context-window-manager 0.1.0

Production-ready LLM context window optimization and management

Context Window Manager

Production-ready LLM context window optimization and management. Automatically handles token counting, message pruning, and content compression to keep conversations within context limits.

Features

• Token Counting: Accurate token counting with tiktoken or approximate fallback
• Automatic Pruning: Multiple strategies (FIFO, relevance, recency-weighted, etc.)
• Content Compression: Truncation, summarization, bullet point extraction
• Priority System: Pin important messages, set priorities for retention
• Model Presets: Built-in configs for GPT-4, GPT-4o, GPT-3.5, Claude
• Usage Statistics: Track token usage and pruning operations
• Zero Dependencies Core: Works without tiktoken (with approximation)

Installation

pip install context-window-manager           # Core (approximate counting)
pip install context-window-manager[tiktoken] # Accurate token counting

Quick Start

Basic Usage

from context_window_manager import ContextWindowManager, create_manager

# Create manager for GPT-4
manager = create_manager("gpt-4")

# Set system message
manager.set_system_message("You are a helpful assistant.")

# Add messages
manager.add_message("user", "Hello, how are you?")
manager.add_message("assistant", "I'm doing well, thank you!")
manager.add_message("user", "Tell me about Python.")

# Get messages for API call
messages = manager.get_messages()

# Check budget
budget = manager.get_budget()
print(f"Used: {budget.used_tokens}/{budget.total_tokens}")

Automatic Pruning

from context_window_manager import ContextWindowManager, WindowConfig, PruningStrategy

config = WindowConfig(
    max_tokens=16000,
    pruning_strategy=PruningStrategy.RECENCY_WEIGHTED,
    pruning_threshold=0.85,  # Start pruning at 85% utilization
    preserve_recent_turns=2   # Always keep last 2 exchanges
)

manager = ContextWindowManager(config)

# Messages are automatically pruned when threshold is reached
for i in range(100):
    manager.add_message("user", f"Message {i}: " + "x" * 500)
    manager.add_message("assistant", f"Response {i}: " + "y" * 500)

print(f"Messages: {manager.conversation.message_count}")
print(f"Utilization: {manager.conversation.utilization:.1%}")

Manual Pruning

from context_window_manager import PruningStrategy

# Prune to specific token count
result = manager.prune(target_tokens=8000)
print(f"Removed {result.removed_messages} messages")
print(f"Saved {result.tokens_saved} tokens")

# Prune with specific strategy
result = manager.prune(strategy=PruningStrategy.RELEVANCE)

Message Priorities

from context_window_manager import Priority

# Add important message
manager.add_message("user", "Critical instruction", priority=Priority.CRITICAL)

# Pin a message (never pruned)
manager.pin_message(0)

# Set priority after creation
manager.set_priority(1, Priority.HIGH)

Content Compression

from context_window_manager import CompressionMethod

# Compress a specific message
result = manager.compress_message(
    message_index=5,
    target_tokens=100,
    method=CompressionMethod.BULLET_POINTS
)

print(f"Compressed from {result.original_tokens} to {result.compressed_tokens}")

Conversation Buffer

from context_window_manager import ConversationBuffer

# Simple buffer with limits
buffer = ConversationBuffer(
    max_tokens=8000,
    max_messages=50
)

buffer.add("user", "Hello")
buffer.add("assistant", "Hi there!")

messages = buffer.get_messages()
print(f"Tokens: {buffer.token_count}")

Model Configurations

from context_window_manager import ModelConfig, ContextWindowManager

# Use preset
config = ModelConfig.gpt4o()
manager = ContextWindowManager(model_config=config)

# Or create custom
custom_config = ModelConfig(
    name="custom-model",
    max_context_tokens=32000,
    max_output_tokens=4096,
    tokenizer=TokenizerType.TIKTOKEN_CL100K
)

Pruning Strategies

Strategy	Description
FIFO	Remove oldest messages first
LIFO	Remove newest (except recent turns)
SLIDING_WINDOW	Keep only most recent N messages
RELEVANCE	Remove by relevance score
IMPORTANCE	Remove by importance score
RECENCY_WEIGHTED	Combine recency + relevance

Compression Methods

Method	Description
TRUNCATE	Cut content at sentence boundary
BULLET_POINTS	Extract key sentences as bullets
EXTRACT_KEY_INFO	Keep first/last paragraphs

API Reference

ContextWindowManager

manager = ContextWindowManager(config, model_config)

# Add messages
manager.add_message(role, content, **kwargs)
manager.set_system_message(content)

# Get messages
messages = manager.get_messages()

# Budget management
budget = manager.get_budget()
snapshot = manager.get_snapshot()

# Pruning
result = manager.prune(target_tokens, strategy)

# Compression
result = manager.compress_message(index, target_tokens, method)

# Message management
manager.pin_message(index)
manager.set_priority(index, priority)
manager.clear()

# Utilities
fits = manager.fits(content)
tokens = manager.tokens_for(content)

WindowConfig

config = WindowConfig(
    max_tokens=128000,
    reserved_output_tokens=4096,
    max_history_ratio=0.7,
    pruning_strategy=PruningStrategy.RECENCY_WEIGHTED,
    compression_method=CompressionMethod.NONE,
    tokenizer_type=TokenizerType.TIKTOKEN_CL100K,
    min_messages_to_keep=2,
    always_keep_system=True,
    preserve_recent_turns=2,
    pruning_threshold=0.85,
    on_prune=callback_function
)

License

MIT License - Pranay M