JioHotstar reveals best ways to make AI coding assistants smarter

JioHotstar engineers recently tested six different methods to help AI coding agents work better with large codebases. They compared tools like Sourcebot, Graphify, CocoIndex Code, Repowise, and their own custom retrieval system against a basic filesystem approach, evaluating performance on established coding benchmarks SWE-bench and CodeScaleBench using Claude Sonnet as the AI model.

The results show there's no magic bullet. For well-defined engineering tasks, a simple filesystem search was actually quite competitive. Semantic retrieval was most helpful when the AI had to interpret vague requirements, but added unnecessary overhead for straightforward fixes. The key takeaway is that teams should pick their code search tools based on the specific task at hand, and that AI agents, much like human developers, prefer simple, familiar operations over complex abstractions.