Skip to main content

Create a catalog

The catalog is the grounded knowledge agents cite. Building it means adding curated source content and running the ingest pipeline for the domains that consume it. For the concepts and lifecycle, see Catalog.

1. Add curated source content​

Curated material lives under app/content/catalog-content/ as JSON. Add a file for your source, for example my-standard.json, containing the reference records you want agents to ground against.

Keep the content curated and stable — this is the deterministic layer. Regulatory excerpts, framework controls, and reference taxonomies belong here; per-run conclusions do not.

2. Point a domain's sources at it​

In the consuming domain's catalog.yaml, add a source that reads your content and normalizes it into a typed entity (see Create a new domain for the full source/normalize shape). For http sources, set the URL; for authored content, use transport: static with inline data.

3. Choose whether to synthesize​

  • synthesis: none — the catalog holds exactly the ingested entities.
  • A synthesis stage — agents derive higher-order entities (for example a risk type) from the ingested facts, cluster by cluster. Synthesis-enabled domains take longer to build because each cluster invokes a Bedrock agent.

4. Build and promote​

make seed                                   # publish content first
make ingest DOMAINS="my-domain"             # build this domain's catalog

The build runs: ingest → (synthesis) → validate → promote. The new version goes live only if it passes the coverage gate.

# Large, synthesis-heavy domains may need a longer poll cap
POLL_TIMEOUT_MIN=90 make ingest DOMAINS="my-domain"
caution

make ingest is long-running and idempotent. If a build exceeds the poll cap it continues in AWS but isn't promoted by that invocation — re-run make ingest DOMAINS="my-domain" to finish and promote. Enable Bedrock model access first.

5. Verify agents can ground against it​

Launch an ecosystem whose agents query your entity_type and confirm:

  • Agent output cites the catalog entities.
  • Grounding indicators in the report show claims are source-backed.

If the catalog is empty (you skipped make ingest), those agents still run but have less to ground against — see the seed-vs-ingest comparison in Catalog.