How to use the LLM committee feature¶

The /committee command consults multiple AI models in parallel, optionally has them review each other's responses, then synthesizes their perspectives into a final answer.

When to use a committee¶

Committees are useful when:

You want multiple perspectives on a complex question
Different models might have different strengths (reasoning, creativity, domain knowledge)
You're making an important decision and want diverse viewpoints
You want to identify areas of agreement and disagreement among AIs

Prerequisites¶

You need API keys configured for the models you want to use:

Click the ⚙️ Settings button
Add API keys for providers (OpenAI, Anthropic, Google, Groq, GitHub)
Or use Ollama for local models (no API key needed)

Each model you select must have a valid API key for its provider.

Basic committee usage¶

Type /committee followed by your question:

/committee What are the main risks of using microservices architecture for a small startup?

Press Enter. A modal appears where you can:

Edit the question if needed
Select committee members (2-5 models)
Choose a chairman model (synthesizes the final answer)
Enable review phase (optional, slower but more thorough)

Click "Start Committee" to begin.

Committee configuration¶

Selecting committee members¶

The modal shows all available models as checkboxes. By default, Canvas Chat pre-selects:

Your currently selected model
Up to 2 other recently used models

Requirements:

Minimum: 2 models
Maximum: 5 models

Tips for selection:

Mix model families for diversity (e.g., GPT-4, Claude, Gemini)
Include both large and small models (small models run faster)
For technical questions, include models known for reasoning (Claude, o1)
For creative questions, include models known for creativity

Choosing a chairman¶

The chairman model receives all committee opinions and synthesizes a final answer. Choose a model that's good at:

Analyzing multiple perspectives
Identifying consensus and disagreement
Providing balanced summaries

Recommendation: Use a large, capable model (GPT-4o, Claude Sonnet 4) as chairman, even if smaller models serve on the committee.

Review phase (optional)¶

When enabled, the committee process includes an extra step:

All models provide their initial opinions (as usual)
Each model reviews and ranks the other opinions
The chairman sees both opinions and reviews when synthesizing

Enable review when:

The question is particularly complex
You want to see how models critique each other
Time is not a constraint (review adds 30-60 seconds)

Disable review when:

You want faster results
The question is straightforward
Cost is a concern (review doubles the API calls)

How the committee works¶

Phase 1: Gathering opinions¶

All selected models respond to your question in parallel. You'll see:

[OPINION] GPT-4o: streaming response...
[OPINION] Claude Sonnet 4: streaming response...
[OPINION] Gemini 1.5 Pro: streaming response...

Each model's opinion streams in real-time, appearing in its own OPINION node on the canvas.

Phase 2: Reviews (optional)¶

If enabled, each model reviews the others' opinions:

[REVIEW] GPT-4o reviewing other opinions...
[REVIEW] Claude Sonnet 4 reviewing other opinions...

Reviews include:

Strengths of each opinion
Weaknesses or gaps
Ranking from best to worst

Phase 3: Chairman synthesis¶

The chairman model receives all opinions (and reviews, if applicable) and synthesizes:

[SYNTHESIS] Claude Opus 4: synthesizing final answer...

The synthesis typically includes:

Areas of agreement among models
Points of disagreement and their significance
The chairman's assessment of the most accurate/helpful answer
A comprehensive final recommendation

Working with committee results¶

The canvas will show:

Your question node (HUMAN type)
Opinion nodes (one per committee member)
Review nodes (if review phase was enabled)
Synthesis node (the final answer)

All nodes are connected with edges showing the flow:

Question → Opinions (solid lines)
Opinions → Reviews (dashed lines)
Opinions/Reviews → Synthesis (solid lines)

Reading the synthesis¶

The synthesis node contains the chairman's analysis. It should answer:

Where did models agree?
Where did they disagree, and why?
What's the recommended answer based on all perspectives?

Follow-up questions¶

Reply to the synthesis node to ask follow-up questions:

Can you elaborate on the disagreement about database scaling?

The AI has access to all committee opinions and can reference specific model perspectives.

Branching from opinions¶

Select text from any opinion or review node and click 🌿 Branch to:

Ask a model to elaborate on a specific point
Challenge an assumption in an opinion
Explore a disagreement in depth

Context-aware committees¶

When you select text or nodes before running /committee, the AI refines your question based on that context.

Example: Building on research¶

Run /research microservices architecture trade-offs
Read the research report
Select the research node
Type /committee should we use this for our startup?
The AI refines to: "Should a small startup use microservices architecture for their initial product, considering the trade-offs discussed in the research?"

All committee members receive the research context, leading to more informed opinions.

Tips for effective committees¶

Ask clear, scoped questions¶

✅ Good:

/committee For a Python data pipeline processing 10M records/day, should we use Apache Spark or stick with pandas + multiprocessing?

❌ Too vague:

/committee which technology is better?

Use review for high-stakes decisions¶

Enable review when:

Making architectural decisions
Evaluating business strategy
Analyzing security/privacy concerns
Choosing between alternatives with significant trade-offs

Mix model strengths¶

Reasoning-heavy questions:

Include: Claude Sonnet 4, o1, Gemini 1.5 Pro
Chairman: Claude Opus 4

Creative questions:

Include: GPT-4o, Claude 3.5 Sonnet, Gemini
Chairman: GPT-4o

Coding questions:

Include: GPT-4o, Claude 3.5 Sonnet, Llama models
Chairman: Claude Sonnet 4

Manage costs¶

Committee runs can be expensive:

3 models without review: ~3-5 API calls
3 models with review: ~7-10 API calls
5 models with review: ~16-20 API calls

Cost-saving strategies:

Use smaller models for opinions (GPT-4o-mini, Haiku)
Reserve large models for chairman role
Disable review for exploratory questions
Use Ollama local models (free) as some committee members

Limits¶

Minimum 2 committee members, maximum 5
All models must have valid API keys (except Ollama)
Cannot stop committee once started (let it complete or refresh the page)
Large committees with review can take 2-3 minutes to complete

Troubleshooting¶

"At least 2 committee models required"¶

You must select at least 2 models
Click checkboxes to select more models

"API key missing for [provider]"¶

One of your selected models needs an API key
Go to Settings and add the missing key, or deselect that model

One opinion fails but others succeed¶

The committee continues with working models
The failed model's opinion won't appear
Synthesis will note which models contributed

Committee stuck on "Gathering opinions..."¶

Check browser console for errors (F12)
One model may be timing out (wait up to 60 seconds)
Refresh the page if it exceeds 2 minutes

Advanced: Combining with other features¶

Committee → Matrix¶

After getting diverse opinions, evaluate them systematically:

Run /committee on a question
Select all opinion nodes
Run /matrix compare these opinions against correctness, practicality, and clarity

Research → Committee → Decision¶

Full decision-making workflow:

/research to gather comprehensive information
/committee to get multiple AI perspectives on the research
Select synthesis + key opinions
Ask final clarifying questions
Make informed decision based on diverse viewpoints

Multiple committees¶

For complex decisions, run sequential committees:

First committee: "What are the main options for solving X?"
Second committee: "Given these options, what are the risks of option A?"
Third committee: "Should we pursue option A or B given these trade-offs?"

Each committee builds on the previous, creating a decision tree.