Reasoning (Thinking) System¶

Boris supports native reasoning trace extraction for local models that support thinking capabilities (e.g., Granite 4.2, Qwen, DeepSeek). The system intercepts reasoning fields sent by the model server and displays them in the chat before the final response.

How It Works¶

Architecture¶

User Message │ ▼ ChatService.sendMessageStream() │ ├── resolveThinkMode() → determines reasoning effort level ├── ThinkContextHolder.setThinkMode(level) │ ▼ Spring AI ChatClient │ ├── buildSpringAiOptions(effectiveThink) │ └── reasoningEffort = level (sent as "reasoning_effort") │ ▼ Ollama / Model Server │ ├── /v1/chat/completions │ └── { model, reasoning_effort, messages, tools } │ ▼ Response with thinking + content fields │ ├── ThinkContextHolder.getLastThinking() ← intercepts via HTTP interceptor └── response.content ← main text

Request Flow¶

1. Boris sends the user message + conversation history to the model server via Spring AI's ChatClient.
2. The reasoning_effort parameter is set in the request options based on the configured think level.
3. The model server (Ollama) processes the request and returns two fields:
4. thinking — the model's internal reasoning trace
5. content — the final response
6. Boris intercepts these fields via ThinkClientHttpInterceptor and ThinkContextHolder.
7. If the model doesn't use the standard thinking fields, Boris falls back to regex extraction from the response content using the pattern: (?:\b|<)thinking|<thought>|<reasoning>|(.*)?(?:\b|<\/)(?:thinking|thought|reasoning|code) ```
8. The reasoning trace is displayed in the chat as: <thinking> [model's reasoning here] </thinking>
9. The final response follows after the thinking block.

Configuration¶

Settings Parameters¶

Mapping to Spring AI¶

Boris maps the settings to Spring AI's OpenAiChatOptions.reasoningEffort:

In-App Control¶

You can toggle thinking on/off directly in the chat:

The system prompt confirms the change with a message: - ● Razonamiento (thinking) activado - ● Razonamiento (thinking) desactivado

Model Requirements¶

For Boris's reasoning system to work, your model must support:

1. OpenAI-compatible /v1/chat/completions endpoint — Boris uses this path.
2. reasoning_effort field support — The model server must accept this field in the request.
3. thinking field in response — The model server must return reasoning traces in a separate field (or Boris will fall back to regex extraction from the content).

Tested Models¶

Ollama Configuration¶

For models with native reasoning (Granite, Qwen), ensure your Ollama server is running:

bash ollama serve

Pull a model with think support:

bash ollama pull granite4.2:8b

Then configure settings.json:

json { "model": { "baseUrl": "http://localhost:11434", "name": "granite4.2:8b", "options": { "think": "high" } }, "thinkingEnabled": true, "thinkingMode": "think" }

Technical Details¶

ThinkClientHttpInterceptor¶

Intercepts HTTP requests/responses to add custom headers (API key) and capture thinking-related fields from Ollama's response.

ThinkContextHolder¶

Thread-local storage for thinking data. Each request sets the think mode, and the interceptor stores the reasoning trace. The ChatService retrieves it after each response.

ThinkMode Resolution¶

Boris resolves the effective think mode as follows:

1. Check settings.options.think — if present and non-blank, use it.
2. Check settings.model.reasoningEffort — if present and non-blank, use it.
3. If thinkingEnabled is true and no mode is found, default to low.
4. If thinkingEnabled is false, set reasoningEffort="none".