Delusion

🔮 Convenient generative models pythonic wrappers 🔮

Description

A small toolkit for generative models, with practical conveniences built in: such as caching, fast imports, syntactic sugars, opinionated architecture, curated providers, and more.

• Message classes with type-safe structured output generics, auto validation (chat)
• Modular: Easily write your implementations or modify existing ones.
• Minimal: Avoids the complexity and commitment of a full framework.

Examples

Chat

Chat history, schema-typed messages, auto-validation:

from delusion.chat.option.ollamax import Ollama
from pydantic import BaseModel, Field

class Country(BaseModel):
    name: str
    capital: str
    languages: set[str] = Field(
        description="Officially recognized languages"
    )

chat = Ollama().cache().serve()
chat.gemma4("e2b").pull()
chat.send("Tell me about Canada, its capital and spoken languages.")

# Linters know all fields
canada = chat.generate(schema=Country)

assert (canada.model.name == "Canada")
assert (canada.model.capital == "Ottawa")
assert (canada.model.languages == {"English", "French"})

Audio

Using 0xShug0/audio.cpp, text-to speech with OmniVoice:

from delusion.audio.cpp import AudioCPP, OmniVoice

audio = AudioCPP(
    model=OmniVoice().download(),
)

speak = audio.tts(text="お水はもう一杯もらえますか")
Path("/tmp/output.wav").write_bytes(speak.wav)

Music generation with Stable Audio 3:

from delusion.audio.cpp import AudioCPP, StableAudio3

audio = AudioCPP(
    model=StableAudio3(quant="q8_0").download(),
)

music = audio.music(
    prompt="piano, flute, inspirational, ambient",
    duration=60,
)

Path("/tmp/output.wav").write_bytes(music.wav)

Getting started

Visit the documentation at https://delusion.tremeschin.com/