Vidhi

Indian legal research workbench — for working advocates. Matters, briefs, and authorities in one place. Ask a question, get a structured legal brief — Issue · Governing Law · Authorities · Discussion · Conclusion · Recommended Actions — backed by live citations from Indian Kanoon, India Code, and curated web sources.

Formerly MiniHarvey. Domain and deploy URLs are mid-migration — links below still resolve.

🔗 Try the live demo →

⚠️ Legal information, not legal advice. Vidhi is under active development. For your matter, consult an advocate registered with the Bar Council of India.

---

What it does

Vidhi turns a plain-English legal question ("What are my rights if arrested without a warrant?") into a memo-style answer with:

• Query rewriting — your natural-language question is rewritten into a keyword-dense search query before hitting the providers.
• Parallel search across Indian Kanoon (case law), India Code (statutes/acts), and Google CSE (curated Indian legal web).
• Source balancing — results are round-robin interleaved by provider, not dominated by whoever returns fastest.
• Streaming LLM answer — a structured memo in the voice of "Harvey", with every section clearly labelled. Responses stream token-by-token.
• Citation chips — AIR, SCC, Section, and Article citations extracted from the answer; click one to flash the matching source card.
• Follow-up questions — three clickable chips at the bottom of each brief let you keep researching with one click.
• Pinnable Workbook — pin any case or statute to keep it across threads.
• Explainer videos — relevant YouTube explainers surfaced alongside sources.
• Light & dark themes — navy/gold chambers mode, or warm cream paper mode.

Inspired by MiniPerplexity — same shape, but repurposed as a domain-specific legal assistant.

---

Screenshots

Sign-in — cinematic landmark-case slideshow

A split-screen login: on the left, a rotating slideshow of landmark Indian cases (Kesavananda Bharati, Maneka Gandhi, Vishaka, Puttaswamy, Navtej Singh Johar) layered over Wikimedia imagery of the Supreme Court; on the right, a Clerk sign-in card styled in chambers navy + gold, with a "Continue as guest" escape hatch and the सत्यमेव जयते motto.

Workbench — three-pane research layout

The three-pane research surface in action. A corporate-merger query streams back a structured brief — Issue → Governing Law → Authorities → Discussion — with the Threads rail on the left, Sources + Explainer Videos on the right, and the composer docked beneath the brief for follow-ups.

• Left sidebar — Threads (history) + Workbook (pinned sources). Each section individually collapsible; the whole rail collapses to a thin icon strip.
• Center — the Brief: a typography-first legal memo with a gold Conclusion callout, clickable citation chips, and a follow-up chip strip.
• Right sidebar — Sources (color-tagged by provider — Kanoon / India Code / Google / Wikipedia) + YouTube Explainer Videos. Also individually collapsible.

---

Architecture

flowchart LR
    subgraph Client["Browser"]
        UI["React 18 · Vite · Tailwind 4<br/>Clerk auth · three-pane workbench"]
    end

    subgraph Backend["FastAPI · Render"]
        direction TB
        RL["⚡ Rate limiter<br/>Token bucket · 30 req/min<br/>keyed by endpoint×user"]
        H["Query handler<br/>/search · /answer (SSE)<br/>/session · /health"]
        QC["Classifier<br/>case_law · statute · general<br/>(keyword heuristic, no LLM)"]
        PM["📝 Prompt manager<br/>• Harvey memo prompt (6-section)<br/>• Rewriter prompt (keyword extractor)<br/>• Output guardrails"]
        FS["🧹 Filter & extractor<br/>• BS4 strip script/nav/footer/aside<br/>• AIR / SCC / § / Art citation regex<br/>• URL dedup + round-robin merge"]
    end

    subgraph External["External services"]
        LLM["Cloudflare AI Workers<br/>LLaMA 3.1 70B"]
        K["Indian Kanoon API"]
        IC["India Code scrape"]
        G["Google CSE"]
        YT["YouTube Data v3<br/>safeSearch: strict"]
    end

    UI -- "HTTPS · SSE" --> RL
    RL -- "429 on burst" --> UI
    RL --> H
    H --> QC --> PM
    PM <-. "rewrite · 8 s timeout" .-> LLM
    PM <-. "memo · stream · 2048 tok" .-> LLM
    H --> K
    H --> IC
    H --> G
    H --> YT
    K --> FS
    IC --> FS
    G --> FS
    FS --> UI
    LLM -- "token stream" --> UI
    YT -- "explainer cards" --> UI

    classDef ours fill:#0f2d4a,stroke:#d4af37,color:#fff
    classDef ext fill:#f5efe0,stroke:#0f2d4a,color:#0f2d4a
    class UI,RL,H,QC,PM,FS ours
    class LLM,K,IC,G,YT ext

