⚖️ Legal Aid Eligibility & Document Preparation Agent

An offline AI agent that tells a citizen whether they qualify for free legal aid in India, cites the governing law, and builds their document checklist — running entirely on a local model.

Track: AAI · Domain: Access to justice & support navigation · Reference: NALSA — Legal Services Authorities Act, 1987 (Section 12) Runs fully offline on a small local model (Ollama) — no API key, no cost, and no applicant data ever leaves the device.

---

⚡ Instant launch — one click, zero setup

🖱️   Just double-click start.command

It does everything for you — starts the local AI engine, checks the models, serves the app, and opens it in your browser:

   ╔══════════════════════════════════════════════════════════╗
   ║      ⚖️   LEGAL AID ELIGIBILITY  &  DOCUMENT AGENT        ║
   ║          NALSA · runs 100% offline · no API key          ║
   ╚══════════════════════════════════════════════════════════╝
   [1/4] Starting the local AI engine (Ollama)…      ✓
   [2/4] Checking local models…                      ✓
   [3/4] Serving the app at http://localhost:8000 …  ✓
   [4/4] Opening the app in your browser…            ✓
   ✓  Ready!   Open →  http://localhost:8000/

No terminal. No config. No API keys. Prefer manual? See Run it below.

_{First run: if macOS says the file is blocked, right-click → Open once (or run chmod +x start.command).}

---

✨ What it does

	Feature	How
✅	Eligibility screening	A short questionnaire → a deterministic rules engine returns an indicative verdict with the governing NALSA citation (e.g. S.12(c) for a woman, S.12(h) for income).
📄	Document checklist	Auto-builds a checklist tailored to the applicant's category and matter type.
🔎	Legal Q&A (vector RAG)	Ask a question → local embeddings retrieve the most relevant NALSA passages → the model answers grounded in and citing them, with clickable sources.
🧾	Document Helper	Upload a legal notice (PDF/DOCX) → plain-language, 6th-grade explanation, on-device.
🛡️	Legal-advice guardrail	Requests for case strategy / outcome predictions are refused with a human hand-off.
🧪	Validation & logs	7 built-in test scenarios (7/7 passing) + a JSON-exportable screening log.
🔒	Privacy by design	100% offline: vendored UI + local model. Sensitive intake data stays on the user's machine.

---

🗺️ How it works

flowchart TD
    A([👤 User: questionnaire / question / document]) --> B[Agent controller]
    B --> C{Missing info?}
    C -- yes --> D[Ask for income / category]
    C -- no --> E
    D --> E[🛠️ Tools & retrieval]
    E --> E1[Rules engine: assessEligibility]
    E --> E2[Checklist builder]
    E --> E3[Vector RAG over NALSA corpus]
    E1 --> F[🛡️ Legal-advice guardrail]
    E2 --> F
    E3 --> F
    F --> G[[📋 Verdict + citation + checklist + grounded answer]]
    G --> H[(🗒️ Log screening · export JSON)]
    style G fill:#f5f1e8,stroke:#8b0000,stroke-width:2px
    style F fill:#fbeaea,stroke:#8b0000

Loading

System architecture

flowchart LR
    UI[🖥️ React SPA] --> CTRL[Agent Controller]
    CTRL --> RULES[⚙️ Deterministic rules engine<br/>NALSA S.12 · income ceilings]
    CTRL --> RAG[🔎 Vector RAG]
    CTRL --> DOC[🧾 Document Helper]
    RAG --> EMB[(nomic-embed-text<br/>768-d index)]
    RAG --> LLM[🧠 Local chat model<br/>Ollama /api/chat]
    DOC --> LLM
    RULES --> OUT[✅ Verdict + citation]
    LLM --> OUT2[💬 Plain-language / cited answer]
    OUT --> GUARD[🛡️ Guardrail]
    OUT2 --> GUARD
    GUARD --> FINAL[[Final answer + logs]]
    style RULES fill:#e8f0e8,stroke:#2d6a4f
    style LLM fill:#f3ecf7,stroke:#6a4c93

Loading

Offline vector RAG pipeline

flowchart LR
    C[📚 data/corpus/*.md<br/>NALSA reference notes] --> K[build_rag.py<br/>chunk + embed]
    K -->|nomic-embed-text| I[(data/rag_index.json<br/>16 × 768-d vectors)]
    Q[❓ User question] --> QE[Embed query]
    I --> R[Cosine top-k<br/>in browser]
    QE --> R
    R --> P[Grounded prompt<br/>with cited passages]
    P --> M[🧠 Local model]
    M --> A([💬 Answer + clickable sources])

Loading

The whole pipeline is on-device: the index is built once offline by build_rag.py; retrieval and generation at query time use only local embeddings + a local model, and every answer cites the passages it used.

---

🧠 Model choice — why not the tiny Qwen?

