Skip to content

Admin Dashboard: Knowledge

Knowledge bases provide vector retrieval over narrative documents. Use them for policies, manuals, FAQs, and other prose; use Knowledge Graphs when agents must filter typed records or resolve exact IDs.

Create an Embedding model under Models. Chat models and embedding models are distinct kinds, and a model’s kind cannot be changed after creation.

  1. Open Knowledge and choose Create Knowledge Base.
  2. Enter a name and description, then select an embedding model.
  3. Open the base and drop TXT, Markdown, CSV, PDF, or DOCX files up to 50 MB into Files.
  4. Wait for each file to move from indexing to ready.
  5. Link one or more agents under Linked Agents.

The expected result is a base whose chunks are available to linked agents through the Knowledge capability.

There is no separate re-index command. To update a document, delete its existing file entry and then upload the new version. Uploading a file with the same name does not replace the earlier entry, so leaving both versions can give the agent conflicting passages. File status, chunk count, and any indexing error appear in the detail panel.

SyntheticBrew retains the searchable passages and file name, not a downloadable copy of the uploaded file. Keep the source document in your own document or source-control system.

Linking a base through the coding-agent MCP tools also ensures the agent has the Knowledge capability. In Admin, verify both the agent link and capability when retrieval is missing.

  • No embedding models: add a model with Kind set to Embedding.
  • File remains indexing: verify that the embedding endpoint is available. If the status changes to error, open the file details for the error message.
  • Agent does not retrieve: confirm the base is linked, the file is ready, and the Knowledge capability is enabled.
  • Exact filters are unreliable: move the structured records to a Knowledge Graph.