GraphRAG and connected document search

Follow connections across documents and return to the evidence for each one.

Planned graph-based retrieval for related entities and evidence across documents, with a planned comparison of indexing approaches.

Planned capability. This page describes product direction. This feature is not available in the current trial.

Where the product is today

Today, workspace search finds text and links matches to their source pages.
Graph indexing, relationship traversal and generated graph-grounded answers are not available.

Connect a supplier, order and invoice

The planned workflow would find the source records connecting a supplier to an order and invoice, distinguish explicit facts from inferred links, and cite each supporting passage.

Availability and evaluation

This is a product roadmap. The current trial does not include GraphRAG.
FastGraphRAG and the separate fast-graphrag library are implementation candidates, not installed product features.
No query-latency, indexing-cost or relationship-accuracy claim has been measured for ParseMill.

What we plan to build

Create evidence-backed entities and relationships from versioned document chunks.
Resolve entity aliases while keeping uncertain or conflicting links visible.
Combine lexical and semantic retrieval with bounded graph traversal.
Compare a faster NLP-based graph index with model-extracted relationships.
Update and delete affected graph evidence when source documents change.

How this should be evaluated

Compare graph retrieval against a simpler hybrid search baseline.
Measure indexing time, query latency, answer support and cost on the same corpus.
Verify tenant isolation, source deletion and conflicting relationships.