Loading

Request pipeline

flowchart TD
    Q["User query<br/>(natural language)"] --> RL{"⚡ Rate limit<br/>token bucket · 30/min<br/>per endpoint×user"}
    RL -- "over limit" --> E429["HTTP 429<br/>+ retry hint"]
    RL -- "pass" --> QC["🏷 Classifier<br/>case_law / statute / general<br/>keyword scoring · no LLM call"]
    QC --> RW["📝 Rewriter prompt<br/>_REWRITE_SYSTEM · Cloudflare LLM<br/>8 s hard timeout"]
    RW --> RWG{"Rewrite guardrails<br/>len 3–200 · strip quotes<br/>reject refusal prefixes"}
    RWG -- "bad / slow / empty" --> RAW["Fall back to raw query<br/>(search never blocks on rewrite)"]
    RWG -- "good" --> PAR["⚡ Parallel fan-out<br/>ThreadPoolExecutor · 3 workers"]
    RAW --> PAR
    PAR --> IK["Indian Kanoon<br/>(authority)"]
    PAR --> ICS["India Code<br/>(statutes)"]
    PAR --> GC["Google CSE<br/>(fill gaps)"]
    PAR --> YTS["YouTube Data v3<br/>'Indian law explained'<br/>safeSearch: strict"]
    IK --> CX["🧹 Content extractor<br/>BS4 · strip script / style / nav<br/>footer / header / noscript / aside<br/>cap 2000 chars"]
    ICS --> CX
    GC --> CX
    CX --> RR["Round-robin merge<br/>Kanoon → IndiaCode → Google<br/>(authority order)"]
    RR --> DD["URL dedup<br/>first occurrence wins"]
    DD --> AP["📝 Harvey memo prompt<br/>_SYSTEM_PROMPT + search context<br/>+ last 3 session queries"]
    AP --> LLM["Cloudflare LLaMA 3.1 70B<br/>stream · max_tokens 2048 · 90 s timeout"]
    LLM -- "SSE chunks" --> UI["Brief renders live"]
    LLM --> CF["Citation extractor<br/>AIR / SCC / § / Art regex<br/>(everything else discarded)"]
    CF --> FIN["Final SSE frame:<br/>citations · suggested_steps<br/>· follow_ups"]
    YTS --> VID["Video cards sidebar"]

Loading

Why not SCI?

The Supreme Court of India's own portal (main.sci.gov.in) was pulled from the provider list — the judgment pages are JavaScript-rendered behind session cookies, the SSL certificate has a habit of expiring, and Indian Kanoon already indexes the full SC corpus with better metadata. If you need SC coverage, Indian Kanoon is the source of truth.

---

Engineering highlights

Concern	How it's handled
LLM latency	Streaming (SSE) so users see the first section in <2 s instead of waiting for the full memo.
Provider dominance	Round-robin merge — no single provider can drown the others by returning first.
Provider failure	Per-provider try/except with empty-list fallback; a dead Kanoon call still leaves India Code + Google results.
Rate abuse	Per-session sliding-window limiter (default 30 req/min), configurable via env.
Server-side state	In-memory session store keyed by session_id with a configurable TTL (default 600 s). No DB dependency.
Hallucinated citations	System prompt pins AIR / SCC / Section / Article formats; citation regex extracts only well-formed references for the chip row.
Cold-start drift	Query rewriter has an 8 s hard timeout; on failure the raw user query is used so the answer endpoint never blocks on rewrite.
CORS & secrets	All tokens server-side; ALLOWED_ORIGINS locks CORS to the Netlify domain in prod.

---

Tech stack

Layer	Choice
Frontend	React 18, TypeScript, Vite 8, Tailwind CSS 4
Auth	Clerk v5
Markdown	react-markdown
Backend	FastAPI, Pydantic, Uvicorn, Gunicorn
HTML extraction	BeautifulSoup 4
Parallelism	concurrent.futures.ThreadPoolExecutor (3 workers)
LLM	Cloudflare AI Workers — @cf/meta/llama-3.1-70b-instruct
Search	Indian Kanoon API · India Code scrape · Google CSE · YouTube Data API
Deploy	Netlify (frontend) · Render (backend)

---

Project layout

vidhi/
├── backend/
│   ├── main.py                         # FastAPI entry, CORS, router mount
│   ├── requirements.txt
│   ├── render.yaml
│   ├── .env.example
│   └── app/
│       ├── api/v1/query_handler.py     # /search, /answer (SSE), /session
│       ├── core/settings.py            # pydantic-settings
│       ├── services/
│       │   ├── legal_search_service.py # Parallel Kanoon + IndiaCode + Google
│       │   ├── content_extractor.py    # BS4 stripping
│       │   ├── query_classifier.py     # case_law / statute / general
│       │   └── language_model.py       # Cloudflare LLM, SSE, query rewrite
│       ├── models/                     # Pydantic request/response models
│       └── utils/                      # rate_limiter, session_store, citations
│
└── frontend/
    ├── package.json
    ├── vite.config.ts
    ├── netlify.toml
    ├── index.html
    ├── public/favicon.svg              # gold scales-of-justice
    └── src/
        ├── App.tsx                     # Root layout, three-pane orchestration
        ├── index.css                   # Theme tokens (light + dark)
        ├── types.ts
        ├── hooks/useTheme.ts           # localStorage + prefers-color-scheme
        ├── services/api.ts             # Streaming fetch client
        └── components/
            ├── LeftSidebar.tsx         # Threads + Workbook
            ├── Brief.tsx               # Legal memo renderer
            ├── SourcesPanel.tsx        # Sources + Videos
            ├── CaseCard.tsx            # Unified source card
            ├── VideoCard.tsx
            ├── SearchBar.tsx
            ├── CitationChip.tsx
            ├── JurisdictionBadge.tsx
            ├── DevBanner.tsx           # "LIVE · DEV" marquee
            ├── ThemeToggle.tsx
            ├── DisclaimerFooter.tsx
            ├── LoginPage.tsx           # Clerk-backed splash
            └── CollapsedRail.tsx       # 48px icon rail

