functualize-ai-pydantic

Status: Published — Independently installable from PyPI.

PydanticAI-backed AI implementation plugin for functualize. Bridges the PydanticAI agent framework and LiteLLM model routing into functualize's AI capability system, providing tool calling, streaming, structured output, multi-turn agent loops, and conversation history continuation — all wired automatically via entry-point discovery.

Installation

pip install functualize-ai-pydantic

Quick Start

Once installed, the plugin registers itself via the functualize.ai_providers entry point. Configure it in your project config:

[ai]
provider = "pydantic"
model = "claude-sonnet-4-20250514"

Then use the AI capability in any functualize job:

from functualize_ai import AI, AIConfig
from functualize_ai_pydantic import PydanticAIProvider

config = AIConfig(model="claude-sonnet-4-20250514", max_tokens=1024)
provider = PydanticAIProvider(config)

result = provider.run("Summarize this document in 3 bullet points.")
print(result.output)
print(f"Tokens used: {result.usage.total_tokens}")

Features

• Tool calling — Translates functualize ToolDef instances into PydanticAI's native Tool format via ToolScopeTranslator, enabling LLM-driven function invocation
• Structured output — Pass a Pydantic model as response_model to get typed, validated responses directly from the LLM
• Streaming — Iterate over response tokens as they arrive using PydanticAIProvider.stream()
• Multi-turn agent loops — PydanticAI.run_agent_loop() executes iterative tool-calling workflows with configurable iteration limits
• Conversation history — PydanticAI.run_with_history() continues conversations from prior message history
• Auto-registration — Plugin hooks into APP_READY to resolve AIConfig, create the provider, and register both AI and PydanticAI types with DI
• Budget tracking — Integrates with functualize's state system for token usage and cost tracking across calls

API Reference

Public classes exported by this plugin:

• PydanticAIPlugin — Entry-point plugin class that registers the provider with functualize's DI system at boot. Implements the plugin callable protocol.
• PydanticAIProvider — AIProvider protocol implementation backed by PydanticAI and LiteLLM. Methods: complete(), run(), stream(), extract().
• PydanticAI — Extended AI capability subclass with PydanticAI-specific methods: run_agent_loop() for multi-turn execution and run_with_history() for conversation continuation.
• ToolScopeTranslator — Converts a list of functualize ToolDef instances into PydanticAI Tool objects using Tool.from_schema().

Development

Run plugin tests:

uv run pytest plugins/functualize-ai-pydantic/tests/ -v

Lint and format:

uv run ruff check plugins/functualize-ai-pydantic/
uv run ruff format plugins/functualize-ai-pydantic/