Knowledge

Knowledge collections store policy and product context that agents can reference to answer questions. The platform uses retrieval augmented generation (RAG) to find relevant chunks and ground responses.

Collections

A collection groups related documents.

{
  "name": "Product Documentation",
  "slug": "product-docs",
  "description": "Product manuals and specifications",
  "settings": {
    "relevance_threshold": 0.7,
    "max_chunks": 5,
    "mode": "reference"
  }
}

Collection Settings

Setting	Description
relevance_threshold	Minimum similarity score for a chunk (0 to 1)
max_chunks	Maximum chunks to include in a response
mode	must_cite, reference, or background

Documents

Supported formats: PDF, TXT, MD, DOCX, HTML.

Documents include metadata for filtering and auditing:

{
  "name": "Return Policy",
  "slug": "return-policy",
  "description": "Returns and refunds",
  "metadata": {"region": "US"}
}

Using Knowledge In Capabilities

Attach knowledge at the capability level:

knowledge:
  - product-docs
  - faq/return-policy

Or use the full form with settings:

knowledge:
  collections:
    - product-docs
  documents:
    - faq/return-policy
  settings:
    relevance_threshold: 0.75
    max_chunks: 4
    mode: reference

Agent Level Knowledge

You can also attach collections directly to an agent. Those collections are available across all capabilities for that agent.

Best Practices

• Use clear, well structured documents
• Keep policies and FAQs up to date
• Avoid duplicates and conflicting content
• Test in Playground to validate answers