`ollama://`¶

Ollama — local LLM runner. Talks to a local Ollama daemon (default http://localhost:11434) over its native chat and embeddings API. Use this when you want to run open-weights models (Llama 3, Mistral, Qwen, Phi-3, etc.) on your own machine or in your cluster without leaving the network. No API key required — auth is purely network-level.

URI shape¶

ollama://chat?model=<name>&baseURL=<url>&temperature=<float>
ollama://embeddings?model=<name>&baseURL=<url>

Parameters¶

Param	Type	Default	Producer/Consumer/Both	Description
path	enum	required	Producer	`chat` or `embeddings`.
`model`	string	`llama3`	Producer	Ollama model tag (must be pulled on the daemon first via `ollama pull`). Common: `llama3`, `mistral`, `qwen2:7b`, `nomic-embed-text` (embeddings).
`baseURL`	string	`http://localhost:11434`	Producer	Ollama daemon URL. Override to talk to a remote daemon or sidecar.
`temperature`	float	`0.7`	Producer	Sampling temperature, passed inside the `options` object.

No apiKey parameter — Ollama doesn't authenticate. No maxTokens parameter — Ollama uses model-default max output unless the model itself caps it.

Producer-only.

Headers¶

Reads:

Header	Effect
`systemPrompt`	Prepended as a `role: system` message.

Sets:

Header	Description
`ollama.model`	Model id from the response.
`ollama.done`	Boolean — generation completed.
`ollama.metrics.promptEvalCount`	Tokens evaluated for the prompt.
`ollama.metrics.evalCount`	Tokens generated.

Examples¶

Basic local chat:

- id: local-chat
  from: direct://chat
  steps:
    - to: ollama://chat?model=llama3

Remote daemon with explicit URL:

- to: ollama://chat?model=qwen2:7b&baseURL=http://gpu-host:11434&temperature=0.2

Embeddings for a RAG pipeline:

- id: embed
  from: direct://embed
  steps:
    - to: ollama://embeddings?model=nomic-embed-text

Behavior notes¶

300s timeout — longer than other providers because local models on CPU can take a long time to load and respond on first call.
Streaming disabled. The component sets stream: false for simplicity; the daemon returns one JSON blob. The first request after a daemon start typically pays a model-load tax of several seconds.
Model must be pulled. If the daemon doesn't have the model cached, the call fails with model 'X' not found, try pulling it first — ollama pull <model> on the daemon host fixes it.
Chat path: POST {baseURL}/api/chat. Embeddings path: POST {baseURL}/api/embeddings (singular prompt, not the OpenAI plural input).
Embedding response shape is {embedding: [...]} — Helix writes the float slice directly to the exchange body.

Multimodal input (vision models)¶

ollama://chat works with any Ollama vision model (llava, llama3.2-vision, granite3.2-vision, qwen2-vl, etc.). Images travel in Ollama's native images: []string (base64) field alongside the text in content.

Body	Headers	What gets sent
`string` or `[]byte` (non-image)	—	Text-only message, no `images` field.
`[]byte` w/ image magic	`HelixLLMSystemPrompt` for the prompt	`images: ["<base64>"]` + text from the system prompt.
Any text body	`HelixLLMImages` = `[]string` of URLs	URLs get inlined as `[image URL: …]` text — Ollama's chat endpoint does not fetch URLs server-side. Use the `file://` step to download first, then pass bytes.
`*exchange.MultipartContent`	—	Pass-through.

- to: ollama://chat?model=granite3.2-vision&baseURL=http://localhost:11434
  # body: PNG/JPEG bytes from upstream

Source¶

pkg/component/ollama/ollama.go
pkg/component/ollama/ollama_test.go

ollama://¶