llm-multitool

llm-multitool is a local web UI for working with large language models (LLM). It oriented towards instruction tasks and can connect to and use different servers running LLMs.

Features

• Aims to be easy to use
• Supports different LLM backends/servers including locally run ones:
  • OpenAI's ChatGPT
  • Ollama
  • and most backends which support the OpenAI API such as LocalAI and Oobabooga.
• "Instruction" templates to simplify certain tasks.
• Chat support for backends which support it.
• Multiple persistent sessions and history

Downloading

Executables for Windows, macOS, and Linux can be downloaded from the Releases page.

Building from source

Instead of downloading a precompiled executable you can also build it from source.

Build Dependencies

• This project uses Taskfile for building. The task executable must be available in your path. You can install the task binary from https://taskfile.dev/ .
• Go compiler
• Recent Node JS version.

Building

Run task prepare and task build to prepare and build llm-multitool. It will first build the web frontend and then the Go based backend. The output executable will be in the backend/ folder and named llm-multitool.

Backends

llm-multitool is just a UI for LLMs. It needs a LLM backend to connect to which actually runs the model. But which one should you use?

Short answer:

• If you don't want to run a LLM locally then you should set up OpenAI's ChatGPT.
• If you do want to run a LLM locally then try Ollama.

The backends in more detail:

• OpenAI's ChatGPT - Support for this backend is the most complete and stable. It does require setting up billing and an API token at OpenAI to use.
• Ollama - This can run LLMs locally, is easy to set up, and supports Linux, Windows, and macOS. Version v0.1.14 or later is required.
• LocalAI - This will also let you run LLMs locally, is easy to set up and supports many different LLMs, but only runs Linux and macOS. llm-multitool supports this quite well via it's OpenAI API.
• Oobabooga text-generation-ui - This backend can be a challenge to install and isn't really meant for end users. It does support have many LLM types. llm-multitool support for this mostly works but is buggy.

Note: It should be possible to connect llm-multitool to most things which support the OpenAI API.

Configuring a backend

Before running llm-multitool you first need to write a small configuration file in yaml to set up which backend(s) it should connect to and how. By default you can name this file backend.yaml. Its name can be specified when starting llm-multitool.

This configuration file consists of 1 or more backend configurations in a YAML list. You may have as many backends configured as you want. At start up llm-multitool will query each backend for its list of models. If a backend is not available, then it is jus skipped.

Below are sections on how to configure the different backends.

OpenAI Configuration

The following example backend.yaml file shows how to connect to OpenAI's ChatGPT model. You need to generate your own API token on OpenAI's website to use in the file.

- name: OpenAI
  api_token: "sk-FaKeOpEnAiToKeN7Nll3FAKzZET3BlbkFJLz8Oume19ZeAjGh3rabc"
  models:
  - gpt-3.5-turbo
  - gpt-4

The name field can be any name you like, but it is best to keep it short.

api_token holds the value of the token you generated at OpenAI.

If you don't want to copy your token directly into your configuration file, you can omit the api_token file and replace it with api_token_from with a value naming a text file from which to read the read token from. The file path is relative to the backend.yaml file.

models is a list of model to permit. OpenAI have many different models and varieties, but only a handful of the the LLMs are useful for use with llm-multitool.

Ollama

llm-multitool can connect to a Ollama server via its own API. The configuration block is as follows:

- name: Ollama
  address: "http://localhost:11434"
  variant: ollama

The name field can be any name you like, but it is best to keep it short.

The address field is the URL of the Ollama server.

The variant field must have the value "ollama".

LocalAI

LocalAI and OpenAI configuration is the same thing except that LocalAI needs an address value to be specified and it doesn't require the token or model values.

- name: LocalAI
  address: "http://127.0.0.1:5001/v1"

The address field is the URL of the LocalAI server.

Oobabooga text-generation-ui

llm-multitool can connect to a running Oobabooga text-generation-ui server via its OpenAI extension.

⚠️ You need to set up and activate the openai extension in your Oobabooga installation. The README.md file under Oobabooga's extensions/openai/ folder gives more details. Also, when starting Oobabooga you need to turn the extension on.

The configuration needed in backend.yaml is typically:

- name: Ooba
  address: "http://127.0.0.1:5001/v1"
  variant: oobabooga

The name field can be any name you like, but it is best to keep it short.

The address field is the URL of the OpenAI end-point running on Oobabooga.

The variant field must have the value "oobabooga".

Running

If you have built llm-multitool from source then the executable will be in backend/llm-multitool and you should have written a minial backend.yaml file. Start up llm-multitool with:

backend/llm-multitool -c backend.yaml

This will start the server and it will listen on address 127.0.0.1 port 5050 by default.

Open your browser on http://127.0.0.1:5050 to use the llm-multitool UI.

Command line reference

usage: llm-multitool [-h|--help] [-c|--config "<value>"] [-s|--storage
"<value>"] [-p|--presets "<value>"] [-t|--templates
"<value>"] [-a|--address "<value>"]

Web UI for instructing Large Language Models

Arguments:

-h  --help       Print help information
-c  --config     Path to the configuration file. Default: backend.yaml
-s  --storage    Path to the session data storage directory. Default: data
-p  --presets    Path to the file containing generation parameter presets.
Default:
-t  --templates  Path to the file containing templates. Default:
-a  --address    Address and port to server from. Default: 127.0.0.1:5050

Custom instruction templates

llm-multitool has a small set of built in templates for instruct type tasks. You can read this yaml file up on GitHub here. It is possible to create your own templates file and tell llm-multitool to use it with the -t command line option.

The format of the templates yaml file is an array of template objects. Each template object has the following fields:

• id - A unique string to identify the template. UUIDs work well here, but any string is accepted
• name - The name of the template. This will be shown in the web UI. For example, "Translate to French"
• template_string - The template for the prompt itself. The string {{prompt}} will be replaced with what ever the user enters as the prompt in the web UI.

If you write an interesting template, consider submitting it to this project for inclusion.

Custom parameter presets

llm-multitool has a small set of built in parameter presets. These control the generation of responses. You can read this yaml file up on GitHub here. It is possible to create your own presets file and tell llm-multitool to use it with the -p command line option.

• id - A unique string to identify the preset.
• name - The name of the preset. This will be shown in the web UI.
• temperature - a numeric value specifying the temperature value to use during generation. For example, 0.8
• top_p - a numeric value specifying the Top P setting to use during generation. For example, 0.1

License

MIT

Author

Simon Edwards simon@simonzone.com