Tool schemas quietly weaken your model's refusal behavior. Here is what a new agent-safety paper found and the checks I now run before shipping any…
RAG or fine-tuning? One question settles it most of the time. What the benchmarks leave out, where retrieval quietly fails, and the hybrid split I…
Prompt tone changes LLM output token count by up to 44%, per a new arXiv paper. Here is what I measured on my own prompts,…
Adding more few-shot examples can quietly make your LLM prompt worse. Here is what the over-prompting research found and how I size the example set…
SLM vs LLM in 2026: why most of your model calls do not need a frontier model, the small-first cascade pattern that cuts cost, and…
My long-context RAG bot kept refusing questions whose answers were right there. The fix was in my anti-hallucination prompt, not my retriever.
I ran Claude's prompt caching in production for six months. Here's the cache_control setup, TTL trade-offs, and the anti-patterns that burned me.
pgvector plus HNSW replaced Pinecone on half my side projects. The schema, the index picks, hybrid search with FTS, and where I still reach for…
I ran an LLM judge for six weeks before I noticed it kept rewarding longer answers. Here is the evaluation metrics stack I actually trust…
Fine-tune an LLM and a small open model can beat GPT-5 on a narrow task, cheaper. A look at when QLoRA wins, when it loses,…