AI Suggestions

Agent Assist generates suggestions in real time as a conversation unfolds. Each suggestion is a self-contained card with an AI-generated answer, source citations, relevance scores, and follow-up questions.

How the Suggestion Flow Works

The end-to-end flow from customer utterance to suggestion card follows these steps:

1. Customer speaks or types. The transcript is captured in real time by the conversation platform.
2. The classifier evaluates the utterance. A lightweight classifier determines whether the message warrants a knowledge base search. Noise messages (greetings, filler) are filtered out.
3. The knowledge base is searched. The RAG pipeline finds matching document chunks using vector similarity search.
4. Sources appear first. Source citations are delivered to your panel as soon as relevant documents are found.
5. The answer streams in. Using the matched chunks as context, the LLM generates a grounded answer that appears progressively in your panel -- you can start reading before it finishes.

This entire process typically completes in two to five seconds, depending on the complexity of the query and the size of the knowledge base.

What a Suggestion Contains

Each suggestion card includes four elements:

Element	Description
Answer	A markdown-formatted response generated by the LLM. May include paragraphs, bullet lists, numbered steps, and bold/italic emphasis.
Sources	One or more source citations linking back to the original documents. Each source shows the document title, page number, a content snippet, and a relevance score.
Relevance Scores	A percentage indicating how closely each source matched the customer's query. Higher scores mean a stronger match.
Quick Replies / Follow-ups	Quick reply chips for chat/messaging conversations, or follow-up question chips for voice calls.

TIP

Suggestions are grounded in your organization's knowledge base. If a question falls outside the indexed content, the system will indicate that no relevant sources were found rather than fabricating an answer.

Sub-Pages

Page	What You Will Learn
Suggestion Cards	Card anatomy, streaming behavior, expand/collapse, and timing indicators.
Source Citations	How to read source cards, relevance scores, and the difference between omni and CCAI sources.
Knowledge Search	How to manually search the knowledge base using the search bar.
Quick Replies	How to use pre-generated reply chips and follow-up questions.