Latin for the legal principle a court articulates when deciding a case. Not the outcome. Not the commentary. The rule that future courts are bound to follow.
Under Article 141 of the Indian Constitution, any ratio declared by the Supreme Court is law across the entire country. This product is named after the thing it actually retrieves.
Whether the appeal was allowed or dismissed is the dispositio. It is not the ratio. Two cases can have identical ratios and opposite outcomes.
Obiter dicta are observations made in passing. Persuasive in future courts, but not binding. Retrieving them as holdings is a structural error.
The ratio is the rule the court articulates as the basis of its decision. Under Article 141, every Supreme Court ratio binds all courts in India. This is what Ratio retrieves.
A single Supreme Court judgment records counsel submissions, lower court findings, distinguished precedents, and the court's own binding holding - often in the exact same paragraph.
Disambiguating counsel submission from binding judicial ratio is not an aesthetic preference - it is a structural legal requirement. Ratio\'s stance detection is built around this distinction.
An interactive map of Ratio's retrieval pipeline. Use the arrows, the dots, or drag to move through it, click any stage to read the engineering rationale.
Structural guarantees where the schema can enforce them, honest measured numbers where it can't - stance labelling reduces hallucinated holdings, it isn't a hard guarantee the way the citation schema is.
Every 450-token chunk is automatically prefixed with a machine-generated 3-sentence summary of its parent judgment - preserving the core issue, ruling, and context in every vector embedding. Chunk size was chosen for defensible reasons; it has not itself been swept against the label set, which is recorded rather than implied.
Standard LLM tokenizers lowercase all text, destroying citation sensitivity ('AIR' → 'air', 'CrPC' → 'crpc'). Ratio indexes each identifier twice: once atomic and case-preserved, once split into lowercase words, so exact and topical matching share one postings list.
Generation prompts enforce a strict JSON output schema. A claim without a source ID cannot be emitted at all, measured at 0% uncited claims and 91.2% citation precision on a 15-query structured-output benchmark - the schema forces attribution, it does not by itself guarantee the citation is right.
Built over the Hugging Face OpenNyAI InJudgements dataset. Supreme Court of India, High Courts, and Tribunals - pre-processed into a 580,939-chunk hybrid search index.
Every chunk carries parent metadata, tribunal tier weight, and full case citations to prevent context dilution during generation.
Ask complex legal questions. Receive verified holdings, verbatim passage citations, and precedent authority scores.
Enter the Workspace→