ollama-ai-provider

Vercel AI Provider for running Large Language Models locally using Ollama

Note: This module is under development and may contain errors and frequent incompatible changes.

All releases will be of type MAJOR following the 0.MAJOR.MINOR scheme. Only bugs and model updates will be released as MINOR. Please read the Tested models and capabilities section to know about the features implemented in this provider.

Installation

The Ollama provider is available in the ollama-ai-provider module. You can install it with

npm i ollama-ai-provider

Provider Instance

You can import the default provider instance ollama from ollama-ai-provider:

import { ollama } from 'ollama-ai-provider';

If you need a customized setup, you can import createOllama from ollama-ai-provider and create a provider instance with your settings:

import { createOllama } from 'ollama-ai-provider';

const ollama = createOllama({
  // custom settings
});

You can use the following optional settings to customize the Ollama provider instance:

• baseURL string

  Use a different URL prefix for API calls, e.g. to use proxy servers. The default prefix is http://localhost:11434/api.

• headers Record<string,string>

  Custom headers to include in the requests.

Models

The first argument is the model id, e.g. phi3.

const model = ollama('phi3');

Examples

Inside the examples folder, you will find some example projects to see how the provider works. Each folder has its own README with the usage description.

Tested models and capabilities

This provider is capable of generating and streaming text and objects. Object generation may fail depending on the model used and the schema used.

At least it has been tested with the following features:

Image input	Object generation	Tool usage	Tool streaming
✅	✅	⚠️	⚠️

Image input

You need to use any model with visual understanding. These are tested:

• llava
• llava-llama3
• llava-phi3
• moondream

Object generation

This feature is unstable with some models

Some models are better than others. Also, there is a bug in Ollama that sometimes causes the JSON generation to be slow or end with an error. In my tests, I detected this behavior with llama3 and phi3 models more than others like openhermes and mistral, but you can experiment with them too.

More info about the bugs:

• ollama/ollama#3851
• ollama/ollama#3785

Remember that Ollama and this module are free software, so be patient.

Tool usage (no streaming)

This feature is not completed and unstable

Ollama does not support tooling, so this provider simulates tool usage with prompt injection. That means that this feature can fail very often. Again, it depends on the model you use, and it is very related to the object generation issues explained in the previous section.

I recommend you use openhermes and mistral or experiment with your preferred models.

Tool streaming

This feature is not completed and unstable

Again, since Ollama does not support tooling, we should simulate the feature. In this case, the problem is worse than in non-streaming tool usage. We don't have the full response before knowing if the model has detected function calling. We are waiting for the first characters before sending the deltas to detect if we are in a tool call flow.

Obviously, this is very buggy and should be used with caution. Right now, you cannot use it in chats and with more than one tool.