Release Notes - v0.2.3

Release Notes - v0.2.3

🎉 New Feature: Structured Output

What's New

Conversation now supports structured outputs via the returns parameter. Get validated, type-safe data directly from LLM responses using Pydantic models.

Usage

from pydantic import BaseModel, Field
import chak

class User(BaseModel):
    name: str = Field(description="User's full name")
    email: str = Field(description="User's email address")
    age: int = Field(description="User's age")

conv = chak.Conversation("openai/gpt-4o", api_key="YOUR_KEY")

# Get structured output automatically
user = await conv.asend(
    "Create a user: John Doe, john@example.com, 30 years old",
    returns=User
)

print(user.name)   # "John Doe"
print(user.email)  # "john@example.com"
print(user.age)    # 30

Key Features

• Type-Safe: Automatic validation using Pydantic models
• Multimodal Support: Works with images, documents, and other attachments
• Seamless Integration: No manual parsing, just specify the model
• Provider Compatible: Works with any LLM that supports function calling

Multimodal Example

from chak import Image

class SceneDescription(BaseModel):
    main_subject: str
    colors: list[str]
    mood: str

# Extract structured data from images
scene = await conv.asend(
    "Analyze this image and describe the scene",
    attachments=[Image("photo.jpg")],
    returns=SceneDescription
)

Examples

• Basic Structured Output: examples/structured_output_simple.py
• Multimodal Structured Output: examples/structured_output_multimodal.py

Technical Details

• API: New returns parameter in asend() method
• Implementation: Uses function calling under the hood
• Validation: Automatic Pydantic validation
• Context Management: Fully integrated with conversation history

Notes

• Requires Pydantic BaseModel subclass
• Currently async only (asend(), not send())
• Some vision models may not support function calling
• Best results with models like OpenAI GPT-4o, Anthropic Claude 3

---

Full documentation: README.md