Hard-Negative Mining for Enterprise Retrieval
Hard Negative Mining for Domain-Specific Retrieval studies training examples that are semantically similar to a query but irrelevant to its actual information need. These challenging negatives help train rerankers for enterprise search and downstream RAG applications.
Authors: Hansa Meghwani, Amit Agarwal, Priyaranjan Pattnayak, Hitesh Laxmichand Patel, Srikant Panda. Venue: ACL 2025 Industry Track.
Retrieval training method
The framework combines multiple embedding models with dimensionality reduction to select useful hard negatives efficiently. The paper evaluates the approach on a cloud-services corpus and additional public domain-specific retrieval datasets.
This work contributes to my interest in grounding AI systems in relevant evidence, including the retrieval components used in conversational and agentic workflows.
Paper and citation
Explore conversational retrieval with RECOR and more publications.