Indian Legal AI Expert

A production-grade RAG system that answers Indian legal queries with source citations, confidence scoring, and full LLMOps observability — running on ₹0/month infrastructure.

---

Architecture Overview

User Query
  ↓
React 18 SPA (Vercel)
  ↓ HTTPS
FastAPI + Uvicorn (Render free tier — 512MB RAM)
  ↓
LangGraph StateGraph (6 nodes)
  ├── classify_node    → query type routing
  ├── reject_node      → abusive query fast-fail
  ├── greet_node       → greeting handling (no Qdrant)
  ├── retrieve_node    → PII mask → keyword expand → dual Qdrant search
  ├── generate_node    → confidence gate → LLM via OpenRouter
  └── post_process_node → MongoDB save → Langfuse log
  ↓
Response + Sources + Confidence Score

---

Tech Stack (₹0/month)

Layer	Technology	Why
Frontend	React 18 + Vite + Tailwind	Vercel free tier
Backend	FastAPI + Uvicorn	Fast async, minimal overhead
Orchestration	LangGraph StateGraph	Conditional routing — not linear chain
Vector DB	Qdrant Cloud (1GB free)	Parent text in payload — no SQL round-trip
Embeddings	Jina AI v2 (1M/month free)	0MB local RAM
LLM	Qwen 3 235B via OpenRouter	Best free-tier quality
Chat History	MongoDB Atlas (512MB free)	Persistent cross-session memory
Cache	Upstash Redis (10K req/day)	Repeated query caching
Auth	Google OAuth 2.0 + JWT	No credential storage
PII Masking	Microsoft Presidio + Custom Regex	Indian PII: Aadhaar + PAN
Observability	Langfuse	Full LLM trace: latency, tokens, confidence

---

Knowledge Base

Document	Chunks (Parent)	Chunks (Child)	Size
Constitution of India 2024	685	2,845	2.4 MB
Bharatiya Nagarik Suraksha Sanhita	555	2,656	2.1 MB
Bharatiya Nyaya Sanhita 2023	271	1,311	0.9 MB
Motor Vehicles Act 1988	262	1,248	1.2 MB
Consumer Protection Act 2019	81	416	1.2 MB
IT Act 2000 (Updated)	83	420	0.8 MB
Total	1,937	8,896	~8.6 MB

---

Performance Metrics

Average response latency:   189ms  (measured across 50 queries)
Minimum observed:           167ms
Maximum observed:           211ms
Infrastructure:             Render free tier, 512MB RAM
Vector search:              768-dim Cosine, 100% precision@3

Confidence zones:
  0–39%   → Fallback response (no LLM call)
  40–65%  → Partial match answer
  65–85%  → Good match with sources
  85–100% → Exact match with direct citation

---

Key Engineering Decisions

1. Parent Text in Qdrant Payload

Instead of storing parent text in PostgreSQL and doing a SQL lookup after vector search, parent text lives directly in the Qdrant point payload. One API call returns both the child match and the parent context. Eliminated one full network round-trip per query.

2. Legal Keyword Expansion

Acronym expansion before embedding:

KEYWORD_MAP = {
    "fir":     "First Information Report Section 173 BNSS",
    "murder":  "Section 103 BNS culpable homicide",
    "arrest":  "Section 35 BNSS arrest without warrant",
    "bail":    "Section 479 BNSS bail conditions",
    "ipc":     "Indian Penal Code BNS equivalent",
}

Users type "FIR" — system searches for the full legal phrase. Precision improvement: significant.

3. Dual Search (Core + User Uploads)

Every query searches two Qdrant collections simultaneously:

• Core corpus (6 legal documents — permanent)
• User temp vectors (documents uploaded by this specific user — session-scoped)

Results merged, deduplicated by parent_id, sorted by score.

4. Zero-Memory PII Masking

Replaced spaCy NLP (~200MB RAM) with custom regex recognizers (0.01MB). Same accuracy for Indian PII patterns (12-digit Aadhaar, PAN format ABCDE1234F).