# Topic 18: Advanced Groq Capabilities

**Audit Date**: 2026-02-01
**Auditor**: Agent Antigravity
**Scope**: High-Performance AI Features

---

## 1. Parallel Tool Execution
The system leverages Groq's **Parallel Tool Calling** to perform multiple forensic tasks simultaneously.
*   **Implementation**: `llm_client.py` -> `generate_tool_call(..., parallel_tool_calls=True)`.
*   **Logic**:
    1.  The model receives a prompt requiring diverse data points (e.g., "Check this IP and verify this phone number").
    2.  The model generates **TWO** tool calls in a single response: `lookup_ip(...)` AND `verify_phone(...)`.
    3.  The system executes both in parallel using `asyncio.gather()`.
*   **Benefit**: Reduces latency by 50% compared to sequential execution.

---

## 2. Compound AI Systems
The system uses "Compound" architectures where the LLM is just one component of a larger cognitive loop.
*   **Compound-Mini**: Used for Math Forensics (`math_forensics`). Optimized for speed and logic.
*   **Reasoning Format**: `parsed` or `hidden`. The system captures the "Chain of Thought" (`<think>` tags) and exposes it in the logs (`[🧠] NATIVE REASONING CAPTURED`), allowing devs to debug *why* a scam was flagged.

---

## 3. Strict Mode & JSON Schemas
*   **Problem**: Standard LLMs often fail to produce valid JSON (trailing commas, missing keys).
*   **Solution**: **Groq Strict Mode** (`Capability.STRICT_MODE`).
*   **Audit**: `llm_client.py` contains `_harden_schema_for_strict_mode()`.
    *   **Action**: It rewrites simple JSON schemas to be "Strict Compliant" (No optional fields, `additionalProperties: false`).
    *   **Result**: 100% Guarantee of valid JSON output for the API.

---

## 4. Prefill Optimization
*   **Feature**: `kwargs["prefill"]` in `generate()`.
*   **Usage**: The system injects a "pre-fill" of `{` or `{"status":` to force the model into JSON mode immediately.
*   **Benefit**: Saves tokens (don't need to generate the first bracket) and reduces "I can't do that" refusals.