Legal AI Assistant – Project Overview & Presentation Script

---

1. Mission

Deliver a professional, Indian-law–focused AI assistant that answers legal questions using the new Bharatiya Nyaya Sanhita (BNS), related legislation, and Supreme Court precedents.
Built with a Retrieval-Augmented-Generation (RAG) backend and a modern chat-style frontend.

---

2. High-Level Architecture

User ↔ Next.js Frontend ↔ FastAPI Backend ↔ LangGraph-powered RAG System ↔
         ├─ Supreme-Court Vector DB (Chroma)          
         └─ Tavily Web Search API

1. Frontend (Next.js 15 + Tailwind) – chat UI, markdown rendering, avatars, loading states, clear-chat, mobile responsive.
2. Backend (FastAPI) – single /ask endpoint; handles CORS and calls RAG pipeline.
3. RAG Pipeline (LangGraph)
   • Embeddings: Sentence-Transformers
   • Vector DB: Chroma (persisted in backend/db_data/sc-chroma)
   • LLM: Llama-3-70B via Groq API
   • Web search: Tavily API
   • Graph nodes: routing → retrieval → grading → generation → grading.

---

3. Backend Walkthrough

Stage	Key Code	Purpose
Initialization	initialize_rag_components()	Loads models, vectorstore, tools, prompts, and compiles the LangGraph workflow.
Routing	route_question_edge	LLM decides whether to use SC vectorstore, web search, or both.
Retrieval Nodes	retrieve_sc_node, retrieve_web_search_node, retrieve_combined_sources_node	Pull relevant documents (k = 5) or search results (k = 3).
Document Grading	grade_documents_node	LLM grades each doc for relevance (truncates to 2 k chars & safe-try).
Query Rewrite	transform_query_node	Improves the question if no good docs found.
Answer Generation	generate_answer_node + RAG prompt	Produces structured legal answer (summary, provisions, analysis, final).
Answer Grading	grade_generation_edge	Checks hallucination + usefulness; may regenerate or rewrite.
Safety Guard	MAX_GENERATION_ATTEMPTS = 5	Stops infinite loops and returns fallback "Sorry, we could not understand your query…".

Endpoint: POST /ask

{
  "question": "What is the procedure for filing an FIR?",
  "chat_history": []   // (reserved for future multi-turn use)
}

Returns: answer (markdown) + sources (URL metadata).

---

4. Frontend Walkthrough

Component: components/ChatInterface.tsx

1. State: messages, input, isLoading, error.
2. UX perks: avatars, markdown via react-markdown + remark-gfm, thumbs-up/down placeholders, clear-chat, loading spinner.
3. Flow:

   await fetch("http://127.0.0.1:8000/ask", {...})
   → push assistant message with answer & sources

4. Responsive layout: flex column → mobile stack; Tailwind prose for typography.

---

5. End-to-End Flow

1. User types a legal query → clicks Send.
2. Frontend POSTs to /ask.
3. Backend's LangGraph pipeline:
   1. Router chooses sc_vectorstore / web_search.
   2. Retrieves docs → grades relevance.
   3. If no good docs, question rewritten → new retrieval.
   4. RAG prompt + docs → draft answer.
   5. Hallucination & usefulness graded.
   6. If bad, regenerate or rewrite (max 5 loops).
4. Final answer & sources returned.
5. Frontend renders markdown-styled answer, shows clickable sources.

---

6. Running Locally (Demo Setup)

# 1) Backend
cd backend
python -m venv venv && venv\Scripts\activate      # Windows example
pip install -r requirements.txt
set GROQ_API_KEY=...        && set TAVILY_API_KEY=...
uvicorn app.main:app --reload

# 2) Frontend (new shell)
cd nextjs_frontend
npm install
npm run dev

Navigate to http://localhost:3000

---

7. Live Presentation Script (≈ 4 min)

1. Opening (30 s)
   "Good day judges. I present the Legal AI Assistant, designed to democratize access to up-to-date Indian legal knowledge."
2. Problem (20 s)
   "With BNS replacing IPC, practitioners need quick, reliable answers—existing search tools are slow and fragmented."
3. Solution Overview (40 s)
   Show architecture slide / diagram above. Emphasise RAG to ground answers in law + cases.
4. Live Demo (90 s)
   • Ask: "Explain the procedure for filing an FIR under BNSS."
   • Highlight bold / italic markdown, structured answer sections, source links.
   • Clear chat, ask "What was the SC view on Triple Talaq?" ‑ show Supreme-Court citations.
5. Safety & Quality (30 s)
   Explain grading nodes, hallucination guard, retry cap.
6. Tech Highlights (20 s)
   Groq Llama-3-70B speed, Chroma vectorstore, LangGraph orchestration.
7. Closing (30 s)
   "Scalable to other jurisdictions, integrates new caselaw instantly via web search. Ready to assist lawyers, journalists, and citizens alike."

---

8. Future Work

• Multi-turn memory using chat_history.
• Feedback buttons hooked to a scoring API for continual improvement.
• Deploy on GPU-backed cloud with authentication & usage analytics.

---

Prepared by Aman Gusain aka Callmeamann