AI Architecture Bootcamp Archives - Page 6 of 10

Day 27: RAG Freshness: Git Webhooks vs Nightly Jobs

Stop serving yesterday’s main: design webhook-driven and nightly RAG reindex paths so hourly merges do not create silent index drift.

Force file:line citations in code RAG and fail closed when the quoted span was never retrieved — stop shipping eloquent guesses.

Draft-then-verify routing: cheap model drafts, expensive model gates on risky paths — tests remain ground truth.

Prompt cache keys must include rules SHA, tools schema, and guardrail version or you serve stale policy after a repo move.

Show token budgets per PR and per engineer next to the diff — soft-warn at 80%, hard-stop at 100%.

Project: ephemeral ECS Fargate coding-agent sandbox with no prod creds, locked egress, and evidence bundle on exit.

OPA/Cedar policy-as-code in front of shell and apply tools — prompts are not controls.

Stop token bombs and recursive invokes with input limits, depth caps, and per-tenant daily budgets.

Pin and verify agent image digests and allowlist model IDs — SLSA-ish provenance before Fargate/Lambda pulls.

Append-only session event logs with artifact hashes so humans can replay agent runs without guessing.