smartcat (sc)

Puts a brain behind cat! CLI interface to bring language models in the Unix ecosystem and allow terminal power users to make the most out of llms.

What makes it special:

• made for power users, tailor the config to reduce overhead on your most frequent tasks;
• minimalist, built according the unix philosophy with terminal and editor intergation in mind;
• good io handling to insert user input in prompts and use the result in cli-based workflows;
• built-in partial prompt to make the model play nice as a cli tool;
• full configurability on which API, LLM version and temperature you use;
• write and save your own prompt templates for faster reccuring tasks (simplify, optimize, tests, etc);
• conversation support;
• glob expressions to include context files.

Currently supports the following APIs:

• Local runs with Ollama or any server compliant with its format, see the Ollama setup section for the free and easiest way to get started!
  Answers might be slow depending on your setup, you may want to try the third party APIs for an optimal workflow.
• OpenAi, Mistral AI, Anthropic, Groq.

Table of Contents

• Installation
• Usage
• A few examples to get started 🐈‍⬛
  • Integrating with editors
    • Example workflows
• Configuration ← please read this carefully
  • Ollama setup ← easiest way to get running for free
• Developping

Installation

On the first run (sc), it will ask you to generate some default configuration files and give pointers on how to finalize the install (see the configuration section).

The minimum config requirement is a default prompt calling a setup API (either remote with api key or local with ollama).

Now on how to get it.

With Cargo

With an up to date rust and cargo setup (you might consider running rustup update):

cargo install smartcat

run this command again to update smartcat.

By downloading the binary

Chose the one compiled for your platform on the release page.

Usage

Usage: sc [OPTIONS] [INPUT_OR_CONFIG_REF] [INPUT_IF_CONFIG_REF]

Arguments:
  [INPUT_OR_CONFIG_REF]  ref to a prompt from config or straight input (will use `default` prompt template)
  [INPUT_IF_CONFIG_REF]  if the first arg matches a config ref, the second will be used as input

Options:
  -e, --extend-conversation        whether to extend the previous conversation or start a new one
  -r, --repeat-input               whether to repeat the input before the output, useful to extend instead of replacing
      --api <API>                  overrides which api to hit [possible values: openai, mistral, groq, anthropic, ollama]
  -m, --model <MODEL>              overrides which model (of the api) to use
  -t, --temperature <TEMPERATURE>  temperature higher means answer further from the average
  -l, --char-limit <CHAR_LIMIT>    max number of chars to include, ask for user approval if more, 0 = no limit
  -c, --context <CONTEXT>...       glob patterns or list of files to use the content as context
                                   make sure it's the last arg.
  -h, --help                       Print help
  -V, --version                    Print version

You can use it to accomplish tasks in the CLI but also in your editors (if they are good unix citizens, i.e. work with shell commands and text streams) to complete, refactor, write tests... anything!

The key to make this work seamlessly is a good default prompt that tells the model to behave like a CLI tool an not write any unwanted text like markdown formatting or explanations.

A few examples to get started 🐈‍⬛

sc "say hi"  # just ask

sc test                         # use templated prompts
sc test "and parametrize them"  # extend them on the fly

sc "explain how to use this program" -c **/*.md main.py  # use files as context

git diff | sc "summarize the changes"  # pipe data in

cat en.md | sc "translate in french" >> fr.md   # write data out
sc -e "use a more informal tone" -t 2 >> fr.md  # extend the conversation and raise the temprature

Integrating with editors

The key for a good integration in editors is a good default prompt (or set of) combined with the -p flag for precising the task at hand. The -r flag can be used to decide whether to replace or extend the selection.

Vim

Start by selecting some text, then press :. You can then pipe the selection content to smartcat.

:'<,'>!sc "replace the versions with wildcards"

:'<,'>!sc "fix this function"

will replace the current selection with the same text transformed by the language model.

:'<,'>!sc -r write_test

will append at the end of the current selection the result of the language model.

Add the following remap to your vimrc for easy access:

nnoremap <leader>sc :'<,'>!sc

Helix and Kakoune

Same concept, different shortcut, simply press the pipe key to redirect the selection to smarcat.

pipe:sc write_test -r

With some remapping you may have your most reccurrent action attached to few keystrokes e.g. <leader>wt!

Example Workflows

For quick questions:

sc "my quick question"

which will likely be your fastest path to answer: a shortcut to open your terminal (if you're not in it already), sc and you're set. No tab finding, no logins, no redirects etc.

To enhance coding:

select a struct

:'<,'>!sc "implement the traits FromStr and ToString for this struct"

select the generated impl block

:'<,'>!sc -e "can you make it more concise?"

put the cursor at the bottom of the file and give example usage as input

:'<,'>!sc -e "now write tests for it knowing it's used like this" -c src/main.rs

...

