Ollama JavaScript Library

The Ollama JavaScript library provides the easiest way to integrate your JavaScript project with Ollama.

Getting Started

npm i ollama

Usage

1import ollama from 'ollama'
2
3const response = await ollama.chat({
4  model: 'llama3.1',
5  messages: [{ role: 'user', content: 'Why is the sky blue?' }],
6})
7console.log(response.message.content)

Browser Usage

To use the library without node, import the browser module.

1import ollama from 'ollama/browser'

Streaming responses

Response streaming can be enabled by setting stream: true, modifying function calls to return an AsyncGenerator where each part is an object in the stream.

1import ollama from 'ollama'
2
3const message = { role: 'user', content: 'Why is the sky blue?' }
4const response = await ollama.chat({ model: 'llama3.1', messages: [message], stream: true })
5for await (const part of response) {
6  process.stdout.write(part.message.content)
7}

API

The Ollama JavaScript library's API is designed around the Ollama REST API

chat

1ollama.chat(request)

• request <Object>: The request object containing chat parameters.

  • model <string> The name of the model to use for the chat.
  • messages <Message[]>: Array of message objects representing the chat history.
    • role <string>: The role of the message sender ('user', 'system', or 'assistant').
    • content <string>: The content of the message.
    • images <Uint8Array[] | string[]>: (Optional) Images to be included in the message, either as Uint8Array or base64 encoded strings.
  • format <string>: (Optional) Set the expected format of the response (json).
  • stream <boolean>: (Optional) When true an AsyncGenerator is returned.
  • think <boolean>: (Optional) When true, the model will think about the response before responding. Requires thinking support from the model.
  • keep_alive <string | number>: (Optional) How long to keep the model loaded. A number (seconds) or a string with a duration unit suffix ("300ms", "1.5h", "2h45m", etc.)
  • tools <Tool[]>: (Optional) A list of tool calls the model may make.
  • options <Options>: (Optional) Options to configure the runtime.

• Returns: <ChatResponse>

generate

1ollama.generate(request)

• request <Object>: The request object containing generate parameters.
  • model <string> The name of the model to use for the chat.
  • prompt <string>: The prompt to send to the model.
  • suffix <string>: (Optional) Suffix is the text that comes after the inserted text.
  • system <string>: (Optional) Override the model system prompt.
  • template <string>: (Optional) Override the model template.
  • raw <boolean>: (Optional) Bypass the prompt template and pass the prompt directly to the model.
  • images <Uint8Array[] | string[]>: (Optional) Images to be included, either as Uint8Array or base64 encoded strings.
  • format <string>: (Optional) Set the expected format of the response (json).
  • stream <boolean>: (Optional) When true an AsyncGenerator is returned.
  • think <boolean>: (Optional) When true, the model will think about the response before responding. Requires thinking support from the model.
  • keep_alive <string | number>: (Optional) How long to keep the model loaded. A number (seconds) or a string with a duration unit suffix ("300ms", "1.5h", "2h45m", etc.)
  • options <Options>: (Optional) Options to configure the runtime.
• Returns: <GenerateResponse>

pull

1ollama.pull(request)

• request <Object>: The request object containing pull parameters.
  • model <string> The name of the model to pull.
  • insecure <boolean>: (Optional) Pull from servers whose identity cannot be verified.
  • stream <boolean>: (Optional) When true an AsyncGenerator is returned.
• Returns: <ProgressResponse>

push

1ollama.push(request)

• request <Object>: The request object containing push parameters.
  • model <string> The name of the model to push.
  • insecure <boolean>: (Optional) Push to servers whose identity cannot be verified.
  • stream <boolean>: (Optional) When true an AsyncGenerator is returned.
• Returns: <ProgressResponse>

create

1ollama.create(request)

• request <Object>: The request object containing create parameters.
  • model <string> The name of the model to create.
  • from <string>: The base model to derive from.
  • stream <boolean>: (Optional) When true an AsyncGenerator is returned.
  • quantize <string>: Quanization precision level (q8_0, q4_K_M, etc.).
  • template <string>: (Optional) The prompt template to use with the model.
  • license <string|string[]>: (Optional) The license(s) associated with the model.
  • system <string>: (Optional) The system prompt for the model.
  • parameters <Record<string, unknown>>: (Optional) Additional model parameters as key-value pairs.
  • messages <Message[]>: (Optional) Initial chat messages for the model.
  • adapters <Record<string, string>>: (Optional) A key-value map of LoRA adapter configurations.
• Returns: <ProgressResponse>

Note: The files parameter is not currently supported in ollama-js.

delete

1ollama.delete(request)

• request <Object>: The request object containing delete parameters.
  • model <string> The name of the model to delete.
• Returns: <StatusResponse>

copy

1ollama.copy(request)

• request <Object>: The request object containing copy parameters.
  • source <string> The name of the model to copy from.
  • destination <string> The name of the model to copy to.
• Returns: <StatusResponse>

list

1ollama.list()

• Returns: <ListResponse>

show

1ollama.show(request)

• request <Object>: The request object containing show parameters.
  • model <string> The name of the model to show.
  • system <string>: (Optional) Override the model system prompt returned.
  • template <string>: (Optional) Override the model template returned.
  • options <Options>: (Optional) Options to configure the runtime.
• Returns: <ShowResponse>

embed

1ollama.embed(request)

• request <Object>: The request object containing embedding parameters.
  • model <string> The name of the model used to generate the embeddings.
  • input <string> | <string[]>: The input used to generate the embeddings.
  • truncate <boolean>: (Optional) Truncate the input to fit the maximum context length supported by the model.
  • keep_alive <string | number>: (Optional) How long to keep the model loaded. A number (seconds) or a string with a duration unit suffix ("300ms", "1.5h", "2h45m", etc.)
  • options <Options>: (Optional) Options to configure the runtime.
• Returns: <EmbedResponse>

ps

1ollama.ps()

• Returns: <ListResponse>

abort

1ollama.abort()

This method will abort all streamed generations currently running with the client instance. If there is a need to manage streams with timeouts, it is recommended to have one Ollama client per stream.

All asynchronous threads listening to streams (typically the for await (const part of response)) will throw an AbortError exception. See examples/abort/abort-all-requests.ts for an example.

Custom client

A custom client can be created with the following fields:

• host <string>: (Optional) The Ollama host address. Default: "http://127.0.0.1:11434".
• fetch <Object>: (Optional) The fetch library used to make requests to the Ollama host.

1import { Ollama } from 'ollama'
2
3const ollama = new Ollama({ host: 'http://127.0.0.1:11434' })
4const response = await ollama.chat({
5  model: 'llama3.1',
6  messages: [{ role: 'user', content: 'Why is the sky blue?' }],
7})

Building

To build the project files run:

1npm run build