Ollama Provider

Use Ollama when you want funcai to call local models through the Ollama API.

Import

import { createAiFn } from "funcai";
import { ollama } from "funcai/providers/ollama";

const ai = createAiFn({
  provider: ollama(),
});

Configure Ollama

Install and start Ollama, then pull a model:

ollama pull gemma4:latest

The default base URL is:

http://127.0.0.1:11434

Use the default local server:

const ai = createAiFn({
  provider: ollama(),
});

Or point at another host:

const ai = createAiFn({
  provider: ollama({
    baseURL: "http://192.168.2.188:11434",
  }),
});

Optional advanced fields:

ollama({
  baseURL: "http://127.0.0.1:11434",
  apiKey: "optional-proxy-key",
  headers: { "x-local-run": "dev" },
  fetch: customFetch,
});

Model IDs

Ollama model IDs are local strings, so OllamaModelId is intentionally open:

const classify = ai.fn({
  model: "gemma4:latest",
  // ...
});

Set OLLAMA_MODEL in examples when your local model ID differs.

Structured output

All funcai calls use schema-backed object generation. Model reliability depends on the local model and quantization. Strong instruction-following models with native JSON support are the best fit.

For multimodal examples, use a model that actually accepts image input through your installed Ollama version.

Smoke test

Start Ollama, pull the model, then run:

Set OLLAMA_BASE_URL and OLLAMA_MODEL, then run pnpm exec vitest run --config vitest.e2e.config.ts tests/e2e/ollama-live.test.ts.

Examples:

Set OLLAMA_BASE_URL and OLLAMA_MODEL, then run pnpm -C examples ollama:vision.

Troubleshooting

Connection refused: start Ollama and confirm the base URL does not include /v1.
Model not found: run ollama pull <model> or use an installed model from ollama list.
Schema parse failures: use a stronger model, simplify the schema, add few-shots, or lower prompt ambiguity.
Multimodal input fails: confirm the selected model and installed Ollama version support image input.

FilesExpand file tree

README.md

Latest commit

History