Goll

Goll is a lightweight command-line interface (CLI) that streamlines your interaction with a local Ollama Generate API. With Goll, you can:

• Customize your prompts in one place using folder structures.
• Easily iterate over multiple prompts to save time and reduce errors.
• Chain multiple prompts for seamless workflow automation.

By automating these common tasks, Goll aims to make your local Ollama CLI experience more efficient and enjoyable.

Yes, naming is hard. goll is unique, short and easy to type.

Background

We wanted a very simple CLI to enable us to break up more complex prompts into a pipeline of smaller, task oriented prompts using local Ollama. We also wanted to be able to easily adjust the model and it's settings for each prompt in the pipeline.

Features

• Support iteration of subfolder prompts in a folder.
• Supports simple chaining of multiple prompts.
• Customize the model for each step in chain.
• Print and log each response with metrics such as tokens per second.
• Support for structured/JSON output.

Prereqs

• Go v1.23 or greater installed.
• Access to an Ollama API. Currently tested on a local instance and only uses generate endpoint.
• Ensure you have pulled whatever Ollama models you reference in your config.

Installation

1. Clone the repository:

 git clone https://github.com/ambitiousfew/goll.git
 cd goll

2. Build the CLI tool:

go build -o goll ./cmd/goll

Usage

./goll -f folder1,folder2,folder3 -p "optional beginning prompt" -v

• -f: Comma-separated list of folder names. You have to provide at least one folder.
• -p: Optional text prompt. Applied to first folder in chain, or to all subfolders if -r is used. If not present we expect a prompt.txt in first folder of a chain, or in all subfolders if -r is used.
• -r: Optional recurse of folder. If used only one folder can be set with -f flag. No prompt chaining. Will iterate over each subfolder in the given folder running each prompt. Only one level supported.
• -v: Optional verbose output. Print results of each step to command line.

Example: Chaining With Prompt

When you want to pass one prompt in and let the pipeline you create pass the output to the next folder as an input prompt.

  ./goll -f improve_prompt,basic -p "why is the sky blue"

• Starts with improving the prompt provided with the -p flag and passes the improved prompt to the basic folder. If we did not use -p flag here we would expect prompt.txt in the improve_prompt folder to read "why is the sky blue".
• Each step will generate an output log with date/time appended.

Example: Folder Recursion With Verbose output

This is useful for testing combinations of prompts/models/configs in one call. Or use the -p flag to pass the same prompt to each model/config in subfolders. Set it all up, run it, and go get coffee.

  ./goll -f improve_prompt_test -r -v

• Iterates over the improve_prompt_test folder and calls generate on each subfolder. We expect prompt.txt in each subfolder since we did not use -p flag to pass in a prompt.
• No prompt chaining supported.
• -v Will cause verbose output to print on command line.
• Each step will generate an output log with date/time appended.

Configuration

Each folder should contain the following files:

• config.json: Configuration file for the model.
• system.txt: System prompt.
• prompt.txt: User prompt.
• format.json: Optional. Contains output format schema per Ollama spec in JSON format.

Note: If chaining, prompt.txt is only required for the first folder in chain if you do not use the -p flag. Each step will write to the next folders prompt.txt file. If calling each folder recursively with -r flag then you can either pass the same prompt to each subfolder with the -p flag or you need to provide a prompt.txt in each subfolder being called.

Examples for each file can be found in prompts folder. In general, the config.json fields match the Ollama generate API spec.

Tool Settings

There is a settings.json in the root that allows you to set some global defaults for the cli.

{
  "api_base_url": "http://localhost:11434/api",
  "folder_base_path": "prompts",
  "timeout": 300
}

These are applied in every step in chain.

Timeout is in seconds and is max time each step/request can take to the Ollama API before being closed.

Known Issues\Limitations

• Some settings might be better in per prompt config. Settings such as api_base_url and timeout might be better served in each prompt's config.json. This would allow for more flexibility. Current approach works but might be too simple for some.
• <think></think> tags and enclosed content are automatically removed when passing prompt. Might be better to make this an option rather than force removal for each prompt.
• Support for limited model options of num_ctx, repeat_last_n, repeat_penalty, and temperature in config.json as it appears not all models support all options.

Notes

• Running tests on MacBook Pro M3 with 12/18 cores and 32GB RAM
• Using Ollama latest
• llama3.2:latest: ~50 tokens per second
• deepseek-r1:8b: ~24 tokens per second
• deepseek-r1:14b: ~12 tokens per second

More Options

• If you need broader support for more than local Ollama, and great community prompts: Fabric AI https://github.com/danielmiessler/fabric
• Chatbox is a pretty cool UI that allows simple model config tweaks and custom system prompts. https://github.com/Bin-Huang/chatbox
• Open WebUI is a powerful self hosted UI that can do everything goll does plus much more. https://github.com/open-webui/open-webui