Sarvam structures enterprise knowledge so agents can navigate and cite it

Its knowledge fabric combines document navigation with a domain ontology, addressing questions that require complete sets of records.

Sarvam has described a knowledge fabric intended to help enterprise AI agents work with information spread across documents and business systems. The live research page dated October 26 presents two layers: a navigable knowledge tree and an ontology that records domain entities, properties and relationships.

The tree gives an agent ways to inspect indexes, move through documents and read relevant pages. The ontology supports structured filtering and joins across records. Sarvam says citations lead back to the source material, including passages and table rows.

Finding examples is different from finding every match

The problem is especially visible in questions that ask for a complete list. Finding several relevant agreements does not show that every qualifying agreement has been found. A structured representation can make completeness easier to test, provided its extraction and classification are themselves accurate.

Sarvam reports benchmark improvements and lower task costs in its own comparisons. Those results are evidence about the company’s tested configurations; they do not establish equivalent performance for an unfamiliar customer’s files.

A measured legal-document task

On Harvey’s LAB benchmark, Sarvam reports that answers using GLM 5.3 met 84.2% of expert rubric criteria, against 55.1% for its agentic RAG comparison. Average task cost fell from $0.64 to $0.28 in that configuration. The simulated law-firm corpus contains 9,288 files across 266 matters for 46 clients.

The example illustrates why completeness is a data-organisation problem: a request for all executed agreements needs correct entities and relationships, not just a handful of relevant passages. These are Sarvam’s results for a specified setup, rather than an independent customer deployment. Keeping extracted relationships and source citations current remains central to the approach.

Sources

- [How we built the knowledge fabric for enterprise agents](https://www.sarvam.ai/blogs/knowledge-fabric-enterprise-agents) — Sarvam AI