search

minimaxir / simpleaichat

Python package for easily interfacing with chat apps, with robust features and minimal code complexity.
MIT License
3.43k stars 224 forks source link
ai chatgpt

readme

simpleaichat

from simpleaichat import AIChat

ai = AIChat(system="Write a fancy GitHub README based on the user-provided project name.")
ai("simpleaichat")

simpleaichat is a Python package for easily interfacing with chat apps like ChatGPT and GPT-4 with robust features and minimal code complexity. This tool has many features optimized for working with ChatGPT as fast and as cheap as possible, but still much more capable of modern AI tricks than most implementations:

Here's some fun, hackable examples on how simpleaichat works:

Installation

simpleaichat can be installed from PyPI:

pip3 install simpleaichat

Quick, Fun Demo

You can demo chat-apps very quickly with simpleaichat! First, you will need to get an OpenAI API key, and then with one line of code:

from simpleaichat import AIChat

AIChat(api_key="sk-...")

And with that, you'll be thrust directly into an interactive chat!

This AI chat will mimic the behavior of OpenAI's webapp, but on your local computer!

You can also pass the API key by storing it in an .env file with a OPENAI_API_KEY field in the working directory (recommended), or by setting the environment variable of OPENAI_API_KEY directly to the API key.

But what about creating your own custom conversations? That's where things get fun. Just input whatever person, place or thing, fictional or nonfictional, that you want to chat with!

AIChat("GLaDOS")  # assuming API key loaded via methods above

But that's not all! You can customize exactly how they behave too with additional commands!

AIChat("GLaDOS", "Speak in the style of a Seinfeld monologue")

AIChat("Ronald McDonald", "Speak using only emoji")

Need some socialization immediately? Once simpleaichat is installed, you can also start these chats directly from the command line!

simpleaichat
simpleaichat "GlaDOS"
simpleaichat "GLaDOS" "Speak in the style of a Seinfeld monologue"

Building AI-based Apps

The trick with working with new chat-based apps that wasn't readily available with earlier iterations of GPT-3 is the addition of the system prompt: a different class of prompt that guides the AI behavior throughout the entire conversation. In fact, the chat demos above are actually using system prompt tricks behind the scenes! OpenAI has also released an official guide for system prompt best practices to building AI apps.

For developers, you can instantiate a programmatic instance of AIChat by explicitly specifying a system prompt, or by disabling the console.

ai = AIChat(system="You are a helpful assistant.")
ai = AIChat(console=False)  # same as above

You can also pass in a model parameter, such as model="gpt-4" if you have access to GPT-4, or model="gpt-3.5-turbo-16k" for a larger-context-window ChatGPT.

You can then feed the new ai class with user input, and it will return and save the response from ChatGPT:

response = ai("What is the capital of California?")
print(response)
The capital of California is Sacramento.

Alternatively, you can stream responses by token with a generator if the text generation itself is too slow:

for chunk in ai.stream("What is the capital of California?", params={"max_tokens": 5}):
    response_td = chunk["response"]  # dict contains "delta" for the new token and "response"
    print(response_td)
The
The capital
The capital of
The capital of California
The capital of California is

Further calls to the ai object will continue the chat, automatically incorporating previous information from the conversation.

response = ai("When was it founded?")
print(response)
Sacramento was founded on February 27, 1850.

You can also save chat sessions (as CSV or JSON) and load them later. The API key is not saved so you will have to provide that when loading.

ai.save_session()  # CSV, will only save messages
ai.save_session(format="json", minify=True)  # JSON

ai.load_session("my.csv")
ai.load_session("my.json")

Functions

A large number of popular venture-capital-funded ChatGPT apps don't actually use the "chat" part of the model. Instead, they just use the system prompt/first user prompt as a form of natural language programming. You can emulate this behavior by passing a new system prompt when generating text, and not saving the resulting messages.

The AIChat class is a manager of chat sessions, which means you can have multiple independent chats or functions happening! The examples above use a default session, but you can create new ones by specifying a id when calling ai.

json = '{"title": "An array of integers.", "array": [-1, 0, 1]}'
functions = [
             "Format the user-provided JSON as YAML.",
             "Write a limerick based on the user-provided JSON.",
             "Translate the user-provided JSON from English to French."
            ]
params = {"temperature": 0.0, "max_tokens": 100}  # a temperature of 0.0 is deterministic

# We namespace the function by `id` so it doesn't affect other chats.
# Settings set during session creation will apply to all generations from the session,
# but you can change them per-generation, as is the case with the `system` prompt here.
ai = AIChat(id="function", params=params, save_messages=False)
for function in functions:
    output = ai(json, id="function", system=function)
    print(output)
title: "An array of integers."
array:
  - -1
  - 0
  - 1
An array of integers so neat,
With values that can't be beat,
From negative to positive one,
It's a range that's quite fun,
This JSON is really quite sweet!
{"titre": "Un tableau d'entiers.", "tableau": [-1, 0, 1]}

Newer versions of ChatGPT also support "function calling", but the real benefit of that feature is the ability for ChatGPT to support structured input and/or output, which now opens up a wide variety of applications! simpleaichat streamlines the workflow to allow you to just pass an input_schema and/or an output_schema.

