Retrieval quality beats model size for most coding helpers. A tiny index of your own docs and past PRs beats a giant general context dump. I keep the corpus narrow and refresh it when the repo moves.