agentsop-hybrid-retrieval

What problem does it solve?

Dense-only retrievers often miss the right document when users paste exact identifiers like error codes, SKUs, API/function names, or citation strings, leading to “searched exact string and got nothing / wrong doc” failures.

Core Features & Use Cases

• Traffic-driven hybrid decisioning: adds sparse (BM25/keyword) only when real query traffic has a non-trivial lexical-identity share, avoiding unnecessary complexity.
• Correct dense+sparse fusion: wires dense and BM25 and fuses results using robust methods (e.g., RRF) rather than naïvely summing incompatible scores.
• Alpha tuning and routing: tunes the dense↔sparse blend per query type (semantic vs lexical), or routes queries to the right configuration to prevent regressions.

Quick Start

Ask your agent to assess your retrieval corpus and traffic, then implement a dense+BM25 hybrid retriever that fuses with RRF and tunes alpha per query type to recover exact-token matches without degrading semantic recall.