Legal_clause-RAG

Legal Contract Analyzer — Backend

A RAG-powered FastAPI service that takes Indian employment contracts (PDF), splits them into clauses, and checks each one against the Constitution of India and the Indian Contract Act, 1872. It returns risk levels, confidence scores, legal reasoning, and lets you chat with the analyzed document afterwards.

Stack: FastAPI · Gemini 2.5 Flash · FAISS · Redis · Docker

---

Problems I Ran Into (and How I Fixed Them)

---

1. Splitting Contracts Into Clauses Is Harder Than It Sounds

The problem: I assumed I could just split on sentences or newlines. Nope. Real employment contracts have nested numbering like 1., (a), (ii), sub-clauses inside sub-clauses, and zero consistency between different PDFs. Sentence splitting was giving me broken clauses — half a legal thought in one chunk, the other half in the next.

What I did: Wrote a multi-pattern regex that tries numbered/lettered clause markers first (1., (a), a), etc.). If that fails (like when the contract is just a wall of text), it falls back to splitting on sentence-ending punctuation (. and ;). It's not perfect, but it handles the majority of real-world contracts I tested with.

---

2. Building the Legal Knowledge Base Was a Manual Grind

The problem: Constitutional articles, Contract Act sections, and Supreme Court judgments all come in completely different formats — some from PDFs, some from legal websites, some from bare text files. There was no clean dataset I could just plug in.

What I did: Manually extracted the relevant provisions and converted everything into a common JSON format. Each entry has text, source, reference (like "Section 27" or "Article 19"), and type (statute vs case law). Not glamorous, but it gave me full control over what goes into the vector index.

Another problem with case law: Supreme Court judgments can be hundreds of pages long. If I indexed them as-is, retrieval would pull back huge irrelevant paragraphs instead of the actual legal principle I needed.

Fix: I hand-picked landmark cases relevant to employment rights (like Vishaka v. State of Rajasthan, Olga Tellis v. BMC, Niranjan Shankar Golikari), extracted just the key holdings and constitutional principles, and stored them as short, citation-ready chunks. Much better retrieval quality.

---

3. Contract Clauses Don't Talk Like the Constitution

The problem: A non-compete clause in a contract will never say "this violates Article 19(1)(g) — freedom of profession." It'll say something like "the employee shall not engage in any competing business for 2 years after termination." Keyword-based search was basically useless here.

What I did: Used semantic embeddings (Gemini's gemini-embedding-001) so the system retrieves based on meaning, not exact words. A clause about restricting future employment now correctly pulls up Article 19(1)(g) and Section 27 of the Contract Act, even though they share zero keywords.

---

4. Retrieval: Too Little Context vs. Too Much

The problem: When I set top_k=1, the system would miss important legal references. When I cranked it up to top_k=10, the LLM got flooded with irrelevant context, slowed down, and started giving wishy-washy answers.

What I did: Settled on top_k=3 after testing. Three relevant documents give enough legal grounding without turning the prompt into a wall of text. The FAISS L2 distance also helps — if the top result is already far away, the system knows retrieval confidence is low.

---

5. LLM Confidence Alone Is Not Trustworthy

The problem: I initially asked Gemini to self-report a confidence score. It would say "95% confident" on answers where the retrieved context was barely relevant. Pure LLM confidence is just vibes.

What I did: Built a hybrid confidence formula that blends two signals:

• Retrieval match (40% weight) — based on how close the FAISS L2 distance was. Closer = more relevant context was found.
• LLM certainty (60% weight) — the model's self-reported confidence, but now anchored by real retrieval quality.

This way, if the model is confident but the retrieval was weak, the final score drops. If both are strong, the score is high. Much more calibrated than trusting either signal alone.

---

6. The 10-Second Wait Was Killing the UX

The problem: When a contract has 15-20 clauses and each one needs an embedding call + LLM generation, processing them one-by-one took 10-15 seconds. The user just stared at a loading spinner with no feedback.

What I did two things:

Concurrency: Wrapped all clause analyses in asyncio.create_task() and used asyncio.Semaphore(5) to process up to 5 clauses in parallel. The semaphore prevents hitting Gemini's rate limits while still being way faster than sequential.

SSE Streaming: Switched from returning one big JSON blob to streaming results via Server-Sent Events (text/event-stream). Now clauses pop up on the frontend one-by-one as they finish. First result shows up in under a second instead of waiting for all 15 to complete. Used asyncio.as_completed() so faster clauses arrive first regardless of their position in the document.

---

7. Getting the LLM to Return Consistent JSON

The problem: Gemini would sometimes return markdown, sometimes JSON with extra fields, sometimes JSON with missing fields. The frontend needs a predictable schema every time.

What I did: Used Gemini's response_mime_type="application/json" with a response_schema set to my Pydantic model (ClauseAnalysis). This forces the model to output JSON that matches my exact schema — clause_summary, potential_violation, applicable_laws, legal_reasoning, risk_level, confidence_score. No more parsing surprises.

---

8. Deploying on Free Infrastructure Was Its Own Battle

The problem: Hugging Face Spaces gives you a free Docker container, but RAM is limited, startup times matter, and some Python packages are huge. I also hit issues with:

• Invalid Host header errors (the HF proxy uses a different hostname than localhost)
• Port mismatches between the Dockerfile and HF's expected app_port

What I did:

• Added *.hf.space to TrustedHostMiddleware allowed hosts
• Matched the Dockerfile EXPOSE and CMD port to 7860 (what HF expects)
• Used python:3.11-slim as the base image to keep the container small
• Set --workers 1 since the free tier doesn't have enough RAM for multiple Uvicorn workers

---

9. Balancing Cost, Speed, and Accuracy

The problem: Gemini embeddings and reasoning give great quality but every API call costs time and adds external dependency. For a portfolio project, I needed it to be fast enough to demo live without awkward pauses.

What I explored:

• Looked into replacing Gemini embeddings with local models like BAAI/bge-small-en-v1.5 (runs on CPU, no API call needed, cuts embedding latency to near-zero)
• Considered Groq-hosted open-source LLMs for faster inference
• Ended up sticking with Gemini for quality, but the architecture is designed so swapping the embedding model or LLM is a one-file change in rag_core.py

---

Running It Locally

# Clone and set up
python -m venv venv
source venv/bin/activate  # Windows: .\venv\Scripts\activate
pip install -r requirements.txt

# Create .env with your Gemini API key
echo "GEMINI_API_KEY=your_key_here" > .env

# Start
uvicorn main:app --host 0.0.0.0 --port 8000 --reload

Swagger docs at http://localhost:8000/docs.

Docker

docker build -t legal-rag-backend .
docker run -p 7860:7860 --env-file .env legal-rag-backend

---

API Endpoints