Model Reasoning: Showing the Work

Ever wonder how an AI gets to its answer? With reasoning-capable models, you can ask them to "show their work." It’s like getting a window into the model's thought process as it breaks down a problem, considers options, and formulates a plan.

This is incredibly powerful for:

• Debugging: See why the model gave a weird answer.
• Transparency: Understand the logic behind a recommendation.
• Building Agents: Use the model's plan to orchestrate complex tool calls.

AIKit normalizes the output from different providers, giving you a consistent way to access this "behind-the-scenes" thinking.

Getting the Full Thought Process

The easiest way to see a model's reasoning is to let it finish completely and then inspect the reasoning property on the result.

This example uses Anthropic's Claude 3.5 Sonnet, which has explicit support for this feature. We just have to enable it with the thinking parameter.

import { createProvider, userText, collectStream } from '@chinmaymk/aikit';

const provider = createProvider('anthropic', {
  apiKey: process.env.ANTHROPIC_API_KEY!,
});

const messages = [userText('Solve this equation: 2 * (x + 3) = 14')];

const stream = provider(messages, {
  model: 'claude-3-5-sonnet-20240620',
  // Just enable thinking to get the goods
  thinking: true,
});

const { text, reasoning } = await collectStream(stream);

console.log('Final Answer:');
console.log(text);
// "The value of x is 4."

console.log('\nHow it got there (Reasoning):');
console.log(reasoning);
/*
  "The user wants me to solve a linear equation.
  1.  First, I need to distribute the 2 on the left side: 2*x + 2*3 = 14, which simplifies to 2x + 6 = 14.
  2.  Next, I'll isolate the term with x by subtracting 6 from both sides: 2x = 14 - 6, which is 2x = 8.
  3.  Finally, I'll solve for x by dividing both sides by 2: x = 8 / 2, which gives x = 4.
  I will now state the final answer."
*/

For other providers, the process is similar:

• OpenAI: Use an o-series model (e.g., 'o1-mini') and set reasoning: { effort: 'medium' }.
• Google: Use a recent Gemini model (e.g., 'gemini-2.5-pro'). Support is automatic but can be experimental.

Watching the Thoughts Form (Streaming)

For a real-time view into the model's head, you can stream its reasoning. Use the processStream helper and provide an onReasoning handler to capture the thought process as it unfolds.

import { createProvider, userText, processStream } from '@chinmaymk/aikit';

const provider = createProvider('anthropic', {
  apiKey: process.env.ANTHROPIC_API_KEY!,
});

const messages = [userText('Solve this equation: 2 * (x + 3) = 14')];

const responseStream = provider(messages, {
  model: 'claude-3-5-sonnet-20240620',
  thinking: true,
});

await processStream(responseStream, {
  onReasoning: text => {
    // In a UI, you could render this to a "thinking..." box
    process.stdout.write(text);
  },
  onText: text => {
    // The final answer still comes through onText
  },
});

The Golden Rules of Reasoning

• It Costs Extra: "Showing your work" takes more effort from the model, which means it uses more tokens and costs more. Use it when you actually need the process, not just the answer.
• Check for Support: Not all models support reasoning. It's generally available on the latest flagship models from OpenAI (o-series), Anthropic (Claude 3+), and Google (Gemini 1.5+).
• It's Not Always Perfect: Sometimes the reasoning is a better answer than the final content! It's a fascinating, powerful, and occasionally quirky feature. Experiment with it.

Now go see what your AI is really thinking! 🚀