KB Search Node

The KB Search node queries a Knowledge Base (KB) to find relevant document chunks, then uses an LLM to generate a grounded response. It combines vector similarity search with AI generation -- a pattern known as retrieval-augmented generation (RAG).

When to Use

• You want the bot to answer questions from your uploaded documents (FAQs, policies, manuals, product catalogs).
• You need responses that are grounded in specific content rather than general LLM knowledge.
• You want to cite sources so the user knows where the information came from.

Configuration

Property	Description	Default
Knowledge Base	The KB to search. Selected from a dropdown of your tenant's Knowledge Bases.	None (required)
Integration	The AI model used to generate the response from retrieved chunks.	Tenant default AI model
Query Template	The search query. Defaults to the user's last message. Supports {{variable}} interpolation.	{{last_message}}
Top K	The maximum number of document chunks to retrieve.	5
Min Score	Minimum similarity score (0.0 to 1.0). Chunks below this threshold are excluded.	0.5
Max Context Tokens	Maximum number of tokens from retrieved chunks to include in the LLM context.	4000
Max Tokens	Maximum number of tokens for the LLM's generated response.	2048
Include Sources	Whether to include source document references in the output.	true

imageKB Search node config panel showing Knowledge Base dropdown, integration selector, query template field, Top K slider, Min Score threshold, and Include Sources toggle

KB Search configuration panel

AI Prompt Management

Like the LLM Response node, you can select a Managed Prompt to control how the LLM generates answers from the retrieved context. The prompt template typically instructs the LLM to answer based only on the provided documents and to cite its sources.

Dynamic KB Lookup

Enable Dynamic KB to select the Knowledge Base at runtime based on the page_label session variable. This is useful when:

• You embed the chat widget on multiple pages, each with its own KB.
• The page label (set via the widget's metadata option) maps to a KB name.

When dynamic lookup is enabled, the node:

1. Reads the page_label variable from the session context.
2. Searches for a KB whose name matches the label (case-insensitive).
3. Falls back to the statically configured KB if no match is found.

Setting the page label from the widget

Pass the page_label in the widget's metadata init option:

<!-- Widget on the pricing page -->
<link rel="stylesheet" href="https://omnibots-widget.web.app/widget.css">
<script src="https://omnibots-widget.web.app/widget.iife.js"></script>
<script>
  window.omnibot.init({
    deploymentKey: 'your-deployment-key',
    metadata: {
      page_label: 'Pricing'
    }
  });
</script>

For single-page apps, update the label on navigation:

// SPA route change
window.omnibot.setContext({ page_label: 'Support' });

TIP

Dynamic KB lookup lets you build a single flow that serves different content depending on which page the user is chatting from. Create one KB per product or topic and name it to match the page label. See Widget Page Context for the full context API.

Agentic Mode

When Agentic Mode is enabled, the KB Search node can call tools during response generation, just like the LLM Response node. The LLM can combine KB results with tool calls (e.g., look up an order, then answer using KB context).

Property	Description	Default
Agentic Mode	Enable tool calling during KB response generation.	true
Enabled Tools	Which tools the LLM can call.	None
Max Iterations	Maximum tool-call loops.	5

Output

The node stores its output in a variable (default: kb_results) containing:

• The generated text response (grounded in KB content).
• An array of source documents with metadata (if Include Sources is enabled), including document name, page, and relevance score.

Use the output in a Message node to send the response to the user:

{{kb_results}}

imageChat widget showing a KB Search response with an AI-generated answer followed by source citation links showing document names and page numbers

Search results with source citations in the chat widget

WARNING

If the KB is empty or no chunks meet the minimum score threshold, the LLM will generate a response without document context. Consider adding a Condition node after the KB Search to check whether relevant results were found.

Handles

Handle	Direction	Description
Input	In	Receives execution from the previous node
Output	Out	Continues to the next node after the search and generation completes