RTI Wiki AI — Free, Private, Open Source

An RTI Wiki AI — chat about the RTI Act, 2005 that runs entirely in your browser. Ask questions in plain English — fees, timelines, grounds for refusal, case law, appeal procedure, state-specific rules, the DPDP amendment — and the chatbot answers, grounded in the pages of this site. Free as of now. No sign-up. No API key. No data leaves your device.

1. Open the page; the app downloads one open-source model — Llama 3.2 1B Instruct, quantised by MLC — about 0.9 GB, once, then cached in your browser.
2. You type a question.
3. Before answering, the app retrieves the five most relevant passages from a 410-entry RTI Wiki corpus — the Act, guides, and case pages — and feeds them to the model as context.
4. The model runs on your device's GPU via WebGPU and streams an answer.
5. Answers that draw on retrieved pages end with auto-linked Sources, plus follow-up question chips. The chat remembers the last few turns for context.

The model is primed with the RTI Act framework — the ten grounds of rejection, §19 appeals, §20 penalties, key sections, and the Digital Personal Data Protection Act, 2023 §44(3) amendment that substituted §8(1)(j) of the RTI Act (in force 14 November 2025) — before your question ever reaches it, and the retrieval layer then anchors every answer to a real page on this site.

• Not legal advice. A 1-billion-parameter model can be wrong. Cross-check with statutory text and the curated pages before you file anything.
• Not a substitute for the statute. When the answer matters, open the section page it cites — the Act index — and read the actual words.
• Not up-to-the-minute. The model's prior knowledge has a training cut-off; the retrieval corpus is refreshed with the site, but new circulars may lag. For the latest, check our blog.

• WebGPU — Chrome 113 or newer, Edge, or any WebGPU-capable browser (Chrome shipped WebGPU by default in version 113, May 2023). Older browsers and some mobile devices cannot run the model.
• First-run download — about 0.9 GB for the model, plus the corpus file. One-time; cached thereafter.

• The question you type never leaves your browser — the model runs locally.
• The model weights come from a public CDN (HuggingFace via WebLLM), but only once — cached afterwards.
• The retrieval corpus loads from this site, exactly like any page.
• No tracking, no analytics, no logging of your questions on our side.

• Llama 3.2 by Meta under the Llama community licence.
• Quantization by MLC AI.
• Runtime — WebLLM (Apache-2.0).

Yes. Free as of now, no sign-up, no API key, no message limits. The model runs on your own device, so there is no per-question server bill to pass on to you.

No. Your question is processed inside your browser. The only network fetches are the one-time model download from a public CDN and the RTI Wiki corpus from this site.

Llama 3.2 1B Instruct (MLC quantisation), served by the WebLLM runtime. Small enough to run on a laptop GPU, large enough to stay coherent on legal questions when anchored by retrieved passages.

The tool needs WebGPU, available in Chrome 113+ and Edge. Update the browser; on Linux and some older machines WebGPU may still be behind a flag.

The system prompt includes the DPDP Act 2023 §44(3) change to §8(1)(j) that took effect 14 November 2025, and retrieval pulls live pages. For court challenges to that amendment, read our DPDP guide.

It explains what to ask for; use the AI RTI Drafter to produce the ready-to-file application, and the PIO Reply Checker to grade a reply you received.

If your mobile browser supports WebGPU, yes. On phones without it, use the Question Builder instead — it works on any browser.

Treat it as a research hint, not a citation. Verify against the linked source page, the Act text, or the FAQ — and remember the model's output is not legal advice.

• Question Builder — describe a problem, get the records to ask for.
• RTI Application Generator — produce a ready-to-file RTI.
• PIO Reply Checker — analyse a PIO reply (also has an AI deep-analysis option).
• First Appeal Builder — if the PIO refuses, draft the §19(1) appeal.
• Timeline Tracker — compute your §7 and §19 deadlines.

• Grounds for RTI rejection
• Ask for records, not answers
• RTI FAQ — 25 most-asked questions
• How to file RTI online
• 10 landmark CIC decisions
• The RTI Playbook book
• State-wise RTI fees

• Right to Information Act, 2005 (as amended 2019, 2023)
• Digital Personal Data Protection Act, 2023 — §44(3), in force 14 November 2025
• Meta Llama 3.2 community licence; MLC AI; WebLLM (Apache-2.0)
• WebGPU in Chrome 113 — developer.chrome.com/blog/webgpu-release

Last reviewed: 26 August 2026. Verified the live tool at righttoinformation.wiki/tools/chat-app.html — single Llama-3.2-1B model, WebGPU runtime and 410-entry retrieval corpus read from the app source, and the Chrome 113 WebGPU baseline against developer.chrome.com.