InferenceFort
Guide
The attribution graph
A map of who actually drove spend. You build it by adding the users, teams, or
models you care about; the graph then pulls in everything they're connected to and totals the cost.
The three layers
Every node is one of three kinds, left to right:
USERa person who made calls (their email / SDK identity)
TEAMthe team a call was attributed to
MODELthe LLM the call used (e.g. gpt-4o, gemma3:4b)
How to use it
1
Search and add a node. Type a user, team, or model in the search box and click a result. It drops onto the canvas. Use the All / Users / Teams / Models toggle to narrow what the search returns.
2
The graph fills in its connections. Adding a model pulls in the users and teams that actually called it; adding a user pulls in the teams and models it used. Add several nodes to compare.
3
Click a node to trace its lineage. Everything connected to it lights up; everything else dims — so you can see exactly what a user, team, or model touches.
4
Read the spend. The panel on the right totals the cost attributed to your current selection, broken down per entity.
5
Clear and start over with the Clear button any time.
Solid vs. dashed lines
This is the important bit — the lines mean two different things:
Solid = actually used. A real call happened on this path. These are what drive the spend totals.
Dashed = allowed but idle. Policy permits this team to use this model, but no call has happened. Capability, not usage.
The canvas only grows along solid (usage) lines. So a model never appears connected to people who never actually called it — what you see is real attribution, not policy theory.
← Close this tab