Knowledge Base

The Knowledge Base (KB) enables agents to answer based on documents, web pages and manual text.

Concept

• An organization can have multiple KBs.
• Each KB has its own instruction that guides retrieval.
• Each KB contains documents.
• An agent can have multiple KBs assigned.

Source types

Site crawl is an ingestion method: it walks a domain and creates url documents, one per page.

Document lifecycle

1. pending: waiting for processing.
2. uploading: upload in progress (PDF).
3. processing: extraction and chunking.
4. indexing_pending: ready, waiting for indexing.
5. ready: document available for retrieval.
6. failed: processing error.

Indexing

After processing, chunks are embedded and indexed. The state can be:

• pending
• indexing_pending
• indexed
• failed
• skipped

Retrieval

At runtime, the agent can use retrieval modes:

• retrieval: relevant chunks.
• full_doc_fallback: the full document when the chunk is not sufficient.
• instructions_only: only the KB instruction is used.
• empty: no documents included.

Assign to an agent

1. Open the agent.
2. Go to the Knowledge tab.
3. Choose the KBs to enable.
4. Save and test with the KB diagnostic tool.

Citations

When the KB is used, the runtime can return citations indicating which document and chunk originated the response.

Next

After configuring the Knowledge Base, validate the agent in the Test Center.