LLM Reasoning (ReAct)

The intelligence of Rez is not just in its underlying model, but in the recursive reasoning loop it executes. This algorithm, based on the ReAct (Reasoning and Acting) framework, allows the agent to dynamically explore its environment.

The Interaction Algorithm

The TradingAgent manages a conversation history that can span up to 6 turns (default) per trading cycle.

Turn 1: The Zero-Shot Context

The agent sends the initial "System Prompt" and "Market Context".

• Prompt: "You are Rez. Your goal is to maximize PnL while adhering to the [RISK_PROFILE]..."

• Context: A broad snapshot of current prices and standard indicators (EMA 20, RSI 14).

Turn 2-N: The Recursive Discovery

The LLM analyzes the context. If it identifies a gap in its knowledge, it emits a fetch_indicator tool call.

# Pseudo-logic of the ReAct turn
while turn < max_turns:
    response = llm.generate(history)
    if response.has_tool_call:
        result = execute_tool(response.tool_call)
        history.append(tool_result)
        turn += 1
    else:
        return parse_final_decision(response)

The Reasoning Engine Prompting

Rez uses "Chain of Thought" prompting to force the model to verbalize its internal logic.

1. Analyze Trends: "Observe the slope of the 4h EMA."

2. Verify Confluence: "Check if the 5m RSI supports the long-term trend."

3. Evaluate Risk: "Assess if the distance to the next major resistance justifies the leverage."

4. Form Decision: Output the final trade parameters.

Output Parsing & Resilience

The final turn must result in a structured JSON object.

The Sanitization Recursive Step

If the primary model returns text instead of JSON, Rez triggers the _sanitize_output algorithm:

1. Extract raw strings that look like JSON.

2. Pass the raw strings to a high-speed "Resilience Model" (e.g., GPT-4o-mini).

3. The resilience model is constrained by a Strict JSON Schema.

4. Validation: The system checks that required fields (action, leverage, tp_price) are present and mathematically valid (e.g., tp_price > entry_price for a buy).

Failure Handling

If the LLM fails to provide a valid decision after the recursion limit or sanitization fails, the algorithm defaults to a hold action to protect the user's capital.