TypedAI

TypedAi is a Python library that simplifies the OpenAI chat completions api using type hints.

Think Typer / FastApi for LLM interactions.

The library has no dependencies other than openai and pydantic.

pip install eidolon-typedai

Easily Define Output Format

Defining json schema by hand error-prone. With TypedAI, you can define the output format with python types that are transformed to JsonSchema.

TypedAI then stores those types with the response to be used later parsing.

from typedai import TypedAI
from pydantic import BaseModel

class Response(BaseModel):
    philosopher: str
    meaning_of_life: str

typed_completion = TypedAI().completions.create(  # returns TypedChatCompletion[Response]
    model="gpt-3.5-turbo",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "What is the meaning of life according to Douglas Adams?"},
    ],
    response_type=Response
)
print(typed_completion.parse_content().meaning_of_life)  # 42

IDE Type Hinting

Sure you can code without them. You can live on rice and beans too, but you sure don't want to.

Define tool calls with function signatures

from typedai import TypedAI
from typing import Literal

def get_meaning_of_life(philosopher: Literal["Douglas Adams"]) -> str:
    """Call this tool to get the meaning of life according to a philosopher."""
    if philosopher == "Douglas Adams":
        return "42"

completion_with_tool_calls = TypedAI().completions.create(
    model="gpt-3.5-turbo",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "What is the meaning of life according to Douglas Adams?"},
    ],
    fn_tools=[get_meaning_of_life],
)

Easily Construct Response Messages

When a llm completion has tool calls, you normally need to parse the response and execute the function.

Since you defined your tools with functions, TypedAI can execute them and build the response messages for you.

from typedai import TypedAI
from above_example import get_meaning_of_life

messages=[
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "What is the meaning of life according to Douglas Adams?"},
]
completion_with_tool_calls = TypedAI().completions.create(
    model="gpt-3.5-turbo", messages=messages, fn_tools=[get_meaning_of_life]
)
if completion_with_tool_calls.has_tool_calls():
    messages.extend(completion_with_tool_calls.build_messages())
    completion_with_tool_calls = TypedAI().completions.create(
        model="gpt-3.5-turbo", messages=messages, fn_tools=[get_meaning_of_life]
    )

Full Streaming Support (even with typed responses)

And this all works great with streaming completions too!

from typedai import TypedAI
from above_example import Response


typed_stream = TypedAI().completions.stream(
    model="gpt-3.5-turbo",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "What is the meaning of life according to Douglas Adams?"},
    ],
    response_type=Response,
)

for chunk in typed_stream:
    print(chunk)  # iterates ChatCompletionChunk just like you would expect

typed_stream.completion()  # aggregated TypedCompletion[Response] (with type hints!)

Purely Additive

TypedAI is purely additive. You can use it as much or as little as you want. It doesn't change the way you use OpenAI's API, it just makes it easier.

The objects used are either the raw openai objects or a minimally altered TypedAI child object with some additional functionality. Even dumping the models will give you unchanged results.

Similarly, no parsing or validation is done without explicit method calls, so the additional functionality will never slow you down.