Navigation-Informed Embeddings: Adapting Dense Retrievers from Agent Traces

A recent paper on arXiv (2608.15956) presents Navigation-Informed Embeddings (NIE), a set of objectives derived from traces that allows dense retrievers to adjust to evolving workflow distributions without the need for new relevance labels, synthetic queries, or judgments from LLMs. This approach utilizes retrieval traces—comprising queries, retrievals, and stopping points—generated during question answering. NIE-Stop considers the stopping document as a soft positive, while NIE-Path incorporates earlier path documents as hard comparisons and applies ordinal constraints with geometric decay. Tests indicate that a BGE encoder, modified from retained source trajectories, enhances support Recall@20 on an independent benchmark from 72.2 to 78.0. NIE-Stop scores 76.9 overall and 52.3 on long paths, whereas NIE-Path boosts long-path performance to 55.4, up from 46.7 for the unmodified encoder. A control using shuffled order under the full path objective results in a 3.2-point drop. The research underscores the promise of utilizing agent search traces for effective retrieval adaptation.