A CLI tool to manage Chat Conversations w/ Markdown instead of a WebGUI.
> python chat.py --help
Usage: chat.py [OPTIONS]
This script provides a chatbot interface using the OpenAI API. You can select different models and modify chat options for customized responses. The chatbot will read from stdin and write to stdout. Useful for chaining prompts.
Options:
-h, --help Show this help message and exit.
--model=MODEL_ID Select the AI model. Available models: gp4, gpt3.5
(Default: gpt3.5)
--mode=RESPONSE_MODE Select how the model generates a response.
Available modes: ss, cot, tot
(Default: ss)
--personality=PERSONALITY_ID Select a personality preset for AI chat responses.
Available presets: creative, balanced, reserved.
(Default: balanced)
--temperature=FLOAT_VALUE Set a custom temperature value for the AI chat responses.
(Overwrites the personality preset temperature value.)
--tokens=INT_VALUE Set a custom maximum token count for the AI chat responses.
(Overwrites the personality preset token value.)
--topp=FLOAT_VALUE Set a custom top_p value for the AI chat responses.
(Overwrites the personality preset top_p value.)
--verbose Enable verbose logging.
--debug Enable debug logging.
--trace Enable trace logging.
--quiet Disable all logging.
Environment Variables:
OPENAI_API_KEY Set the OpenAI API key for authentication.
OPENAI_ORGANIZATION Set the OpenAI organization for API access.
Examples:
1. Using the script with default options and personality preset:
cat ./chat.md | python chat.py --model=gpt3 --personality=balanced > ./completed_chat.md
2. Customizing chat options for a more creative response:
cat ./chat.md | python chat.py --model=gpt3 --temperature=1.5 --topp=0.9 --tokens=4096 > ./completed_chat.md
Response Modes:
ss - Single Shot
This is the standard response mode & generally the fastest. The LLM will generate a single, direct response to the input.
Use Single Shot when you want a quick response, have a simple prompt or are exploring a topic or thought process.
cot - Chain of Thought
Take more time & generate a thoughtful response. CoT is a framework that enhances an LLM's problem-solving capabilities by employing an ordered sequence of reasoning steps that collectively lead to a (more) comprehensive solution. For more information see https://arxiv.org/pdf/2201.11903.pdf
Use Chain of Thought when you want a more comprehensive response, have a complex prompt or have a problem statement that is well defined, well scoped & has concrete actionables/questions. CoT is best applied after narrowing down the scope using Single Shot.
tot - Tree of Thoughts
NOT YET IMPLEMENTED
Take alot of time & search the entire solution space for a more optimal response. [ADD MORE DESCRIPTION HERE]
Use Tree of Thoughts when [TBD]Note: This is currently a WIP. Formatting is a bit rigid.
- Wrap Messages w/ a
<!-- start { "role": "OpenAI Role" } -->&<!-- stop -->HTML Comment- Pad your messages with a newline between the comments
- Add metadata to the JSON object in the
startcomment. At a minimum include the OpenAI Role (system,assistantoruser).- The tool will include a
modelkey in the metadata ofassistantmessages: ex<!-- start {"role": "assistant", "model": "gpt-3.5-turbo-0613"} -->
- The tool will include a
- Seperate Messages with a line break:
---- Pad line breaks with a newline before & after them
Syntax Example:
<!-- start {"role": "system"} -->
You are a highly experienced marine biologist. You are participating as part of a QA. Answer questions precisely & accurately. Keep your prose concise & your language simple.
<!-- end -->
---
<!-- start {"role": "user"} -->
Can you tell me about Dolphins?
<!-- end -->