Skip to content

Knowledge

Knowledge holds the runbooks, notes and skills Polo reads when it works on a task. Open it from Apps → Knowledge. An agent that has Search workspace knowledge switched on searches it during a task and cites the document, version and lines it used.

Type Use it for
Runbook A procedure for a service. Tag the service and environment so alert investigations find it before broader searches.
Service notes What a service is, who owns it, how it behaves
Team notes Conventions and how your team works
Architecture How systems fit together
Incident notes Records of past incidents
Skill A procedure written as instructions and checks. An agent is given one exact version of a skill.
Learned fact A short statement an agent recorded while it worked, with its evidence
  1. Open Knowledge and press New document.
  2. Choose the Document type and enter a Title.
  3. Write or paste the Content. Use workspace text without credentials or secrets.
  4. If it applies to one, set the Environment and Service. Optionally add the Source HTTPS URL it came from.
  5. Press Save document.

Documents are versioned. Edit current version saves a new version; earlier versions stay readable, so a citation in an old task still points to the text the agent read. Archive document removes it from active knowledge and search.

Administrators and operators can edit documents. Only administrators can edit a skill.

A skill is a knowledge document that describes a procedure. To use one:

  1. Create a document of type Skill.
  2. Edit the agent on the Agents page and switch on the skill under its knowledge tools. It appears as the skill’s title with Skill version and a number.

The agent receives that exact version. After you save a new version of a skill, open the agent and check which version it holds. The Terraform skill is a different thing: it is set up on the agent itself. See Agents.

An agent can record what it learns for later tasks. To allow it, edit the agent and tick Record facts it learns for later tasks.

  • A fact is a title and a short statement of what is true and how the agent knows.
  • Polo attaches where it came from: the task, and up to five tool observations the agent had made by then. Only a one-line summary of each observation is kept, never its content.
  • A fact recorded before the agent had observed anything is marked as having no evidence.
  • An agent can record at most three facts per run. A fact whose title is already known is not recorded twice, and credentials are refused.
  • No approval is needed to record a fact, because it changes nothing outside Polo.

Later tasks find learned facts when they search knowledge. Polo tells the agent to treat a fact as a lead to check again with a current read, not as proof.

Choose the document type Learned fact in the library to review them. Each fact shows where it came from, its evidence and whether it was confirmed. The audit flags a fact when it:

Flag Meaning
No evidence It was recorded before the agent had observed anything
Not confirmed for 90 days Nobody has confirmed it recently
Duplicate It says the same as an earlier fact
Status note It reads like something temporary, not something that stays true

For each flagged fact, press Still true to confirm it, which restarts its 90 days, or archive it. Nothing is removed automatically.

The Search tab shows what Polo would find for a question.

  1. Enter a query, such as “How do we roll back checkout?”.
  2. Optionally filter by document type, environment and service.
  3. Press Search. Each result shows the matching section and opens the original version.

Search indexing on the Documents tab shows whether semantic search is working. Semantic search needs an embedding service configured by whoever runs Polo. When it is not configured or not available, the page says so. If a document fails to index, press Retry indexing.

  • An investigation started by an alert or a monitor reads the runbooks tagged with its service and environment first, and the task names the runbooks that apply.
  • An agent searches within the scope of its task.
  • Text in a document is information for the agent. It cannot give the agent tools or permissions it does not have.
  • Agents cannot edit or create documents. The only thing an agent writes to knowledge is a learned fact.
  • Each line of a skill’s instructions must stay under 1,500 bytes.
  • Do not put credentials in knowledge. Use the vault for keys and tokens.