---

Running locally

Prerequisites

• Python 3.10+
• Node 20+ and npm
• A Cloudflare AI Workers account (API token + account ID)
• An Indian Kanoon API token
• A Google API key with Custom Search + YouTube Data v3 enabled, and a CSE cx
• A Clerk project (publishable key + secret key)

1. Backend

cd backend
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env    # then fill in your tokens
uvicorn main:app --reload --port 8000

Health check: curl http://localhost:8000/api/v1/health → {"status":"ok"}

2. Frontend

cd frontend
npm install
cp .env.example .env    # set VITE_API_URL and VITE_CLERK_PUBLISHABLE_KEY
npm run dev

Open http://localhost:5173, sign in, and ask a question.

---

Environment variables

Backend (backend/.env)

Variable	Required	Purpose
CLOUDFLARE_API_TOKEN	✅	Cloudflare AI Workers auth (also accepts CLOUDFLARE_API_KEY)
CLOUDFLARE_ACCOUNT_ID	✅	Cloudflare account ID for the Workers endpoint
INDIAN_KANOON_API_TOKEN	✅	Indian Kanoon REST API token
GOOGLE_API_KEY	✅	Google Custom Search + YouTube Data v3
GOOGLE_SEARCH_CX	✅	Custom Search Engine ID
YOUTUBE_API_KEY		Optional — falls back to GOOGLE_API_KEY
CLERK_SECRET_KEY	✅	Server-side Clerk verification
ALLOWED_ORIGINS		Comma-separated CORS origins (default: http://localhost:5173)
ENV		dev or prod
SESSION_TTL_SECONDS		In-memory session TTL (default 600)
RATE_LIMIT_CALLS_PER_MIN		Per-session rate limit (default 30)
MAX_SEARCH_RESULTS		Per-provider cap (default 10)

Frontend (frontend/.env)

Variable	Required	Purpose
VITE_API_URL	✅	Backend base URL
VITE_CLERK_PUBLISHABLE_KEY	✅	Clerk client publishable key

---

API

POST  /api/v1/search/{session_id}   → rewritten-query + results + videos
POST  /api/v1/answer/{session_id}   → SSE stream of tokens + final frame
DELETE /api/v1/session/{session_id} → clear server-side session state
GET   /api/v1/health                → liveness probe

/answer SSE schema:

data: {"chunk": "The offence of cheque dishonour…"}
data: {"chunk": " falls under Section 138…"}
…
data: {"done": true, "citations": [...], "suggested_steps": [...]}

---

The Harvey persona

The system prompt asks the LLM to produce a memo with a fixed structure:

1. Issue — the legal question, formally framed
2. Governing Law — Acts, Articles, Sections
3. Authorities — cases in AIR / SCC format (one per line)
4. Discussion — how the law applies to the facts
5. Conclusion — a single callout paragraph
6. Recommended Actions — 3–5 concrete, imperative next steps
7. Follow-up Questions — 3 related questions under 12 words each (rendered as chips)

Never invent citations. Always end with the disclaimer.

---

Roadmap

• Citation verifier — re-fetch every AIR/SCC/Section citation and programmatically check the case name / section number exists; flag hallucinated cites in the UI.
• Semantic cache — Redis vector cache over recent queries to cut repeat-query LLM spend.
• Hybrid retrieval — BM25 + embedding + cross-encoder rerank over the merged result set, scored with nDCG@10.
• Eval harness — 30–50 golden questions with expected authorities, run in CI.
• Observability — OpenTelemetry traces and a Grafana board for p99 latency and token cost per query.

---

Deploy

• Frontend → Netlify. netlify.toml points to frontend/ with npm run build.
• Backend → Render. render.yaml declares a web service from backend/ with gunicorn -k uvicorn.workers.UvicornWorker main:app.

Update ALLOWED_ORIGINS on Render to your Netlify domain before promoting.

---

Credits

Built by Paritosh Tripathi. Architecture borrowed from MiniPerplexity; repurposed for Indian law.

License

MIT — see LICENSE.