ibm/task-aware-embedding-refinement

IBM Task-Aware Embedding Refinement is a research tool that improves how AI systems search and classify text, specifically in situations where the AI has not been trained on the specific task, a setting called zero-shot. The core idea is to make search queries smarter at test time, without retraining anything. Here is the problem it solves. Embedding models convert text into numbers that represent meaning. These representations are efficient, but they sometimes miss subtle nuances in what a user actually wants. This tool takes a query, retrieves some candidate results, asks a stronger AI model to judge how relevant those candidates are, and then uses that feedback to nudge the query's numerical representation closer to the right answer through a process called gradient descent. The result is a more accurate search or classification without touching the underlying model. It supports multiple AI providers for the feedback step, including a LiteLLM proxy, the OpenAI API, or local models. It includes evaluation across several benchmark datasets covering tasks like literature search, intent detection, key point matching, and instruction following retrieval. This is a research codebase from IBM aimed at machine learning practitioners who want to boost retrieval quality without retraining. The project is written in Python. The full README is longer than what was shown.