Questions

I have 50k labeled support tickets to route into 12 categories. Prompting GPT 5 works but costs add up at volume. When does fine tuning a small model pay off?

When a tool returns an error my agent just calls it again with the same args, indefinitely. I expected it to adapt. Using a basic ReAct loop.

Building retrieval over long technical docs. Fixed 512 token chunks are simple but cut mid sentence and hurt recall. Is semantic chunking worth the complexity, or is overlap enough…

Sharing a snippet that cut hallucinations noticeably in my eval set. Works across Claude and GPT. Curious if others have a tighter version.

I need guaranteed parseable JSON. Response format json object helps but it still occasionally wraps things in prose or markdown fences. What is the current best practice?

My agent narrates every single tool call ("Now I will read the file...") which bloats the transcript. I want it to just act and only speak when it has something for the user. What…