The collection graph

Every knowledge collection gets a graph: its documents, their sections, the people, organizations and places they mention, and the sections that say similar things. It shows what a collection holds and where an answer’s passages came from. It is on the collection’s Graph and search tab.

Who can open it

Role	What they can do
Admin	Explore the graph, ask and search the collection, Rebuild graph, and change People in the graph.
Builder	The same as an Admin.
Auditor	Explore the graph. Cannot ask, search, rebuild or change the setting.
Reviewer	Cannot open it.
End user	Cannot open it.
Platform admin	Everything an Admin can, for any tenant.

What is in the graph

Node	What it is
Collection	The collection itself.
Document	Each document loaded into it.
Section	Each section of a document, from its headings down to the fourth level; a section shorter than 200 characters joins the one before it. A text with no headings is cut into parts of about 4,000 characters, shown as Part 1, Part 2… The text before a document’s first heading is Opening of document.
Entity	A named thing a section mentions: Person, Organization, Place, Product, Event, Law or policy, Work, Group of people, or Other entity.
Group	A cluster of closely linked nodes, used when the graph is too large to draw at once.

Link	What it means
Structure	A collection holds a document; a document holds a section; a section holds a subsection.
Mentions	A section mentions an entity.
Similar	Two sections of different documents say similar things. Inferred.
Related	Two entities are mentioned together. Inferred.

Similarity comes from the embeddings the collection already holds: two sections are linked when their cosine similarity is 0.75 or more, or the embedding model’s own retrieval threshold when that is higher. No language model is involved in building the graph, and building it spends no tokens.

Named entities are found by the guardrails service. When it does not answer, the graph is built with documents, sections and similar sections only, and says Named entities need the guardrails service; entities join at the next build once the service answers.

What you see

On the collection’s Overview tab, a Graph card summarizes it: when it was Built, its Size (nodes, links and groups), its Structure (documents and sections), and its Entities by kind. Rebuild graph is there for Admins and Builders.

The Graph and search tab has two regions:

1. The canvas, with its controls: Zoom in, Zoom out, Fit to the view, Clear highlight and Full screen; the Legend; and the relations drawn, one checkbox each for Structure, Mentions, Similar and Related, plus Only explicit (hide inferred). Inferred links are drawn lighter. Hovering a node shows its name, its kind and where it sits.
2. The side panel: Ask and Raw vector search at the top, for Admins and Builders, then Selected node. Until you choose a node it says: “Choose a node in the graph, or a search result, to see what it is and what it connects to.”

When a node is selected, the panel shows what it is and what it connects to: a document’s sections, a section’s subsections, Entities mentioned, Similar sections, an entity’s Mentioned in and Related entities, with its chunk, page and mention counts. Long lists end with “And n more”.

States:

• Building the graph…: a build is running.
• No graph yet, “Built after the next change, or rebuild now.”, with Rebuild graph.
• The graph is empty: the last build found no documents.
• The last build could not make the graph: the reason, and the graph built before it, which stays shown.
• The graph could not be loaded.

What you can do

Explore it

1. Open the collection, then Graph and search.
2. Choose a node, or a search result, to select it and see its details in the side panel.
3. Choose Show its connections to draw only the node and its neighbours. Under How far, choose One step or Two steps. Back to the whole graph returns.
4. On a section or a document, choose Open document to open it on the collection’s Documents tab.

Explore a large graph

A graph of more than 5,000 nodes opens on its groups instead: “This graph has n nodes, more than the 5,000 drawn at once, so it opens on its n groups.” The side panel lists the 24 largest groups.

1. Choose a group, on the canvas or in Groups, to draw its nodes.
2. On a selected node, Show its group opens the group it belongs to.
3. Back to all groups returns to the overview.

See where an answer came from

1. In the side panel, use Ask or Raw vector search. See Collections.
2. The sections the matched passages came from light up in the graph. For an answer, choose Show in the graph on a source to select its section.

If an answer’s passages do not say which section they came from, the panel says so; a raw search always lights its matches.

Rebuild the graph

Olympusx AI rebuilds a collection’s graph by itself after the collection changes: a pipeline run that loaded or removed something, an upload, a pipeline deleted, or a change to People in the graph. It waits until the collection has been quiet for 20 seconds, and builds no more often than once every 5 minutes.

To rebuild now:

1. Open the collection.
2. Choose Rebuild graph on the Overview tab or the Graph and search tab.

Olympusx AI says “The graph is being rebuilt. It shows here when it is done.” If a build is already running, it says so and starts no second one. A failed build never fails what triggered it.

Leave people out

1. Open the collection’s Settings tab.
2. Under Graph, turn People in the graph off.

The next build leaves out the names of people the documents mention. Searching, asking and the documents themselves are unchanged. Turn it back on to include them at the next build.

Code collections

A collection the Code Assistant indexes has no graph of its own here. The tab says A code collection’s graph is on its repositories’ pages, and keeps the search. See Repository view.

Limits

What	Limit
Nodes drawn at once	5,000; larger graphs open on their groups
Groups listed in the side panel	The 24 largest
Neighbours drawn per step in Show its connections	60 per column; the rest are counted
Section size without headings	About 4,000 characters per part
Similar sections	Cosine similarity of 0.75 or more, or the model’s own threshold when higher
Automatic rebuild	After 20 quiet seconds, at most once every 5 minutes

Related

• Collections
• Pipelines and connectors
• Guardrails
• Repository view
• Knowledge and grounding