Recourse

Free web tool for Indian tenants to draft a statute-grounded legal notice when a landlord refuses to return a security deposit.

v1 state coverage:

• Full state-law grounding: Karnataka (KRA 2001), Maharashtra (MRCA 1999), Delhi (DRCA 1958), Tamil Nadu (TNRRRLT 2017), Telangana (Buildings Act 1960), West Bengal (WBPTA 1997), Rajasthan (RCA 2001), plus the Model Tenancy Act 2021 for Andhra Pradesh / Uttar Pradesh / Assam
• Pan-India fallback (Consumer Protection Act 2019 + Contract Act s.74 + CPC s.80) for any other state, with a limitation flag noted in the draft.

Retrieval quality: 43 verbatim Bare Act section chunks in the Upstash Vector index; 15/15 canned smoke queries hit the expected act + section on the top-5 retrieval.

Stack

Next.js 16 (App Router, Node runtime), TypeScript, Tailwind 4, Vercel AI SDK v6 + @ai-sdk/google, Gemini 2.5 Pro + gemini-embedding-001 (1536-dim), Upstash Vector + Redis + Ratelimit, @react-pdf/renderer, vitest + @playwright/test.

Local setup

1. cp .env.example .env.local and fill in Google AI Studio + Upstash Vector + Upstash Redis credentials.
2. npm install
3. Populate the corpus (see Bare Act text below), then ingest and smoke-test:

   npm run corpus:ingest
   npm run corpus:smoke   # must pass 10/10 before agent will produce usable drafts

4. npm run dev

Scripts

• npm run corpus:fetch — best-effort fetch of statute text from indiacode.nic.in into data/corpus/. Sections that cannot be auto-extracted are left as [PLACEHOLDER — human must paste verified Bare Act text here.] markers.
• npm run corpus:ingest — embed each section with gemini-embedding-001 and upsert to Upstash Vector namespace v1. Refuses to run if any placeholder is present.
• npm run corpus:smoke — 10 canned queries; asserts each pulls the expected Act + Section into top-5.
• npm test — vitest unit + integration.
• npm run e2e — Playwright.

Bare Act text

data/corpus/*.md must contain verbatim Bare Act text under ## SECTION N headers. The citation verifier compares the model's citations against these exact chunks — paraphrasing breaks grounding.

For each of the eleven acts:

• Karnataka Rent Act, 2001 (ss.9, 12, 27, 30, 34)
• Model Tenancy Act, 2021 (ss.11, 12, 21, 32)
• Maharashtra Rent Control Act, 1999 (ss.7, 24, 33, 55)
• Consumer Protection Act, 2019 (ss.2(7), 2(11), 2(42), 35, 38)
• Indian Contract Act, 1872 (s.74)
• Code of Civil Procedure, 1908 (s.80)
• Delhi Rent Control Act, 1958 (ss.4, 6, 14, 27, 50)
• Tamil Nadu R.R.R.L.T. Act, 2017 (ss.4, 6, 21, 30)
• Telangana Buildings (LR&E) Control Act, 1960 (ss.4, 8, 9, 10, 15)
• West Bengal Premises Tenancy Act, 1997 (ss.6, 7, 17, 21, 26)
• Rajasthan Rent Control Act, 2001 (ss.9, 10, 13, 22)

Open the source_url in data/corpus/manifest.json, copy each section from the official PDF, and replace the [PLACEHOLDER …] markers. Verify with:

grep -l "PLACEHOLDER" data/corpus/*.md && echo "still have placeholders" || echo "OK"

Non-negotiables

• Every legal citation is verified against retrieved corpus chunks before rendering. If verification fails 3× the draft is rejected.
• No output contains "legal advice", "lawyer", "attorney", "counsel" — enforced at prompt + post-generation sanitizer.
• Session-only. No auth, no payments. 24h Redis TTL, 10 drafts/IP/day.

Deploy

GitHub first; Vercel auto-deploys from main. Do not run vercel deploy manually. Vercel env vars required: GOOGLE_GENERATIVE_AI_API_KEY, UPSTASH_VECTOR_REST_URL, UPSTASH_VECTOR_REST_TOKEN, UPSTASH_REDIS_REST_URL, UPSTASH_REDIS_REST_TOKEN.

Not legal advice.