You can construct a schema using a pydantic BaseModel.

from pydantic import BaseModel, Field

ai = AIChat(
    console=False,
    save_messages=False,  # with schema I/O, messages are never saved
    model="gpt-3.5-turbo-0613",
    params={"temperature": 0.0},
)

class get_event_metadata(BaseModel):
    """Event information"""

    description: str = Field(description="Description of event")
    city: str = Field(description="City where event occured")
    year: int = Field(description="Year when event occured")
    month: str = Field(description="Month when event occured")

# returns a dict, with keys ordered as in the schema
ai("First iPhone announcement", output_schema=get_event_metadata)
{'description': 'The first iPhone was announced by Apple Inc.',
 'city': 'San Francisco',
 'year': 2007,
 'month': 'January'}

See the TTRPG Generator Notebook for a more elaborate demonstration of schema capabilities.

Tools

One of the most recent aspects of interacting with ChatGPT is the ability for the model to use "tools." As popularized by LangChain, tools allow the model to decide when to use custom functions, which can extend beyond just the chat AI itself, for example retrieving recent information from the internet not present in the chat AI's training data. This workflow is analogous to ChatGPT Plugins.

Parsing the model output to invoke tools typically requires a number of shennanigans, but simpleaichat uses a neat trick to make it fast and reliable! Additionally, the specified tools return a context for ChatGPT to draw from for its final response, and tools you specify can return a dictionary which you can also populate with arbitrary metadata for debugging and postprocessing. Each generation returns a dictionary with the response and the tool function used, which can be used to set up workflows akin to LangChain-style Agents, e.g. recursively feed input to the model until it determines it does not need to use any more tools.

You will need to specify functions with docstrings which provide hints for the AI to select them:

from simpleaichat.utils import wikipedia_search, wikipedia_search_lookup

# This uses the Wikipedia Search API.
# Results from it are nondeterministic, your mileage will vary.
def search(query):
    """Search the internet."""
    wiki_matches = wikipedia_search(query, n=3)
    return {"context": ", ".join(wiki_matches), "titles": wiki_matches}

def lookup(query):
    """Lookup more information about a topic."""
    page = wikipedia_search_lookup(query, sentences=3)
    return page

params = {"temperature": 0.0, "max_tokens": 100}
ai = AIChat(params=params, console=False)

ai("San Francisco tourist attractions", tools=[search, lookup])
{'context': "Fisherman's Wharf, San Francisco, Tourist attractions in the United States, Lombard Street (San Francisco)",
 'titles': ["Fisherman's Wharf, San Francisco",
  'Tourist attractions in the United States',
  'Lombard Street (San Francisco)'],
 'tool': 'search',
 'response': "There are many popular tourist attractions in San Francisco, including Fisherman's Wharf and Lombard Street. Fisherman's Wharf is a bustling waterfront area known for its seafood restaurants, souvenir shops, and sea lion sightings. Lombard Street, on the other hand, is a famous winding street with eight hairpin turns that attract visitors from all over the world. Both of these attractions are must-sees for anyone visiting San Francisco."}
ai("Lombard Street?", tools=[search, lookup])
{'context': 'Lombard Street is an east–west street in San Francisco, California that is famous for a steep, one-block section with eight hairpin turns. Stretching from The Presidio east to The Embarcadero (with a gap on Telegraph Hill), most of the street\'s western segment is a major thoroughfare designated as part of U.S. Route 101. The famous one-block section, claimed to be "the crookedest street in the world", is located along the eastern segment in the Russian Hill neighborhood.',
 'tool': 'lookup',
 'response': 'Lombard Street is a famous street in San Francisco, California known for its steep, one-block section with eight hairpin turns. It stretches from The Presidio to The Embarcadero, with a gap on Telegraph Hill. The western segment of the street is a major thoroughfare designated as part of U.S. Route 101, while the famous one-block section, claimed to be "the crookedest street in the world", is located along the eastern segment in the Russian Hill'}
ai("Thanks for your help!", tools=[search, lookup])
{'response': "You're welcome! If you have any more questions or need further assistance, feel free to ask.",
 'tool': None}

Miscellaneous Notes

Roadmap

Maintainer/Creator

Max Woolf (@minimaxir)

Max's open-source projects are supported by his Patreon and GitHub Sponsors. If you found this project helpful, any monetary contributions to the Patreon are appreciated and will be put to good creative use.

License

MIT