Refresh ingestion

Use a focused refresh after a reviewed summary changes. Start without PDF downloads, then widen the scope only when the task requires full text.

If you are building a store for the first time, follow Ingest reviewed content instead.

Execution status

Documentation CI strictly builds this page, but it does not run the credentialed Cognee refresh or retrieval procedures. Run and verify them manually against the intended store; Drive-backed refreshes additionally require authenticated Drive access.

Validate and preview

uv run cortex-validate
uv run cortex-ingest --paper <slug-prefix> --skip-drive --dry-run

--paper matches the summary path case-insensitively. Choose a prefix specific enough to identify the intended record and confirm the dry-run count.

Refresh from committed summaries

uv run cortex-ingest --paper <slug-prefix> --skip-drive

This path needs an Anthropic API key but no Drive credentials. It adds the selected summary text to Cognee, updates the graph and vector state, and runs post-ingestion alias resolution.

Include source PDFs when necessary

Omit --skip-drive only when full paper text is part of the intended corpus:

uv run cortex-ingest --paper <slug-prefix>

PDF retrieval additionally requires CORTEX_DRIVE_ACCOUNT, a valid drive_id in papers/index.yaml, and authenticated access through gog. Downloaded PDFs use a temporary cache and are not documentation assets.

Refresh the full corpus

Remove --paper to process every indexed entry:

uv run cortex-ingest --skip-drive --dry-run
uv run cortex-ingest --skip-drive

Treat a full refresh as an operator event: record the source commit, selected store path, flags, and result so users can determine the currency of the local index.

Use reset only for an intentional rebuild

--reset prunes Cognee data and metadata before ingestion. Confirm the configured store and retain any required recovery state before running it. A content refresh normally does not need this flag.