Creating and maintaining knowledge

Once the bundle exists and the agent is set up, the actual work amounts to two sentences in the chat: name the source, approve the plan. Everything else — writing files, updating the index, appending to the log, validating and committing — is handled by the agent.

Sources

Anything the agent has access to works as a source: collections in companyRAG, documents and transcripts from M365, attached files, GitHub repositories. For content from the web the WebFetch tool is added; in the example, naming a product page along with the instruction to turn it into documentation is enough. PDFs and other document types work the same way.

Note

WebFetch is an MCP server. Its individual functions carry the names of the underlying service (firecrawl_scrape, firecrawl_map, firecrawl_search, firecrawl_crawl, firecrawl_check_crawl_status, firecrawl_extract); for simply reading a page, firecrawl_scrape is sufficient.

The plan comes first

Before anything is written, the agent replies with a list of the intended concepts — file name, type and one line of description per entry:

This intermediate step is the right moment for objections: one concept too many, a wrongly chosen type, a file that belongs in a subfolder — here the correction costs one sentence, after writing it costs a clean-up run.

Write, validate, commit

After confirmation the agent works through the plan and makes the steps visible in the chat:

1. The concepts are written.
2. The index is updated.
3. The change log is appended to.
4. The bundle is validated.
5. The changes are written back.

For a bundle in pull request commit mode the run ends with a proposed change for review instead of a direct commit. That lets the knowledge base be treated exactly like source code: the agent proposes, a human approves. See OKF bundles.

The result

Every generated file has the same shape: a header block with type, tags, resource, title, description and timestamp, below it the content with headings, and a references section at the end.

The content stays short and points to the related concepts instead of repeating them — which keeps the files individually readable and individually updatable.

Automatic indexing

Because the bundle is connected as a collection in companyRAG, the new files appear under Files with the status Indexed without any further action:

As soon as a file is written it is indexed; when it changes, the index is updated. That closes the loop: the agent writes the knowledge, and the same knowledge is immediately findable through search — for users in CompanyGPT and for other agents through their tools.

Note

index.md, log.md and README.md are indexed as well. To match only the actual concepts in search, filter on the type and tags metadata.

What next?

• Files – check the state of individual files and re-index them.
• Using it in CompanyGPT – put the collection to work in chats and agents.