AI bot - Agents

[Discourse]

This guide explains how to set up and manage Agents for Discourse AI - AI bot.

Required user level: Administrator

The Discourse AI bot includes a flexible system for managing Agents. This guide covers:

• Interaction methods for AI Agents
• Configuration options
• Advanced settings
• Supported tools
• Customization possibilities

Learn about the AI bot and how to enable it in this guide.

What are agents?

Agents are a powerful feature that allows you to customize the behavior of the AI engine in your Discourse forum. They act as a ‘system message’ that guides the AI’s responses and interactions, helping to create a more personalized and engaging user experience.

Why use agents?

With agents, you can tailor the AI’s behavior to better fit the context and tone of your forum. Whether you want the AI to be more formal for a professional setting, more casual for a community forum, or even embody a specific character for a role-playing game, agents give you the flexibility to do so.

Group-specific access to agents

Moreover, you can set it up so that certain user groups have access to specific agents. This means you can have different AI behaviors for different sections of your forum, further enhancing the diversity and richness of your community’s interactions.

Interacting with Agents

Agents can be interacted with in multiple ways:

1. Private message (PM)
2. Chat direct message (DM)
3. Chat channel mentions
4. Mention in a public topic via @example_bot_username

Administrators can control which interaction methods are allowed for each agent via four separate toggles:

• Allow personal messages: Allow users to start PMs with this agent (enabled by default)
• Allow topic mentions: Allow @mention in topics to trigger a response
• Allow chat direct messages: Allow chat DMs with this agent (requires chat plugin)
• Allow chat channel mentions: Allow @mention in chat channels (requires chat plugin)

Configuration

To configure Agents, navigate to discourse.example.com/admin/plugins/discourse-ai/ai-agents.

Key configuration options include:

• System prompt: Steers the conversation
• Vision support: Enable for Vision powered LLMs, i.e- GPT-4 Turbo, Claude 3
• Default model: Required for mentionable agents and agents with force default LLM enabled
• Allowed groups: Define permissions using Discourse group security
• Uploads: Index text files for additional knowledge (using RAG - Retrieval-Augmented Generation)
• Enabled tools: Control access to specific tools
• Execution mode: Choose between “default” and “agentic” mode (agentic mode enables multi-turn tool use with configurable token limits and compression)

Here’s a look at a complete AI agent configuration screen:

Advanced configuration options

The agent editor offers several advanced options:

• Temperature and top_p (nucleus sampling): Control creativity vs. predictability
• Max context posts: Limit the number of posts supplied to the LLM for history - fewer posts will reduce costs, but cause the AI to “forget” about older content (not available in agentic mode)
• Execution mode: “Default” for single-pass responses, “Agentic” for multi-step tool use with:
  • Max turn tokens: Token budget for agentic execution
  • Compression threshold: Percentage threshold for context compression (20-99%)
• Show thinking: Display the AI’s reasoning process to users (enabled by default)
• Require approval: Require human approval before certain actions are executed
• Force default LLM: Prevent users from switching LLMs for this agent
• Response format: Configure structured output format
• Examples: Few-shot examples to guide the AI’s behavior (pairs of user/assistant messages)
• Upload options:
  • Chunk tokens: Control document splitting
  • Chunk overlap tokens: Set overlap between chunks
  • Search conversation chunks: Number of document fragments added to user queries
  • Question consolidator LLM: Choose the LLM for determining user questions
  • RAG LLM model: Choose a specific LLM for RAG processing

Prerequisites for upload support

To use upload support:

1. Enable the ai_embeddings_enabled setting
2. Configure an embedding model via ai_embeddings_selected_model

Discourse AI supports various embedding models. Hosted customers get free access to the bge-large-en model.

Self-hosters or people wanting more choice can self host an embedding model or use models by Open AI, Google (Gemini), and more.

Text splitting in Discourse AI

Discourse uses a Recursive Character Text Splitter to keep paragraphs, lines, and words together when splitting text. You can control text splitting using the [[metadata YOUR METADATA HERE]] separator:

[[metadata about cats]]
a long story about cats
[[metadata about dogs]]
a long story about dogs

This method prevents “chunk contamination” by ensuring only relevant data is included in each chunk.

Debugging AI agents

To debug your agent:

1. Add users to the group specified in the ai_bot_debugging_allowed_groups setting
2. Users in this group will see an AI debugging button at the bottom of AI posts
3. Click the button to access the debugging UI modal

Remember: If you provide vague or unhelpful information to an LLM, it will have great difficulty converting it into useful information.

