Knowledge Graph

← User Guide · 简体中文

The graph index extracts entities (vertices) and relations (edges) from documents, complementing the vector index: vectors recall by semantic similarity, graphs organize by entity relations — good for multi-hop questions ("which policies cover the department of A's product lead").

Generating a Graph

Graphs are extracted per document, consuming tokens of the base's ingest model. Three entry points:

Entry	Where
Check "Graph"	The index dialog when importing
Graph button	Document list actions (disabled until extracted)
"Click to generate graph"	The segment management page's document card

Using the segment page:

1. Click "Click to generate graph";
2. The confirmation names the model (translated): "The knowledge graph will be extracted using the base's ingest model '{model}'; this consumes that model's tokens. Start?";
3. Status shows extracting (background task) and refreshes on completion.

Graph confirm

Viewing the Graph

1. Click View graph from the document list or segment management;
2. The canvas shows vertices and relations;
3. Large graphs load in pages: "Vertices {loaded}/{total}, relations {loaded}/{total}" with a Load more button; Re-layout re-arranges the canvas;
4. Other states show hints ("extracting, refresh later" or "no graph data yet").

Knowledge graph

Graph Citations in Answers

After asking against a base with a graph index:

1. Click the Graph button under the answer;
2. The dialog shows the graph fragments referenced (entities with name/description, relations).

Characters linked to the base support this too. Together with Citations (hit segments) this fully traces the answer — see Chat Window · Memory & Citations.

Maintenance Notes

WARNING

Graphs do not auto-update with body edits: after switching segmentation mode or editing the body, re-extract manually, or answers keep citing the old graph.

• Disabling a segment deletes its graph footprint; re-extract after re-enabling if needed (see Segment Management);
• Graph cost scales with document length — prioritize entity-dense material (org structures, product specs, policy clauses); narrative text is fine with vectors only.

---

Previous: Segment Management · Next: Knowledge Base Q&A