🧭JioHotstar Tests Six Retrieval Methods for AI Coding Agents on SWE-bench

Your coding agent might not need that vector database

Bottom line: there's no universal retrieval layer for coding agents. JioHotstar tested six approaches and found that the right choice depends entirely on the task, not on what's trendy this quarter. If you're wiring up Claude-based agents to a large monorepo, this matters. The filesystem baseline stayed competitive on concrete tasks, so bolting on a retrieval layer may just add overhead and token spend. Semantic retrieval earned its keep only when requirements were ambiguous or cross-cutting. The numbers: Sourcebot beat the baseline on harder SWE-bench Verified tasks while costing less. On simpler SWE-bench Lite tasks, retrieval mostly added overhead. CodeScaleBench showed semantic retrieval pulling ahead when agents had to interpret fuzzy problem descriptions. Claude Sonnet was the solving model throughout. Worth watching: the team argues agents should know more than code. Architecture, decisions, incidents, and outcomes could be the next layer of context. That's a bigger bet, and this study doesn't prove it yet.