Ingest reviewed content
This tutorial builds an isolated local store from a reviewed paper summary. It starts with a dry run and skips Drive, so you can inspect the exact content selection before making external model calls.
Execution status
Documentation CI strictly builds this page, but it does not run the credentialed Cognee ingestion or retrieval commands. Run and verify the procedure manually with a disposable store and your own Anthropic key.
Prerequisites
Complete the runtime install from the repository root:
Set an Anthropic key for the real ingestion step:
Choose a disposable directory outside the tracked source tree for this tutorial, then point Cortex at it:
The selected paper must already be registered in papers/index.yaml, and its
summary file must contain reviewed text. This tutorial uses the existing
fengler-2021 slug prefix as an example.
1. Validate the publication record
Resolve errors before ingestion. Warnings describe incomplete metadata such as
a missing Drive identifier; decide whether they matter for the selected path.
Use --strict when you need warnings to fail the command.
2. Preview the selected inputs
The dry run identifies matching summaries and reports how many characters would be sent to the backend. It does not configure Cognee or write a store.
3. Ingest the reviewed summary
--skip-drive limits the input to the committed summary rather than downloading
the corresponding PDF. Cortex writes the resulting graph, vector, and metadata
state to COGNEE_DATA_PATH, then checks new nodes against the alias registry.
4. Verify retrieval
You now have a local store suitable for the first-query tutorial. For later updates to that store, follow Refresh ingestion.
Reset is destructive
cortex-ingest --reset prunes the selected Cognee data and metadata before
rebuilding it. Confirm COGNEE_DATA_PATH before using that flag. A normal
focused refresh does not require it.