Skip to main content
AI methodology

How NyayNow works today

We describe the deployed architecture and its current limits plainly. This page is updated as retrieval, evaluation, and data partnerships are built and validated.

STEP 01

Understand the question

The API receives the user’s question and selected language. The assistant can also receive a short conversation history and user-provided case context.

STEP 02

Retrieve when configured

When an Indian Kanoon API key is configured, selected routes search Indian Kanoon and pass a small number of returned snippets and links to the model. Without that key, this retrieval step is skipped.

STEP 03

Generate and disclose limits

The backend calls Google Gemini through a model fallback helper. The answer is generated text and can be wrong or incomplete; users should open cited sources and ask a qualified advocate about case-specific decisions.

What is implemented

  • Gemini-based generation in the Node API.
  • Optional Indian Kanoon search in selected legal routes, when configured.
  • Some routes request Google Search grounding through Gemini; availability depends on model and API configuration.
  • A separate FastAPI service exists, but the main assistant route inspected here uses the Node service.

What is not yet substantiated

  • No public, reproducible benchmark currently supports a legal accuracy percentage.
  • The inspected assistant path does not implement a custom vector database or embedding retrieval pipeline.
  • Corpus size, update cadence, and comprehensive court coverage are not independently verified here.
  • AI output is not a substitute for legal advice and does not predict a court outcome reliably.

Evaluation before stronger claims

A credible legal AI evaluation should publish the dataset provenance and cutoff date, task definitions, citation-level scoring rules, expert review process, sample size, baselines, error categories, and confidence intervals. NyayNow will publish performance claims only after that protocol is run and independently reviewable.

User safety

Treat generated content as an informational starting point. Verify every statute and judgment against its source, especially before filing, signing, paying, or taking urgent action. Do not enter highly sensitive personal information unless you have reviewed the applicable privacy terms.

Read the legal disclaimer
AI Methodology & Evaluation | NyayNow