sibila 0.4.5

Structured queries from local or online LLM models

Sibila

Extract structured data from remote or local LLM models. Predictable output is important for serious use of LLMs.

• Query structured data into Pydantic objects, dataclasses or simple types.
• Access remote models from OpenAI, Anthropic, Mistral AI and other providers.
• Use vision models like GPT-4o, to extract structured data from images.
• Run local models like Llama-3, Phi-3, OpenChat or any other GGUF file model.
• Sibila is also a general purpose model access library, to generate plain text or free JSON results, with the same API for local and remote models.

No matter how well you craft a prompt begging a model for the format you need, it can always respond something else. Extracting structured data can be a big step into getting predictable behavior from your models.

See What can you do with Sibila?

Structured data

To extract structured data, using a local model:

from sibila import Models
from pydantic import BaseModel

class Info(BaseModel):
    event_year: int
    first_name: str
    last_name: str
    age_at_the_time: int
    nationality: str

model = Models.create("llamacpp:openchat")

model.extract(Info, "Who was the first man in the moon?")

Returns an instance of class Info, created from the model's output:

Info(event_year=1969,
     first_name='Neil',
     last_name='Armstrong',
     age_at_the_time=38,
     nationality='American')

Or to use a remote model like OpenAI's GPT-4, we would simply replace the model's name:

model = Models.create("openai:gpt-4")

model.extract(Info, "Who was the first man in the moon?")

If Pydantic BaseModel objects are too much for your project, Sibila supports similar functionality with Python dataclasses. Also includes asynchronous access to remote models.

Vision models

Sibila supports image input, alongside text prompts. For example, to extract the fields from a receipt in a photo:

from pydantic import Field

model = Models.create("openai:gpt-4o")

class ReceiptLine(BaseModel):
    """Receipt line data"""
    description: str
    cost: float

class Receipt(BaseModel):
    """Receipt information"""
    total: float = Field(description="Total value")
    lines: list[ReceiptLine] = Field(description="List of lines of paid items")

info = model.extract(Receipt,
                     ("Extract receipt information.", 
                      "https://upload.wikimedia.org/wikipedia/commons/6/6a/Receipts_in_Italy_13.jpg"))
info

Returns receipt fields structured in a Pydantic object:

Receipt(total=5.88, 
        lines=[ReceiptLine(description='BIS BORSE TERM.S', cost=3.9), 
               ReceiptLine(description='GHIACCIO 2X400 G', cost=0.99),
               ReceiptLine(description='GHIACCIO 2X400 G', cost=0.99)])

Another example - extracting the most import elements in a photo:

photo = "https://upload.wikimedia.org/wikipedia/commons/thumb/3/32/Hohenloher_Freilandmuseum_-_Baugruppe_Hohenloher_Dorf_-_Bauerngarten_-_Ansicht_von_Osten_im_Juni.jpg/640px-Hohenloher_Freilandmuseum_-_Baugruppe_Hohenloher_Dorf_-_Bauerngarten_-_Ansicht_von_Osten_im_Juni.jpg"

model.extract(list[str],
              ("Extract up to five of the most important elements in this photo.",
              photo))

Returns a list with the five strings:

['House with red roof and beige walls',
 'Large tree with green leaves',
 'Garden with various plants and flowers',
 'Clear blue sky',
 'Wooden fence']

Local vision models based on llama.cpp/llava can also be used.

⭐ Like our work? Give us a star!

Docs

The docs explain the main concepts, include examples and an API reference.

Installation

Sibila can be installed from PyPI by doing:

pip install -U sibila

See Getting started for more information.

Examples

The Examples show what you can do with local or remote models in Sibila: structured data extraction, classification, summarization, etc.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgements

Sibila wouldn't be be possible without the help of great software and people:

• llama.cpp
• llama-cpp-python
• Hugging Face model hub and TheBloke (Tom Jobbins)

Thank you!

Sibila?

Sibila is the Portuguese word for Sibyl. The Sibyls were wise oracular women in ancient Greece. Their mysterious words puzzled people throughout the centuries, providing insight or prophetic predictions, "uttering things not to be laughed at".

Michelangelo's Delphic Sibyl, in the Sistine Chapel ceiling.