Supported tools

Discourse AI supports various tools, including:

Always available

• Categories
• Time
• Search
• Read
• Database Schema
• Search Settings
• Setting Context
• Random Picker
• Web Browser
• Search Meta Discourse
• JavaScript Evaluator
• Researcher
• Flag Post
• Close Topic
• Unlist Topic
• Lock Post
• Delete Topic
• Edit Post
• Edit Category
• Set Topic Timer
• Set Slow Mode
• Move Posts
• Grant Badge

GitHub integration

• GitHub File Content
• GitHub Pull Request Diff
• GitHub Search Files

These three tools are always available. Additionally:

• GitHub Search Code requires the ai_bot_github_access_token setting to be filled

Conditional tools

• Tags and Edit Tags: Available when tagging is enabled on your site
• Image generation (Image, Create Image, Edit Image): Requires a custom image generation tool to be configured via the admin UI
• Google search: Requires the ai_google_custom_search_api_key and ai_google_custom_search_cx settings to be filled
• Artifacts (Create Artifact, Update Artifact, Read Artifact): Available when ai_artifact_security is set to lax, hybrid, or strict
• Assign: Available when the Assign plugin is installed
• Mark as Solved: Available when the Solved plugin is installed

Some tools require additional configuration or API keys.

Custom tool support

You can create your own tools using custom tool support.

Customizing system agents

1. Create a new agent
2. Copy and paste the system prompt from the agent you wish to customize
3. Modify as needed
4. Optionally disable the original system agent to avoid duplication

Customizing search

To customize search for a new agent:

1. Create a new agent
2. Add the “search” tool
3. Configure search options, including:
   • Base search query
   • Maximum number of search results
   • Include private topics option
     
     

This allows you to delegate search permissions, increase the number of results, or restrict search to specific sections of your site.

Last edited by @dax 2026-03-13T17:21:31Z

Check document

Perform check on document:

[MachineScholar]

@Discourse

Read

What exactly does the Read tool do?

Also, mentions on posts are currently broken per this bug report.

[sam (Sam Saffron)]

[Untitled AI bot PM]

In a nutshell it allows the LLM the ability to read topics and posts on your forum.

[Jagster (Jakke Flemming)]

Basically it reads a topic/post when url, name or ID is given and otherwise, or still, it uses embeds and what, when, why and how depends model itself, prompting etc?

[sam (Sam Saffron)]

Oh sorry was replying to a different post

It uses topic ID or topic ID + Post Number to find the raw markdown for a topic/post with some guards for max length.

Does not depend on embedding.

[Jagster (Jakke Flemming)]

A bit off-topic, but because I don’t have anything to ask and you guys will clean these topics every now and then, I don’t care.

Automatic titles in PMs with AI are now in the right language. Not a biggie, but its psychological value is big. Thanks!

[jrgong]

is there any way to add the Bot users to a group? I basically want to add an avatar flair to those users, but when adding them to a group, they don’t show up

[Jagster (Jakke Flemming)]

They don’t show up, but are still there. You can be sure by trying add an AI person second time. Or you see needed group on settings of those pseudo users.

[MachineScholar]

Which part of the Discourse AI repo should I check out if I want to precisely know how the AI bot’s tools work?

[sam (Sam Saffron)]

Tool, Playground, Bot are all good areas to start with.

See this for example: Understanding Tools in Discourse AI System - AI Conversation - Discourse Meta

[MachineScholar]

Precisely the answer I was looking for — thank you kind sir! For anyone it may pertain to: I recommend reading through the code in this repo to be able better steer your prompts for your persona’s tools.

[jrgong]

I am currently trying to fetch all posts from a topic via the “Read” tool, but it only provides a maximum of 100 posts to the LLM model. I already set “Max posts” to 999 in persona settings, but the issue persists.

Could there be any server-side API request limitation causing the issue?

[jrgong]

@sam any idea on this one what could be the cause for the 100 post limit that I am encountering?

[MachineScholar]

I haven’t checked the codebase, but I suspect it may be a hard limit set in the code itself. May I ask what it is that you’re trying to achieve with this? That is a key factor here.

And perhaps it may be best to use the Forum Researcher persona?

[jrgong]

We have “journal/log” category which often contain several hundred posts where the topic owner reports about his project progress over a span of 4-5 months.

My goal is to fetch all posts within that topic, filter them by author = topic owner and create a summary.

We considered using the researcher persona, but cuttently it doesn’t support vision. We also want to process users’ images.

[sam (Sam Saffron)]

Oh you got to use forum researcher for this, it’s an exact use case for the persona

[jrgong]

@sam

Currently, there is no option to send images to the research LLM. This will be considered in future versions.

@sam I would love to, but currently it doesn’t support images. Is that something on the roadmap or not a priority at the moment?

[sam (Sam Saffron)]

Sure I can add that , not too big of a change