Thoughts on Azure, networking, and cloud infrastructure
Prompts are code. They need version control, testing, deployment pipelines, and rollback strategies. Most teams treat them like config files and wonder why their AI quality is inconsistent. Here is how to manage prompts properly.
Most teams instrument their AI apps with latency and error rate dashboards. Those are necessary but not sufficient. The failures that hurt production AI are invisible to standard APM tooling. Here are the four signals that actually matter.
Pure vector search sounds like the obvious choice for RAG. In practice, it misses exact matches, product codes, and proper nouns that keyword search catches trivially. Hybrid search using RRF consistently outperforms either approach alone.
Chunking is the step most teams rush past when building RAG systems. The wrong chunking strategy makes your search retrieve the right document but the wrong passage — and the LLM hallucinates the rest. Here is what works.
Not every Azure OpenAI deployment should look the same. Direct API, APIM gateway, and private VNet topologies each solve a different set of requirements. Here is when to use each one.
