API Reference¶

Complete API documentation for summarization-pydantic-ai.

Quick Reference¶

Python

from pydantic_ai_summarization import (
    # Main exports - Summarization
    SummarizationProcessor,
    create_summarization_processor,

    # Main exports - Sliding Window
    SlidingWindowProcessor,
    create_sliding_window_processor,

    # Main exports - Context Manager Middleware (requires [hybrid] extra)
    ContextManagerMiddleware,
    UsageCallback,
    create_context_manager_middleware,

    # Utilities
    count_tokens_approximately,
    format_messages_for_summary,

    # Types
    ContextSize,
    ContextFraction,
    ContextTokens,
    ContextMessages,
    TokenCounter,

    # Constants
    DEFAULT_SUMMARY_PROMPT,
)

Modules¶

Module	Description
Processor	`SummarizationProcessor` class and factory function
Sliding Window	`SlidingWindowProcessor` class and factory function
Middleware	`ContextManagerMiddleware` class and factory function (requires `[hybrid]` extra)
Types	Type definitions and aliases

Processors Comparison¶

Feature	SummarizationProcessor	SlidingWindowProcessor	ContextManagerMiddleware
LLM Cost	High	Zero	Per compression
Latency	High	~0ms	Low (tracking) / High (compression)
Context Preservation	Intelligent	None	Intelligent
Token Tracking	No	No	Built-in
Usage Callbacks	No	No	Yes
Tool Output Truncation	No	No	Yes
`model` param	Required	Not needed	Optional (for summaries)
Requires extra	No	No	`[hybrid]`

Quick Examples¶

Summarization¶

Python

from pydantic_ai import Agent
from pydantic_ai_summarization import create_summarization_processor

processor = create_summarization_processor(
    trigger=("tokens", 100000),
    keep=("messages", 20),
)

agent = Agent(
    "openai:gpt-4.1",
    history_processors=[processor],
)

Sliding Window¶

Python

from pydantic_ai import Agent
from pydantic_ai_summarization import create_sliding_window_processor

processor = create_sliding_window_processor(
    trigger=("messages", 100),
    keep=("messages", 50),
)

agent = Agent(
    "openai:gpt-4.1",
    history_processors=[processor],
)