.
• Back-end: Python service implementing ingestion, retrieval, and generation stages.
• Data layer: a local vector store (SQLite/Chroma) plus contract storage.
• External service: LLM API (e.g., OpenAI) for answer generation.
File structure
Top level:
• backend/ – Python notebooks (prototypes) under notebooks/, core service under src/
• database/ – vectorstore.py, DB access wrappers
• frontend/ – React app (public/, src/, package.json)
• environment.yml, requirements.txt – Python dependencies
Key back-end modules (backend/src):
• upload.py – handles PDF/text ingestion
• retriever.py – queries vector store
• generator.py – wraps LLM call
• main.py – HTTP API entry point
• constants.py, init_conf.py – configuration
Data files: backend/src/uploads/contract.pdf/.txt, notebooks/db/chroma.sqlite3
Insights from README
• Entry point: python src/main.py
• Two phases of implementation: prototype (notebooks) vs. production code (src)
• React UI demonstrates chat interface components (ChatHistory, ChatInput, TopBar)
Components & interactions
a. Main components
1. Client (React SPA)
2. API Layer (Flask/FastAPI in main.py)
3. Ingestion Module (upload.py)
4. Vector Store (database/vectorstore.py + SQLite/Chroma)
5. Retrieval Module (retriever.py)
6. Generation Module (generator.py → LLM API)
7. Notebooks (research/proto, optional)
b. Relationships & data flow
1. User types question in ChatInput → React sends HTTP POST /ask → API Layer
2. API calls retriever → queries Vector Store → returns top-k passages
3. Passages + user query → generator → calls external LLM → returns answer
4. API returns JSON answer → React updates ChatHistory
5. Ingestion: Admin uploads contract via HTTP POST /upload → upload.py extracts text → vectorizes embeddings → stores in Vector Store
c. Architectural patterns
• Separation of concerns: Ingestion, Retrieval, Generation as distinct modules
• RAG pipeline: “retrieve then generate”
• SPA ↔ API pairing
• Data store abstraction (vectorstore.py)
d. Technologies/frameworks
• Front-end: React, JavaScript, CSS (create-react-app)
• Back-end: Python (Flask or FastAPI), Jupyter notebooks for prototyping
• Data store: Chroma (SQLite) for vector embeddings
• LLM integration: OpenAI (or similar) via generator.py
• Env management: requirements.txt, environment.yml