EmbeddingGemma 2 and Qwen3-Embedding 0.6B against bge-m3 on one workload, clustering news articles from Vietnamese newspapers into stories. Quality was level, each model has its own cosine scale, and the one real gap covers under one percent of pairs.

I put four model profiles on LiteLLM, sent back two comments and two issues, read how 9Router handles the same cases, and still run 3,000 lines of my own. Here is what worked, what did not, and why.

A two-second capacity dip, four retry policies and one retry budget, simulated in about a hundred lines of Go: how retries multiply load, why jitter helps, and what a budget changes.

Why a custom logits processor that runs once per request can dominate tail latency even when the GPU batches perfectly, and the batch-level rewrite that fixes it.

Open-weight models are not catching up to frontier APIs on benchmarks. They are already ahead on the number that decides most production budgets: cost per request at scale.

Saving to the database and publishing an event are two writes to two systems that fail independently. The transactional outbox pattern makes the event part of the same database transaction, so what commits is what gets published - here is a working Go and PostgreSQL implementation, including the retry, ordering, latency and cleanup details most guides skip.