The agent runs on any local Ollama model — pick it live in the sidebar dropdown. We benchmarked several on the plain-language + grounded-RAG tasks:

Model	Size	Verdict
qwen3.5:0.8b	~1 GB	A reasoning ("thinking") model. At 0.8 B it spends its token budget on a hidden chain-of-thought and often returns empty or garbled answers for our structured tasks — and its thinking mode can't be reliably disabled over the API. Kept as the "it literally runs on a 1 GB model" proof, not for real answers.
llama3.1	8 B	Clean and reliable — a solid middle option.
gemma4:latest ⭐	~9.6 GB	Best results for this use case, and the default. A modern encoder-free "Unified" decoder-only transformer, and — empirically — it emits the grounded, cited, plain-language answer straight away with no thinking-mode detour, which is exactly what a legal-aid explainer needs.

Architecture note — what "Unified" (encoder-free) means. Most multimodal models bolt dedicated encoders in front of the LLM to pre-process images/audio. Gemma 4 12B "Unified" removes those encoders entirely, projecting raw image patches and audio waveforms directly into the LLM's embedding space through lightweight linear layers. Everything flows into a single decoder-only transformer, which lowers multimodal latency and lets the whole model be fine-tuned in one pass. (It is still a transformer — just encoder-free, not "transformer-less".) For this text-only task the practical win is its clean, direct instruction-following.

Key design point: the eligibility verdict is decided by the deterministic rules engine, not the model — so switching models never changes correctness, only the wording of the explanation. That's what makes it safe to run on any local model, tiny or large.

---

🚀 Run it

Prerequisites: Ollama and Python 3.

# 1. Pull models (chat + embeddings)
ollama pull gemma4            # default chat model (best results)
ollama pull llama3.1          # good alternative
ollama pull nomic-embed-text  # embeddings for the Legal Q&A RAG

# 2. Build the RAG index (one-time / after editing data/corpus)
python3 build_rag.py          # writes data/rag_index.json

# 3. Serve + open  ── easiest: double-click start.command (macOS)
python3 serve.py              # then open http://localhost:8000/

⚠️ Open it at http://localhost:8000/, not by double-clicking the file. Ollama's CORS policy rejects the file:// origin, and browser localStorage needs an http origin. If you just pulled a change and the UI looks stale, hard-reload (⌘⇧R).

To edit the app: change src_app.jsx → node build.js → reload.

---

📸 Screenshots

Eligibility + checklist	Mutually-exclusive categories
Legal Q&A — grounded, cited RAG	Validation — 7/7 scenarios

---

🧩 AI / agent component (where AI is used, and why)

• Deterministic core (software logic): eligibility + checklist are pure functions — repeatable, auditable, unit-tested. The right engineering choice for a sensitive legal domain.
• AI layer (local LLM): rephrases the rules-engine verdict into plain, supportive language; powers the Document Helper; and answers Legal Q&A grounded in retrieved NALSA passages (true vector RAG). The model never decides eligibility.
• Guardrail: advice/outcome requests are detected and refused, with a hand-off to a human lawyer/PLV.

---

📊 Results & insights

• Rules engine passes 7 / 7 built-in scenarios (also in data/sample_intake.csv).
• Every verdict cites its governing provision; every RAG answer cites its sources (clickable).
• Runs on a 1 GB model as proof; gemma4:latest gives the cleanest explanations.
• Because eligibility is deterministic, model choice never changes the verdict — only the wording.

⚠️ Limitations & responsible use

Indicative screening only — not legal advice; income ceilings vary by state and change over time; the Legal Services Authority makes the final decision; small local models are slower on CPU. Full notes: docs/limitations_and_responsible_use.md. In an emergency, call the NALSA helpline 15100.

🔮 Future improvements

Runtime-editable rules/ceilings · expand the RAG corpus with full NALSA circulars + per-state rules and add re-ranking · multilingual (Hindi & regional) output · OCR for scanned documents · one-click PLV hand-off with the screening log attached.

📁 Repository structure

├── index.html            # the app shell (loads vendored libs + compiled app)
├── src_app.jsx           # editable source  →  build.js  →  vendor/app.js
├── build.js              # JSX → vendor/app.js (classic React runtime, no bundler)
├── build_rag.py          # offline RAG indexer (local embeddings)
├── serve.py              # localhost static server (stdlib, no-cache)
├── start.command         # one-click launcher (Ollama + server + open)
├── data/                 # NALSA reference CSVs + corpus/ + rag_index.json
├── docs/                 # project report, architecture, responsible-use
├── screenshots/          # working-prototype captures
├── vendor/               # React, mammoth, pdf.js, compiled app (all local)
└── requirements.txt

👥 Team

• Chandandeep Sharma
• Amey Chorge
• Neal Balsara

---

_{Educational AAI capstone · not affiliated with NALSA · provides indicative information only, not legal advice.}