# Topic 26: Groq Reasoning Implementation

**Audit Date**: 2026-02-01
**Auditor**: Agent Antigravity
**Scope**: Advanced Cognitive Capabilities

---

## 1. Native Reasoning Integration
The system leverages Groq's support for "Reasoning Models" (like DeepSeek-R1, Qwen 2.5, GPT-OSS) to perform complex forensic analysis *before* generating a response.

### 1.1 Supported Models
The `ModelRegistry` tags specific models with `Capability.REASONING`:
*   `openai/gpt-oss-20b` (DeepSeek-R1 Distill)
*   `qwen/qwen3-32b` (QwQ)

### 1.2 Reasoning Formats
The `GroqClient` in `llm_client.py` supports all three modes documented by Groq:

| Format | Implementation | Usage |
| :--- | :--- | :--- |
| **`raw`** | Returns `<think>...</think>` tags in content. | Used for "Glass Box" debugging logs. |
| **`parsed`** | Separates reasoning into `message.reasoning`. | **DEFAULT**. Used for storing thought process in DB. |
| **`hidden`** | Discards reasoning. | Used for user-facing responses (when speed matters). |

**Code Evidence**:
```python
# app/core/llm_client.py
if reasoning_format:
    payload["reasoning_format"] = reasoning_format
elif include_reasoning is not None:
    payload["include_reasoning"] = include_reasoning
```
*Note: The code explicitly handles the mutual exclusivity of `reasoning_format` and `include_reasoning`.*

---

## 2. Reasoning Effort Configuration
For models that support variable compute (like `qwen3-32b`), the client allows setting `reasoning_effort`:
*   **`low`**: For quick scam classification.
*   **`high`**: For complex "Math Forensics" (checking suspicious investment returns).

---

## 3. The "Glass Box" Audit
Unlike standard LLMs that are "Black Boxes", Sentinel captures the **Chain of Thought (CoT)**.
*   The `_native_reasoning` field is extracted and saved to the `Intelligence` table in the database.
*   This allows auditors to see *exactly* why the AI classified a message as a threat.

**Status**: **FULLY IMPLEMENTED**. The system treats Reasoning as a first-class citizen embedded in the forensic pipeline.