Open WebUI Knowledge Search Tool

On-demand knowledge base search tool for Open WebUI. The LLM intelligently decides when to access organizational knowledge instead of constantly injecting context into every message.

Features

• LLM decides when to access knowledge bases
• Token efficient - only retrieves when necessary
• Full citation support with inline [1], [2] references
• Works with existing Open WebUI knowledge bases
• No external dependencies required
• Supports multiple knowledge bases

Installation

1. Copy the contents of knowledge_search_tool.py
2. In Open WebUI, go to Workspace → Tools → Create Tool
3. Paste the code and save
4. Configure default_knowledge_bases with your knowledge base IDs (comma-separated)
5. Go to Workspace → Models → Edit your model → Enable the tool

Finding Knowledge Base IDs

Go to Workspace → Knowledge, click on a knowledge base, and check the URL:

http://localhost:8080/workspace/knowledge/[KNOWLEDGE_BASE_ID]

Alternatively, enable the tool and ask: "List available knowledge bases"

Usage

When users ask questions requiring organizational knowledge, the LLM automatically calls search_knowledge():

User: What's our password policy?
LLM: According to the security documentation [1], passwords must be at least
     12 characters and include uppercase, lowercase, numbers, and special
     characters [2].

For general questions, the tool is not invoked:

User: What's 2+2?
LLM: 4

Configuration

Setting	Description	Default
default_knowledge_bases	Comma-separated knowledge base IDs	""
top_k	Number of document chunks to retrieve	5
relevance_threshold	Minimum similarity score (0.0-1.0)	0.0

How It Works

1. The LLM sees the tool as an available function
2. When a question requires knowledge, the LLM calls search_knowledge(query)
3. The tool queries knowledge bases using Open WebUI's internal query_collection()
4. Results are formatted with XML <source> tags (same as native RAG)
5. The LLM generates a response with inline citations [1], [2]
6. Citations appear in the source panel

Comparison: Tool vs. Native Knowledge

Feature	Native Knowledge	This Tool
Retrieval	Every message	Only when needed
Token Usage	High	Efficient
LLM Decision	No	Yes
Citations	Yes	Yes
Multiple KBs	One per model	Any combination

Troubleshooting

"Error: No knowledge bases specified" Configure default_knowledge_bases in tool settings or have the LLM pass the knowledge_base_ids parameter.

"No relevant information found" The query didn't match knowledge base content. Try different search terms.

Tool not being called

• Verify tool is enabled in model settings
• Ensure your LLM supports function calling
• Try explicit requests: "Search our documentation for..."

Technical Details

The tool uses Open WebUI's internal modules:

• open_webui.retrieval.utils.query_collection() - Direct vector database querying
• open_webui.models.knowledge.Knowledges - Knowledge base access
• app.state.EMBEDDING_FUNCTION() - Same embedding function as native RAG

License

MIT License