compress-lightreach 1.0.4

AI cost management SDK with intelligent model routing, prompt compression, and real-time token tracking

Compress Light Reach

AI cost management SDK with intelligent model routing, prompt compression, and real-time token tracking

Compress Light Reach is a Python SDK that provides intelligent model routing and prompt compression for LLM applications, reducing token usage and costs while maintaining quality.

Features

• Intelligent Model Routing: Automatically selects optimal model based on quality requirements (HLE) and available provider keys
• Token-aware Compression: Replaces repeated substrings with shorter placeholders
• Dual Algorithms:
  • Fast greedy (~99% optimal) for daily use
  • Optimal DP (O(n²)) for critical prompts
• Lossless: Perfect decompression guaranteed
• Output Compression: Optional model output compression support
• Cloud API: Uses Light Reach's cloud service for compression and routing
• Multi-provider Support: OpenAI, Anthropic, Google, DeepSeek, Moonshot
• BYOK: Provider API keys managed securely in dashboard (never passed through SDK)

Installation

pip install compress-lightreach

Quick Start

The SDK uses intelligent model routing and targets POST /api/v2/complete.

• Authenticate with your LightReach API key (env var PCOMPRESLR_API_KEY or LIGHTREACH_API_KEY)
• Manage provider keys (OpenAI/Anthropic/Google/etc.) in the dashboard (BYOK)
• System automatically selects optimal model based on your requirements

from pcompresslr import PcompresslrAPIClient

client = PcompresslrAPIClient(api_key="your-lightreach-api-key")

result = client.complete(
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Explain quantum computing in simple terms."},
    ],
    desired_hle=30,  # Quality preference (0-40, where 40 is SOTA)
)

print(result["decompressed_response"])
print(f"Selected: {result['routing_info']['selected_model']}")
print(f"Token savings: {result['compression_stats']['token_savings']}")

With Output Compression

result = client.complete(
    messages=[{"role": "user", "content": "Generate a long report..."}],
    desired_hle=25,
    compress_output=True,
)

print(result["decompressed_response"])

Intelligent Model Routing

The system automatically selects the optimal model based on quality requirements and your available provider keys:

from pcompresslr import PcompresslrAPIClient

client = PcompresslrAPIClient(api_key="your-lightreach-api-key")

# Cross-provider optimization: system picks cheapest model meeting your quality bar
result = client.complete(
    messages=[{"role": "user", "content": "Explain quantum computing"}],
    desired_hle=30,  # Quality preference (0-40, where 40 is SOTA)
)

# Check what was selected
print(result["routing_info"]["selected_model"])           # e.g., "gpt-4o-mini"
print(result["routing_info"]["selected_provider"])        # e.g., "openai"
print(result["routing_info"]["model_hle"])                # e.g., 32.5
print(result["routing_info"]["model_price_per_million"])  # e.g., 0.15

Provider-Constrained Routing

Optionally constrain to a specific provider:

# Only use OpenAI models, but pick the cheapest one meeting HLE 35
result = client.complete(
    messages=[{"role": "user", "content": "Write a poem"}],
    llm_provider="openai",  # Optional: constrain to one provider
    desired_hle=35,
)

HLE Cascading with Admin Controls

Admins can set quality ceilings via the dashboard (global or per-tag) to control costs. Your desired_hle is a preference; if it exceeds an admin-set ceiling, the request will silently clamp to the ceiling and proceed.

from pcompresslr import PcompresslrAPIClient, APIRequestError

client = PcompresslrAPIClient(api_key="your-lightreach-api-key")

# Admin set global HLE ceiling to 30%
# Requesting above the ceiling will be clamped to 30 (no error)
result = client.complete(
    messages=[{"role": "user", "content": "Process payment"}],
    desired_hle=35,  # Will be clamped down to 30
    tags={"env": "production"},
)

# Correct usage: request within ceiling
result = client.complete(
    messages=[{"role": "user", "content": "Process payment"}],
    desired_hle=25,  # OK: below ceiling of 30
    tags={"env": "production"},
)

# Check if your HLE was lowered by an admin ceiling
if result["routing_info"]["hle_clamped"]:
    print(f"HLE lowered from {result['routing_info']['requested_hle']} "
          f"to {result['routing_info']['effective_hle']} "
          f"by {result['routing_info']['hle_source']}-level ceiling")

HLE Ceiling Logic:

• effective_hle = min(desired_hle, tag_hle, global_hle) - most restrictive ceiling wins
• Lower ceiling = force cheaper models (better cost control)
• Requests above the ceiling are clamped down to the ceiling
• Tag-level ceilings can override global ceiling (lowest wins)

Using Message and CompressionConfig Classes

For more structured code, use the Message and CompressionConfig dataclasses:

from pcompresslr import PcompresslrAPIClient, Message, CompressionConfig

client = PcompresslrAPIClient(api_key="your-lightreach-api-key")

# Using Message dataclass
result = client.complete(
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Hello!"},
    ],
    desired_hle=30,
    compress=True,
    compress_output=False,
    compression_config={
        "compress_system": False,
        "compress_user": True,
        "compress_assistant": False,
        "compress_only_last_n_user": 1,
    },
    temperature=0.7,
    max_tokens=1000,
    tags={"env": "production"},
)

print(result["decompressed_response"])
print(f"Model used: {result['routing_info']['selected_model']}")

Compression Only (No LLM Call)

from pcompresslr import PcompresslrAPIClient

client = PcompresslrAPIClient(api_key="your-lightreach-api-key")

# Compress text without making an LLM call
compressed = client.compress(
    prompt="Your text with repeated content here...",
    model="gpt-4",      # Model for tokenization
    algorithm="greedy", # 'greedy' or 'optimal'
    tags={"env": "dev"} # Optional tags
)

print(compressed["llm_format"])
print(f"Compression ratio: {compressed['compression_ratio']:.2%}")

# Decompress later
decompressed = client.decompress(compressed["llm_format"])
print(decompressed["decompressed"])

Command Line Interface

# Set your API key
export PCOMPRESLR_API_KEY=your-api-key

# Compress a prompt
pcompresslr "Your prompt with repeated text here..."

# Use optimal algorithm only
pcompresslr "Your prompt here" --optimal-only

# Use greedy algorithm only
pcompresslr "Your prompt here" --greedy-only

API Reference

PcompresslrAPIClient

Main API client for intelligent model routing and compression.

