How kltsai/legal-rag-chatbot works
Project type and purpose
• This is a Python-based Retrieval-Augmented Generation (RAG) chatbot specialized for legal queries.
• It ingests raw legal conversation data, computes embeddings, stores them in a vector database, and at runtime retrieves relevant context to feed into an LLM (Gemini) to generate answers.
File structure
• Top-level entry points:
• main.py – orchestrates batch pipelines or interactive sessions
• app.py / legal-rag-chatbot.py – likely CLI or API wrappers
• Core package: src/legal_rag_chatbot/
• components/
– data_ingestion.py
– data_embedding.py
– vector_storage.py
– gemini_client.py
– rag_sys.py
• config/configuration.py
• constants/
• entity/
• logging/
• utils/common.py
• Artifacts folder holds sample datasets and precomputed embeddings.
README.md
• Project title only, but name implies “Legal-RAG-Chatbot.”
Main components & interactions
a. Entry point (main.py)
• Calls Configuration loader
• Instantiates service modules
• Triggers ingestion/embedding pipeline or starts interactive chat
b. DataIngestion
• Reads raw JSON (train.json, test.json, val.json)
• Produces Document entities
c. DataEmbedding
• Takes Document entities → computes vector embeddings
d. VectorStorage
• Upserts embeddings into vector DB (local or cloud)
• Executes similarity searches at query time
e. GeminiClient
• Wraps external LLM API calls (prompt + context → completion)
f. RagSys
• At runtime:
1. Accepts user query
2. Calls VectorStorage.query() to retrieve top-k docs
3. Builds prompt (using utils/common, constants, templates)
4. Calls GeminiClient.generate()
5. Returns answer
Architectural patterns & design principles
• Modular “service” pattern: each concern isolated in its own component
• Pipeline for offline data prep (Ingestion → Embedding → Storage)
• RAG design at runtime: Retrieval (VectorStorage) + Generation (GeminiClient)
• Dependency inversion: RagSys depends on abstract interfaces of storage & LLM