ollama

name	ollama
description	This skill should be used when modifying api/internal/embedding/ollama/ or api/internal/llm/ollama/, or when asked "why did the server fail to start over Ollama", "how does streaming work here", "add timeout handling to the LLM client". Documents the fail-fast model check and why the LLM client has no HTTP timeout.

Ollama clients: fail-fast startup and streaming

Fail-fast model check

Both the embedding client (embedding/ollama) and the LLM client (llm/ollama) call Ollama's /api/tags at construction time and fail immediately if the configured model isn't pulled. This means a missing ollama pull nomic-embed-text (or the configured OLLAMA_COMPLETION_MODEL) breaks server startup, not the first request. When debugging a server that won't start, check this before assuming a code regression.

No HTTP client timeout on the LLM client — deliberately

llm/ollama's http.Client has no Timeout set, unlike a typical Go HTTP client setup. This is intentional: http.Client.Timeout bounds the entire request including body read, and a long streaming completion would be truncated mid-stream by any fixed timeout. Duration is bounded by the caller's ctx instead. Do not add a Timeout back to this client without switching streaming callers to a context-based cancellation that doesn't conflict with in-flight token streaming.

Streaming error surface

Stream issues and status-checks the HTTP request synchronously — a bad model or unreachable Ollama surfaces as a normal returned error from Stream itself. Once the request succeeds, the response body is handed to a goroutine; any error that occurs after that point (e.g. a connection drop mid-stream) arrives as a StreamChunk{Error: ..., Done: true} value sent on the channel, not as a returned Go error. Callers must check StreamChunk.Error on every received chunk, not just handle Stream's initial return value.