Inventory

Inventory ingests one or more source repositories into a knowledge graph — files, functions, types and the relations between them — that you can filter, browse and ask questions of. Like Wikis, it lives at its own Inventory area, addressed by the Inventory toolkit that built the graph.

Prerequisites

• An Inventory toolkit with at least one source added.
• A source toolkit of an allowed type — by default GitHub or ADO Repos (a deployment can widen this with ELITEA_INVENTORY_SOURCE_TYPES; GitLab and Bitbucket are not supported as sources even when connected as toolkits elsewhere).
• An llm_model set on the Inventory toolkit — ingestion is refused without one.

Build a knowledge graph

1. Add an Inventory toolkit

   If the project has none yet, add one — the Inventory area prompts you to when it finds none for the project.

2. Add a source

   Under Sources, add a source toolkit (GitHub or ADO Repos, unless the deployment allows more) and its branch.

3. Run ingestion

   Select Run ingestion. Progress shows per source: Pending, Ingesting…, Done, or Error. You can Stop ingestion on a source that’s running.

Inventory entity browser with graph filtersscreenshot pending (inventory-browser)

Browse the graph

Filter entities by source, type and layer, and search by name. Select an entity to see its file location and its relations — incoming and outgoing edges to other entities. A source removed from the toolkit but still represented in the graph shows as “No longer configured — its entities are still in the graph,” rather than disappearing silently.

Ask about the inventory

Select Ask this inventory to open a question panel over the graph. Answers name the entities they came from, rather than answering from general knowledge.

Statistics and maintenance

The Stats view breaks the graph down by layer and entity type, and reports cache size and cached-graph count. Maintenance normalizes type-name variations (for example “Feature” vs. “feature”) and rebuilds the name, type, file and source indices.

Common issues

Note

Inventory runs on the deployment’s engine sidecar, not the Rust/Python agent worker — ingestion and browsing work the same regardless of which worker your deployment runs for agents and toolkits.

Note

“This toolkit configures no LLM model, so ingestion is refused” means you must set llm_model in the Inventory toolkit’s settings before Run ingestion will do anything.

Note

If the ingestion status can’t be read, the per-source states shown are stale, not necessarily current — re-check after a refresh rather than assuming a source is still Pending.