kb-guide git:20260315.e56102cA

Immutable. This exact content never changes and is served at /api/v1/blob/44f6cb391f72d1b3.

---
name: kb-guide
description: Guide for using MCP-connected knowledge base tools — search, read, create, and update knowledge entries
type: integration
---

# Knowledge Base Integration Guide

When MCP knowledge base servers are configured, you can use their tools to search, read, create, and update knowledge entries.

## Common MCP Knowledge Base Tools

MCP KB servers typically expose tools with these patterns:

### Search / Query
- `search_documents(query, limit?)` — full-text search across the knowledge base
- `search_notes(query)` — search personal notes or wiki entries
- `query_knowledge(query, filters?)` — filtered search with metadata constraints

### Read
- `get_document(id)` — retrieve a document by ID
- `read_page(path)` — read a wiki page by path
- `list_documents(folder?, tag?)` — list documents in a folder or by tag

### Create / Update
- `create_document(title, content, metadata?)` — create a new entry
- `update_document(id, content)` — update an existing document
- `create_note(title, body)` — create a quick note

## When to Use Knowledge Base

- **Before answering domain questions**: Search the KB first — it may contain authoritative answers
- **When the user asks "do we have docs on…"**: Search the KB and return results
- **After resolving an issue**: Suggest creating a KB entry to capture the solution
- **For onboarding/FAQ questions**: Check if there's an existing KB article

## Usage Pattern

1. Identify the user's intent (search, read, write)
2. Call the appropriate MCP tool by name
3. If a tool is not available, inform the user that no KB server is configured
4. Present results clearly — include titles, snippets, and links when available

## Notes

- MCP server configuration is defined in `golem.yaml` under the `mcp` key
- Available MCP tools depend on which servers are configured — use tool discovery to check
- Always prefer searching the KB before generating answers from your own training data