VectorMatrix.wiki

Featured practices

• Turn the reasoning channel off on Gemma 4, in every request

  Send chat_template_kwargs {enable_thinking: false} with every call, even though the server defaults it off. A grammar-constrained call went from 600 tokens and 8.7 s to...

• Catch the transport error family and set a short connect timeout

  A stopped service does not always refuse; it can hang until a connect timeout, which is a sibling of the connection error, not a subclass. Catch the family, use about...

• Automated security testing runs only against systems you own, with explicit authorization, never production

  Any agent that executes real attacks against an application must confirm authorization before every run, be scoped to development or staging targets, and keep its tooling...

• A JSON schema constrains the answer, not the thinking

  A response_format guarantees valid JSON only if the model produces content at all. A model that finds the input ambiguous can spend its whole budget reasoning and return...

• Parse JSON candidates from the end

  Reasoning models leave rejected drafts in their output. The first JSON object is usually not the answer. Strip think blocks, then walk candidates backwards and take the...

All practices