Best Practices for Using Knowledge Agents

👥
Access Required
You must be an Admin, Creator, or relevant custom role to create a new Knowledge Agent.
You must be an Admin, Knowledge Agent Owner, or relevant custom role to manage Knowledge Agents.

Key Recommendations

Write clear, descriptive summaries that explain each Knowledge Agent's purpose and intended use cases
Enable Research Mode for Knowledge Agents that need to provide comprehensive, well-researched answers
Optimize for both search and chat by configuring agents to excel in conversational interactions
Customize each Knowledge Agent to reflect a unique task, audience, or business goal
Use clear and focused prompts with limited scope to increase answer precision
Review agent behavior using the AI Agent Center and refine prompts over time
Use formatted lists or steps to improve clarity
Encourage prompt examples that align with your team's tone and use cases

Watch the Overview: 3 Best Practice Tips for Using Knowledge Agents

For a deeper dive, check out the recording of a past Knowledge Agent Workshop

Writing Effective Knowledge Agent Descriptions

Clear descriptions help users choose the right Knowledge Agent for their needs, especially when multiple agents are available.

Description Best Practices

Be specific: Explain what the agent specializes in and what questions it answers best
Mention key features: Note if Research Mode is enabled or if it's optimized for chat conversations
Set tone expectations: Indicate if the agent uses formal, friendly, or technical language
Keep it current: Update descriptions when sources or capabilities change

💡
Example Description
"Sales Enablement Assistant - Specializes in battlecards, competitive intelligence, and call preparation. Research Mode enabled for comprehensive market analysis. Optimized for conversational chat interactions."

Optimizing Knowledge Agents for Chat Conversations

Knowledge Agents excel in chat by enabling natural conversations that build on previous questions.

Chat Best Practices

Configure conversational tone: Set prompts that encourage natural back-and-forth interactions
Encourage follow-up questions: Train users to explore topics deeper within the same chat
Enable voice input: Users can ask complex questions hands-free during conversations
Use the AI Agent Center: Monitor chat threads to understand how users interact with agents

💡
Tip
Train users to use search for quick answers and chat for deeper exploration and research.

Maximizing Research Mode

Research Mode enables Knowledge Agents to perform deep research combining internal knowledge with web search for comprehensive answers.

Research Best Practices

Enable strategically: Turn on Research Mode for agents handling complex queries or needing current information
Guide initial queries: Encourage detailed, specific questions when starting research for better results
Allow plan review: Users should review the generated research plan before starting
Use for complex analysis: Examples include competitor research, policy analysis, or market studies

💡
Pro tip
Break complex research topics into smaller, focused questions to improve answer accuracy.

Best Practices for Web Search Configuration

Define your search scope strategically:

Use domain restrictions for industry-specific agents (e.g., only search regulatory sites for compliance agents)
Enable full web search for research and competitive intelligence agents
Consider your team's needs for real-time vs. curated information

Choose authoritative domains:

Prioritize official documentation sites
Include trusted industry publications
Add vendor help centers for technical support agents

Monitor and refine:

Review web search citations in the AI Agent Center
Adjust domain lists based on answer quality
Remove domains that consistently provide outdated or irrelevant content

💡
Tip
Start with a narrow domain list and expand based on user feedback and answer quality metrics.

Extending Knowledge Agent capabilities with MCP

Guru supports MCP (Model Context Protocol) in two ways:

Guru's MCP Server (External tools → Guru)

Guru's MCP Server lets external AI tools like Claude Desktop, Cursor, or ChatGPT connect to your Knowledge Agents. This is ideal for:

Developers accessing documentation directly in their IDE
Content creators using AI tools with your company knowledge
Automation workflows that need to query Guru

Learn more about Guru's MCP Server

Third-party MCP Integration (Guru → External tools)

Third-party MCP integration lets your Knowledge Agents connect to external tools like Asana, Slack, and Salesforce. This enables agents to:

Retrieve real-time data from connected tools
Execute actions like creating tasks or updating records
Combine internal and external information in answers

When to use third-party MCP:

Your team needs to access data from multiple external systems
Agents should take actions in external tools (create tasks, send messages)
You want to reduce context-switching between Guru and other platforms

Setup overview:

Workspace Admins enable MCP servers at the workspace level
Agent Owners configure which MCPs each agent uses
Users authenticate and control permissions for their connections

Learn more about third-party MCP integration

Creating Effective Custom Prompts for Your Knowledge Agent

Generate a custom prompt in less than 1 minute with our Prompt Architect!

Custom prompts define the behavior, tone, and constraints of your Knowledge Agent. When structured thoughtfully, they help ensure responses are clear, consistent, and tailored to your audience—while remaining grounded in your trusted content.

Prompts Define How, Sources Define What

It's important to understand the distinction between what information the AI can use and how it presents that information:

Sources control the content: Your Knowledge Agent only uses information from the sources you've enabled—such as Guru Cards, documents, or Help Center articles. These sources determine what facts are available to answer a question.
Prompts control the presentation: Prompts influence the style, tone, and structure of the response. They don't prioritize which sources are used or provide search direction, they shape how the AI delivers the information it finds in the sources.

Follow Guru's Custom Prompt Template

Include the following sections in every custom prompt. AI generally reads text best in Markdown format, so try to structure it in that way. (If you use the above Prompt Architect, it will generate the prompt in markdown!)

Role

You are a helpful assistant who works for [company name].

Add any relevant details about team or product context.

Task

Your primary goal is to [describe what the agent is helping with].

Context

Define who is asking the questions and what they're typically about. Specify if responses are for internal or external use, and include any additional constraints.

Format

Explain how the response should be structured:

Provide answers in [bullet points / steps / full paragraphs]. 
Responses should be [concise / under 200 words, etc.].

Examples

Add representative Q&A pairs that model the structure, tone, and level of detail you expect in responses.

Tone

Use a [friendly / professional / helpful] tone of voice.
Do not mention [e.g., AI, GPT, internal tools].

Mandatory Instructions

Always include the following:

Responses must be based solely on the provided documents
Do not hallucinate
Do not repeat text

Knowledge Agent Templates

Knowledge Agents can be adapted to meet specific needs across your org:

Enable Conversational AI: Natural chat interactions that build on previous questions
Automate Responses: Instant answers to common questions across Slack, web app, or browser extension
Conduct Deep Research: Use Research Mode for comprehensive analysis combining internal and external sources
Customize Search: Configure agents to search specific collections, boards, or tags
Enhance Accuracy: Use the AI Agent Center to track and improve response quality

Watch Setup in Action: Knowledge Agent for Customer Responses

Watch Setup in Action: Knowledge Agent for Sales

Watch Setup in Action: Knowledge Agent for RFPs.