Reasoning Models
The Tekimax SDK provides first-class support for Reasoning Models (also known as "Thinking" models), such as DeepSeek R1 and newer OpenAI o-series models.
These models expose their internal "chain of thought" process, allowing you to display the reasoning steps to the user before the final answer is generated.
Enabling Thinking
To enable reasoning, pass think: true in your chat options.
const response = await client.text.chat.completions.create({
model: 'deepseek-r1',
messages: [{ role: 'user', content: 'Solve this logic puzzle...' }],
think: true // Tells the provider to return reasoning data
});Accessing Reasoning Data
The reasoning content is available in the thinking field of the message.
const thinking = response.message.thinking;
const content = response.message.content;
console.log('--- Reasoning ---');
console.log(thinking);
console.log('--- Answer ---');
console.log(content);Streaming Support
Thinking is also supported in streaming mode. Use createStream — the thinking delta is emitted separately from the content delta so you can render them in different UI areas.
// createStream returns an AsyncGenerator — no await.
const stream = client.text.chat.completions.createStream({
model: 'deepseek-r1',
messages: [{ role: 'user', content: 'Solve: 2x + 3 = 11' }],
think: true,
});
for await (const chunk of stream) {
// Thinking deltas arrive first, then content deltas.
if (chunk.thinking) {
process.stdout.write(`[think] ${chunk.thinking}`);
}
if (chunk.delta) {
process.stdout.write(chunk.delta);
}
}Provider Support
The SDK handles three distinct implementation patterns for reasoning, so your code stays the same regardless of provider.
| Provider | Support | How It Works |
|---|---|---|
| Ollama | ✅ Native | Ollama returns a dedicated thinking field in its response. The SDK passes think: true directly to the Ollama API. |
| OpenAI | 🚧 Model-dependent | Models like o1 encapsulate reasoning internally. Extraction depends on the model version and API support. |
| Generic | 🔄 Emulated | Models that output <think> XML tags in their content (like DeepSeek R1 via OpenRouter) are parsed by the SDK using the parseThinking utility. |
Utility: Parsing Raw Thinking
If you're processing raw model output that contains <think> tags — for example, from an OpenRouter response — the SDK provides a parseThinking utility that handles three edge cases:
import { parseThinking } from 'tekimax-ts';
// Case 1: Complete think block
const complete = parseThinking("<think>Calculating...</think> The answer is 42.");
// { think: "Calculating...", rest: "The answer is 42." }
// Case 2: Streaming — tag opened but not closed yet.
// This happens mid-stream when the model is still reasoning.
const streaming = parseThinking("<think>Still working on it...");
// { think: "Still working on it...", rest: "" }
// Case 3: No think tag at all
const none = parseThinking("Just a normal response.");
// { think: null, rest: "Just a normal response." }This utility is used internally by the SDK, but it's exported so you can use it in custom streaming implementations.