NyayaAI, an AI-powered legal assistant developed by researchers at UPES Dehradun, is designed to answer legal questions in plain language. A lawyer researching a case precedent, a law student navigating a constitutional provision or a member of the public who cannot afford legal consultation can type a question on topics ranging from tenancy rights to criminal procedure and contract disputes, and receive a referenced answer.

The system can also summarise lengthy documents, identify relevant precedents and generate preliminary drafts of legal documents.

These developments have been captured in a paper titled ‘NyayaAI: An AI-powered legal assistant using multi-agent architecture and retrieval-augmented generation’, published in arXiv.

Legal help in India remains out of reach for most citizens. Consultation costs are prohibitive for many and legal services are concentrated in cities. The statutes are in English, a language in which the majority of Indians are not fluent. India has about two million registered advocates, most of whom are in urban centres.

Existing platforms such as Manupatra, SCC Online and Indian Kanoon operate on keyword search across digitised legal documents. While paid platforms have been adding AI-assisted search in recent years, none offers the integrated multi-agent task handling or RAG-grounded response generation that NyayaAI provides. AI legal assistants available internationally, including Harvey AI, CoCounsel and LexisNexis Protege, are built around American or British law.

Beyond keywords

NyayaAI addresses a documented failure mode in legal AI: Systems relying on training data alone have generated responses citing cases that do not exist, leading to sanctions against lawyers in US courts.

Rather than generating from memory, the NyayaAI system searches a vector database of Indian legal documents using semantic similarity matching, retrieving provisions and precedents relevant to the query’s meaning rather than its keywords, before generating a response grounded in those documents.

The system uses a multi-agent architecture orchestrated through the Mastra TypeScript framework. A main agent classifies incoming queries by legal domain, covering constitutional, criminal, civil, family and corporate matters, and routes complex queries to specialist sub-agents handling legal research, document summarisation, case analysis and drafting assistance. The drafting sub-agent generates preliminary versions of legal documents, including petitions, notices and agreements, which lawyers or users can review and modify. A compliance module intercepts all responses before delivery, validating them against legal, ethical and jurisdictional criteria and appending disclaimers where needed.

Success rates

In evaluations, the system achieved 72 per cent response accuracy and 74 per cent retrieval precision. Domain classification precision reached 70 per cent across the five categories.

Criminal law and constitutional law recorded the highest precision, at 75 per cent and 73 per cent, respectively, which the authors attribute to cleaner terminology and stronger knowledge base coverage.

Corporate law recorded the lowest precision at 65 per cent. The reason is jurisdictional overlap: A question about a company’s acquisition, for instance, may fall simultaneously under the Companies Act, SEBI regulations, RBI guidelines and FEMA provisions, each administered by a different regulatory body with its own rules. The system must determine which framework applies before it can retrieve the right documents.

Causes of failure

Error analysis identified legal jargon complexity as the primary source of failure, accounting for 35 per cent of errors, given that the system is designed partly for users without specialist legal training. Jurisdictional ambiguity contributed 28 per cent, reflecting the difficulty of handling queries that span multiple regulatory regimes. Context misunderstanding, where the system responded to a different question than the one asked, contributed 22 per cent. Out-of-domain queries accounted for the remaining 15 per cent.

The paper’s authors identify several directions for future development, including integration with the Supreme Court judgement repository and the India Code portal, multilingual support for Hindi and other regional languages, real-time update pipelines to keep the knowledge base current with evolving legislation, predictive analytics for legal outcomes based on historical case law, and support for user-uploaded documents to enable AI-assisted review of case-specific materials. The code has been made available to the public.

Published on June 15, 2026