Eliza-GPT

Eliza, the classic chatbot from the 1960s, running on OpenAI's Chat Completions API.

Why?

The main reason is that this seemed like a fun project to help me acquire a greater understanding of the Chat Completions API from OpenAI, which has become a sort of standard for turn-based conversation between a human and a machine.

But also, it is great to have a fast and free service with negligible resource consumption that can be used as a stand-in for a real LLM when the quality of the responses do not matter, such as while doing development or testing.

How To Use

Eliza-GPT is written in Python and can be installed with pip:

$ pip install eliza-gpt

Once installed, run it to start it as a local service on your computer:

$ eliza-gpt
Eliza-GPT is running!
Set base_url="http://127.0.0.1:5005/v1" in your OpenAI client to connect.

Finally, if you have a custom client that talks directly to the chat completions endpoint, configure http://127.0.0.1:5005/v1 as your endpoint.

Configuration

Run with --help to learn about configuration options, which include:

• Setting the listening IP address and port
• Adding an API key for authentication
• Changing the (simulated) response times, or removing all artificial delays.

Connecting With a Chat Client

If you are using the official Python client from OpenAI, add the base_url option to it as follows:

openai_client = OpenAI(base_url='http://127.0.0.1:5005/v1')

If you use Langchain, then use the following to connect to Eliza-GPT:

chat = ChatOpenAI(base_url='http://127.0.0.1:5005/v1')

Eliza-GPT supports both the direct and streaming interfaces offered by these clients.

Examples

The examples directory includes demonstration applications implemented with the OpenAI client and Langchain.

Implementation Details

Eliza-GPT implements a portion of the OpenAI Chat Completions API, ignoring anything that isn't useful.

The following aspects of the API are currently supported:

• The chat history given in the messages argument. Only messages with role user are used to "prime" Eliza so that it provides a reasonable response.
• The stream argument, which determines if the response should be given in a single JSON object or as a stream of Server-Sent Events.
• The seed option, which makes it possible to have deterministic responses.

These are not used or implemented:

• Any messages in the chat history that have a role different than user.
• Model names. Eliza-GPT returns the requested model name in the response.
• The n argument, which controls how many responses are returned. Eliza-GPT always returns a single response.
• The max_tokens argument.
• Temperature and other LLM-specific tuning parameters.
• The stop argument.
• Anything related to tools or functions that the model may call.
• Anything related to log probabilities.