Worked Example: Developer & Creator Tools

A code assistant needs to understand an entire repository, not just the file currently open, generate code that fits existing architecture and conventions, and pass the project's own tests. This is RAG and agency composed together, and the composition is the entire design.

Retrieval here works exactly like Tier 8, with one specific adaptation: code chunking is not text chunking. Splitting by semantic unit, whole functions, whole classes, and always including a function's type signature and imports alongside its body, is what makes retrieved code actually usable. A function body without its type signature is close to useless context, since the model cannot correctly call something it cannot see the contract for.

Generation then runs through the exact sandboxed generate-execute-fix loop from Tier 9's sandboxed code execution chapter: generate code in context, run the project's real tests, and if they fail, feed the actual error back to the model and iterate, capped at a fixed number of attempts, same as every other iterative loop this course has built.

A platform processing 500,000 posts an hour currently catches only 40% of policy violations, with a 15% false positive rate wrongly removing legitimate content. The target: 95%+ violations caught, under 3% false positives. Running every single post through a full LLM call would be prohibitively expensive at this volume, so the design has to be a funnel, not a single stage.

This is the same precision-recall tradeoff from Phase 1's evaluation chapter, deliberately exploited on purpose. Stage 1's job is not to be right. Its job is to never let a real violation through, even at the cost of some false alarms, since stage 2 is specifically built to correct exactly that kind of mistake. Optimizing stage 1 for precision instead of recall would be a serious design error here, since a violation stage 1 misses never reaches stage 2 at all.