Writ combines a fine-tuned language model with an encrypted on-device knowledge base — judgments, statutes, curated legal notes — retrieved through hybrid keyword and neural search. No account. No cloud. Queries never leave your machine.
Capabilities
Every architectural decision prioritises privacy, depth, and precision over convenience and cloud dependency.
The complete inference pipeline — model loading, retrieval, generation, and re-ranking — runs on your local hardware via Ollama. Nothing is transmitted to any external server. Your research stays yours.
Judgments, acts, and curated legal notes stored encrypted on-device. No plaintext exposure at rest.
BM25 keyword search combined with neural re-ranking surfaces the most relevant passages, not just the most frequent matches.
Writ runs Vidhi (Llama 3.2 3B fine-tune), a model trained specifically on Indian legal texts. It understands the structure of judgments, identifies precedents, and reasons about statutory provisions in context.
Save and restore multi-turn research sessions. Annotate retrieved passages and build a structured research trail — stored locally.
A focused interface designed for long-form legal reading. Light and dark modes. No telemetry, no banners, no feature clutter.
Architecture
Four steps from your prompt to a grounded, cited legal response — all on your machine.
Submit a natural-language legal question or a case-law lookup. No query syntax required.
BM25 keyword search across the encrypted knowledge base is fused with a neural embedding pass and candidates are scored.
A cross-encoder reorders results by semantic relevance, surfacing the most on-point paragraphs from judgments and statutes.
Vidhi receives the top passages as context and generates a cited, structured response entirely on your hardware.
"The right to access justice should not require trusting a third-party server with the most sensitive details of a legal matter." — Writ, built for privacy-first legal research
Get Started
Free for personal, non-commercial use. No account required. Windows, macOS, and Linux.
Writ-Setup-1.0.0-Windows.exeWrit-1.0.0-Mac.dmgWrit-1.0.0-Linux.tar.gz
Prerequisites. Writ requires Ollama running locally for model inference. On first launch, the Vidhi model (~2.5 GB) is downloaded automatically. The knowledge base (~1.6 GB) is fetched during installation if not already present.
Status
Writ is a working proof of concept. The framework — offline inference, hybrid retrieval, encrypted on-device knowledge base — is real and functional. But it is not a polished, production-grade product. Expect rough edges, inconsistent outputs, and missing coverage.
The retrieval model and fine-tune were trained on a limited dataset. The knowledge base does not cover all Indian statutes or judgments comprehensively. Results will vary depending on the specificity of your query.
If you are interested in a properly built, production-ready version — see below.
Legal Disclaimer
Writ does not provide legal advice. Nothing generated by this tool constitutes legal counsel, legal opinion, or a substitute for advice from a qualified advocate.
Always consult a licensed lawyer before acting on any legal matter. AI-generated responses can be wrong, incomplete, outdated, or contextually inapplicable to your specific situation.
Citations surfaced by the retrieval system must be independently verified against authorised sources before reliance.
License
Writ is licensed for personal, non-commercial use. You may install and run the software on your own device for personal legal research. Redistribution, modification, or commercial use requires explicit written permission from the author.
Licensing inquiries →Collaboration
The architecture is proven. If you represent a legal firm, institution, or organisation that needs a production-grade offline legal intelligence system — expanded knowledge base, custom fine-tuning, enterprise deployment — reach out.