The tree organizes a corpus into folders, indexes, document summaries, page ranges and tables; the ontology represents domain entities and their relationships.
According to the company, the tree is served over MCP. An agent can list folders, read indexes, search the corpus, open page ranges and inspect tables. Sarvam says citations link back to a passage, image or table row in the original source. Its parsing pipeline converts files to Markdown and uses Sarvam Vision 2.1 for scanned pages, complex layouts, tables and Indic scripts.
The post reports 84.2% of expert-rubric criteria met on Harvey LAB with GLM 5.3, compared with 55.1% for agentic RAG; it reports average task cost of $0.28 versus $0.64. On Sierra's tau-knowledge banking benchmark, it says the fabric matched baseline accuracy with 64% fewer tokens using GLM 5.2. It also reports navigation gains across four public benchmarks, including 91.3% versus 75.5% on BrowseComp-Plus.
The source does not establish general availability, pricing, customer deployment details, or independent validation of the benchmark comparisons.