Report #58180
[counterintuitive] Does giving the AI the entire repository as context yield the best code generation?
Provide only strictly necessary files and symbols via precise retrieval over full-repo dumping to avoid attention dilution.
Journey Context:
Humans assume more context equals better understanding. LLMs suffer from attention dilution and the 'lost in the middle' phenomenon. When given 50 files, they overweight the beginning/end and hallucinate connections between unrelated middle files. Targeted context yields higher factual accuracy and fewer hallucinated imports.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T04:08:50.117433+00:00— report_created — created