JioHotstar compared six ways of helping AI coding agents navigate large repositories and found that extra retrieval infrastructure did not consistently improve results. Standard filesystem search remained competitive for clearly defined tasks, while indexed structural search helped on harder code changes and semantic retrieval was more useful when requirements were ambiguous.
The evaluation covered filesystem tools, Sourcebot, Graphify, CocoIndex Code, Repowise and an internal retrieval-augmented system. Tests used subsets of SWE-bench and CodeScaleBench plus questions from JioHotstar’s own codebase, with Claude Sonnet as the solver. The team measured accuracy, token use, agent turns and how tools were used.
Sourcebot beat the baseline on harder SWE-bench Verified tasks while costing less, but retrieval often added overhead on simpler SWE-bench Lite work. Agents also tended to favor familiar operations such as text search, symbol lookup and file reading. The study’s subsets and probabilistic model behavior limit broad conclusions, yet the practical lesson is clear: teams should match retrieval to task type and preserve a simple baseline instead of assuming a larger knowledge layer will automatically help.