JioHotstar Benchmarks Six Code Retrieval Approaches for AI Coding Agents

JioHotstar tested six code retrieval methods for AI coding agents using the SWE and CodeScale benchmarks plus an internal question-answering harness. Structural search performed best on concrete tasks, while semantic retrieval helped more with ambiguous requirements. The results show trade-offs in accuracy, token consumption, and tool design.

Why it matters: Teams building or deploying coding agents need to match retrieval tools to task type, since the choice directly affects accuracy and token cost.