To have a full conversation with a llm from a markdown file:

vim problem_solving.md

> write your question as comment in the markdown file then select your question
> and send it to smartcat using the aforementioned trick, use `-r` to repeat the input.

If you wan to continue the conversation, write your new question as a comment and repeat
the previous step with `-e -r`.

> This allows you to keep track of your questions and make a nice reusable document.

Configuration

• by default lives at $HOME/.config/smartcat
• the directory can be set using the SMARTCAT_CONFIG_PATH environement variable
• use #[<input>] as the placeholder for input when writing prompts
• the default model is a local phi3 ran with ollama but I recommend trying the latest ones and see which one works best for you;
• the prompt named default will be the one used by default.
• you can play with the temperature and set a default for each prompt depending on its use case;

Three files are used:

• .api_configs.toml stores your credentials, you need at least one provider with API with key or a local ollama setup;
• prompts.toml stores you prompt templates, you need at least the default prompt;
• conversation.toml stores the latest chat if you need to continue it, it's automanaged but you can make backups if you want.

.api_configs.toml

[ollama]  # local API, no key required
url = "http://localhost:11434/api/chat"
default_model = "phi3"

[openai]  # each supported api has their own config section with api and url
api_key = "<your_api_key>"
default_model = "gpt-4-turbo-preview"
url = "https://api.openai.com/v1/chat/completions"

[mistral]
api_key_command = "pass mistral/api_key"  # you can use a command to grab the key
default_model = "mistral-medium"
url = "https://api.mistral.ai/v1/chat/completions"

[groq]
api_key_command = "pass groq/api_key"
default_model = "llama3-70b-8192"
url = "https://api.groq.com/openai/v1/chat/completions"

[anthropic]
api_key = "<yet_another_api_key>"
url = "https://api.anthropic.com/v1/messages"
default_model = "claude-3-opus-20240229"
version = "2023-06-01"

prompts.toml

[default]  # a prompt is a section
api = "ollama"  # must refer to an entry in the `.api_configs.toml` file
model = "phi3"  # each prompt may define its own model

[[default.messages]]  # then you can list messages
role = "system"
content = """\
You are an extremely skilled programmer with a keen eye for detail and an emphasis on readable code. \
You have been tasked with acting as a smart version of the cat unix program. You take text and a prompt in and write text out. \
For that reason, it is of crucial importance to just write the desired output. Do not under any circumstance write any comment or thought \
as you output will be piped into other programs. Do not write the markdown delimiters for code as well. \
Sometimes you will be asked to implement or extend some input code. Same thing goes here, write only what was asked because what you write will \
be directly added to the user's editor. \
Never ever write ``` around the code. \
"""

[empty]  # always nice to have an empty prompt available
api = "openai"
# not mentioning the model will use the default from the api config
messages = []

[write_tests]
api = "anthropic"
temperature = 0.0

[[write_tests.messages]]
role = "system"
content = """\
You are an extremely skilled programmer with a keen eye for detail and an emphasis on readable code. \
You have been tasked with acting as a smart version of the cat unix program. You take text and a prompt in and write text out. \
For that reason, it is of crucial importance to just write the desired output. Do not under any circumstance write any comment or thought \
as you output will be piped into other programs. Do not write the markdown delimiters for code as well. \
Sometimes you will be asked to implement or extend some input code. Same thing goes here, write only what was asked because what you write will \
be directly added to the user's editor. \
Never ever write ``` around the code. \
"""

[[write_tests.messages]]
role = "user"
# the following placeholder string #[<input>] will be replaced by the input
# each message seeks it and replaces it
content ='''Write tests using pytest for the following code. Parametrize it if appropriate.

#[<input>]
'''

see the config setup file for more details.

Ollama setup

1. Install Ollama
2. Pull the model you plan on using ollama pull phi3
3. Test the model ollama run phi3 "say hi"
4. Make sure the serving is available curl http://localhost:11434 which should say "Ollama is running", else you might need to run ollama serve
5. smartcat will now be able to reach your local ollama, enjoy!

⚠️ Answers might be slow depending on your setup, you may want to try the third party APIs for an optimal workflow.

Developping

Some tests rely on environement variables and don't behave well with multi-threading. They are marked with #[serial] from the serial_test crate.

State of the project

Smartcat has reached an acceptable feature set. The focus is now on upgrading the codebase quality as I hadn't really touched rust since 2019 and it shows.

TODO

• make it available on homebrew
• handle streams
• automagical context fetches (might be out of scope)
• add RAG capabilities