Skip to main content
Knowledge stores project documentation, policies, and reference material as versioned sources. It is separate from Agent Memory: Knowledge is explicitly managed and source-grounded, while Memory captures durable facts and preferences learned by an agent. Each source has a trust level, immutable versions, parsed chunks, citations, and an agent allowlist. Retrieval checks project and agent access before a result can enter the runtime context.

​
Source types

Polpo Cloud currently accepts:
  • pasted UTF-8 text;
  • public HTTP or HTTPS URLs.
URL ingestion rejects local, private, link-local, metadata, credential-bearing, oversized, and unsafe redirect targets. File and Connection-backed sources are not yet available in the hosted ingestion path.

​
Ingestion lifecycle

When durable ingestion is enabled for a project, creating or reindexing a source queues bounded background work and returns the source in a pending state. The source progresses through pending, indexing, and indexed, or ends as failed with a sanitized diagnostic. Reindexing does not remove the current indexed version while the replacement is processing. Duplicate deliveries are idempotent, failed jobs can be retried, and deleting a source cancels its queued work. Polpo waits for the durable queue to accept each new job. If dispatch is unavailable, the request fails, the job becomes visibly failed, and its staged payload is retained for an explicit retry. A lost queue response cannot overwrite a job that a worker has already claimed or completed. The managed dashboard shows source status, the latest ingestion attempt, and a retry action. Durable ingestion availability is controlled by the project’s hosted rollout; when disabled, the existing bounded ingestion behavior remains unchanged.

​
API

Hosted projects with observable durable ingestion can also inspect GET /v1/brain/sources/{sourceId}/ingestion-jobs. In enforcing modes, retry a failed job with POST /v1/brain/ingestion-jobs/{jobId}/retry.

​
Runtime use

Grant a source only to the agents that need it. Authorized agents can retrieve relevant chunks through Knowledge tools or automatic runtime context when that policy is enabled. Retrieved text remains clearly separated from system instructions and preserves source, version, trust, and citation metadata.