Iterative Retrieval-Generation

Iter-RetGen enhances Minos LLMs by integrating retrieval-augmented generation (RAG) in an iterative process. Unlike traditional RAG, where retrieval and generation are loosely coupled, Iter-RetGen actively involves the LLM in refining retrieval queries and processing retrieved knowledge holistically.

This iterative synergy enables AI agents to dynamically incorporate external knowledge, reason over complex trading strategies, and mitigate hallucinations making it ideal for tasks requiring real-time data and multi-step reasoning in order to execute trades.

Iter-Retgen workflow

1. Initial Generation: When user creates a trading strategy request Minos LLM produces an initial output (e.g., a partial answer or hypothesis) based on a query and its parametric knowledge.

2. Context-Aware Retrieval: The initial output informs a retrieval step, where the system fetches relevant external knowledge (on-chain data) using the output as context to improve relevance.

3. Iterative Refinement: The retrieved knowledge is fed back into the LLM, which generates an improved output. This cycle repeats, with each iteration refining retrieval queries and outputs until a satisfactory result is achieved.

4. Generation-Augmented Retrieval Adaptation: The LLM can further adapt the retriever by fine-tuning its relevance modeling based on generated outputs, enhancing future retrieval accuracy.