When a Knowledge Base Helps

• Technical support. "How do I reset my device?" The agent can pull the exact steps from your product manual.
• Policy and procedure questions. Warranty terms, return windows, service-level commitments, eligibility rules.
• Long-tail FAQs. The dozens of questions that do not make it into a top-ten list but still come up on real calls.
• Anything document-shaped. PDFs, help-center pages, internal wikis — content that already exists somewhere in writing.

If the answer is a single sentence, put it in the agent's regular company info. If the answer lives in a document, put the document in the Knowledge Base.

How It's Wired to the Agent

Once a Knowledge Base is attached to an agent and active, the agent gets a single tool: Search Knowledge Base. During a call, when the caller asks a question the agent cannot answer from its training data, it calls the search tool with the question, reads the retrieved passages, and gives a conversational answer based on what it found.

Under the hood, searches use a retrieval model to find the most relevant passages, then a language model summarizes them into a natural reply. The caller hears a conversational answer, not a raw document quote.

Adding the Tool to a Flow

Add the Search Knowledge Base tool to any state where the agent might need to answer a question from your documents. Usually this is the first "Greeting / Assisting" state. For a dedicated support agent, you may also want it on a specific "Technical Support" state that the flow transitions into when a caller's question signals they need deeper help.

1. Step 1. Go to the Flows tab and click on a state where you want the agent to use this knowledge (e.g., "Assisting" or "Supporting").
   
2. Step 2. Enable Knowledge Base in that state's properties panel. The agent will then reference your documents when responding in that state.
   

A typical multi-state support flow:

• State 1: Triage — light questions, transfer options, basic company info.
• State 2: Technical Support — Search Knowledge Base, plus an "Escalate to human" transition when the caller is not satisfied.

Content That Works Well

• Structured documents with clear headings. The retriever picks better passages when sections are labeled.
• Short paragraphs. A procedural document with one step per paragraph retrieves more precisely than a wall of text.
• Plain language. If the document is written for customers, the agent's answers will sound like customer support. If it is written for engineers, expect engineering-sounding answers.

Content That Works Poorly

• Scanned PDFs without text. If the PDF is an image, the retriever has nothing to read.
• Spreadsheets and pricing tables. Tabular data reads poorly aloud. For pricing, put the numbers in the agent's company info as text.
• Outdated documents. The agent will cheerfully quote an expired policy. Keep the Knowledge Base current — schedule a review cadence.

Tuning the Experience

• Pair the Knowledge Base search with a comfort message like "Let me look that up for you" so the caller knows something is happening while the search runs. See AI Agent | Comfort Messages for Tool Latency.
  
• Watch for repeat questions the agent is getting wrong. Usually the fix is a small edit to the source document, not a change to the agent.
• For business-critical answers (warranty terms, legal disclaimers), test them by call before trusting the agent with them live.

When to Escalate Instead

A Knowledge Base is not a substitute for a human when the caller needs judgement, empathy, or account-specific action. Build an explicit "Escalate to human" transition on your support state. The goal is to answer the easy 80% confidently and hand the hard 20% off gracefully, not to pretend the agent can do everything.