Constructor

PcompresslrAPIClient(
    api_key: str = None,  # Falls back to env vars
    api_url: str = None,  # Default: https://api.compress.lightreach.io
    timeout: int = 120    # Request timeout in seconds
)

Parameters:

• api_key (str, optional): LightReach API key. Falls back to LIGHTREACH_API_KEY or PCOMPRESLR_API_KEY env vars.
• api_url (str, optional): Override base API URL. Falls back to PCOMPRESLR_API_URL env var.
• timeout (int, optional): Request timeout in seconds. Default: 120 (2 minutes for LLM calls).

Methods

complete(messages, ...)

Messages-first completion with intelligent routing (POST /api/v2/complete).

Parameters:

Parameter	Type	Default	Description
messages	list[dict]	required	Conversation history with role and content
llm_provider	str	None	Provider constraint: "openai", "anthropic", "google", "deepseek", "moonshot". Omit for cross-provider optimization
desired_hle	float	None	Quality preference (0-40, where 40 is SOTA). If above an admin ceiling, it is clamped down
compress	bool	True	Whether to compress messages
compress_output	bool	False	Whether to request compressed output from LLM
algorithm	str	"greedy"	Compression algorithm: "greedy" or "optimal"
compression_config	dict	None	Per-role compression settings (see below)
temperature	float	None	LLM temperature parameter
max_tokens	int	None	Maximum tokens to generate
tags	dict[str, str]	None	Tags for cost attribution and tag-level HLE ceilings
max_history_messages	int	None	Limit conversation history length

compression_config options:

{
    "compress_system": False,         # default
    "compress_user": True,            # default
    "compress_assistant": False,      # default
    "compress_only_last_n_user": 1,   # default (None = compress all)
}

Response (dict):

{
    "decompressed_response": str,     # Final decompressed LLM response
    "compression_stats": {
        "original_size_chars": int,
        "compressed_size_chars": int,
        "original_tokens": int,
        "compressed_tokens": int,
        "compression_ratio": float,
        "token_savings": int,
        "token_savings_percent": float,
        "processing_time_ms": float,
    },
    "llm_stats": {
        "prompt_tokens": int,
        "completion_tokens": int,
        "total_tokens": int,
    },
    "routing_info": {
        "selected_model": str,          # Model chosen by system
        "selected_provider": str,       # Provider chosen by system
        "model_hle": float,             # HLE score of selected model
        "model_price_per_million": float,
        "requested_hle": float | None,
        "effective_hle": float | None,  # Effective HLE after admin ceilings
        "hle_source": str,              # "request", "tag", "global", or "none"
        "hle_clamped": bool,            # True if admin ceiling lowered desired_hle
    },
    "warnings": list[str],
    "cost_estimate": float | None,
    "savings_estimate": float | None,
}

Deprecated parameters (ignored in v1.0.0):

• model: System now selects models automatically
• hle_target_percent: Use desired_hle instead
• min_hle_score: Use desired_hle instead
• auto_select_by_hle: Always auto-selects now
• same_provider_only: Use llm_provider instead

compress(prompt, model, algorithm, tags)

Compression-only (POST /api/v1/compress).

Parameters:

• prompt (str, required): Text to compress
• model (str, optional): Model for tokenization. Default: "gpt-4"
• algorithm (str, optional): "greedy" or "optimal". Default: "greedy"
• tags (dict, optional): Tags for attribution

Response (dict):

{
    "compressed": str,
    "dictionary": dict[str, str],
    "llm_format": str,
    "compression_ratio": float,
    "original_size": int,
    "compressed_size": int,
    "processing_time_ms": float,
    "algorithm": str,
}

decompress(llm_format)

Decompress an LLM-formatted compressed prompt (POST /api/v1/decompress).

Parameters:

• llm_format (str, required): The llm_format string from a compress response

Response (dict):

{
    "decompressed": str,
    "processing_time_ms": float,
}

health_check()

Check API health status (GET /health).

Response (dict):

{
    "status": str,
    "version": str,
}

Data Classes

Message

from pcompresslr import Message

msg = Message(role="user", content="Hello!")
msg.to_dict()  # {"role": "user", "content": "Hello!"}

Roles: "system", "developer", "user", "assistant"

CompressionConfig

from pcompresslr import CompressionConfig

config = CompressionConfig(
    compress_system=False,        # default
    compress_user=True,           # default
    compress_assistant=False,     # default
    compress_only_last_n_user=1,  # default (None = compress all)
)
config.to_dict()

Environment Variables

Variable	Description
PCOMPRESLR_API_KEY	Your LightReach API key (primary)
LIGHTREACH_API_KEY	Your LightReach API key (alternative)
PCOMPRESLR_API_URL	Override the API base URL (advanced/testing)

Exceptions

Exception	Description
PcompresslrAPIError	Base exception class
APIKeyError	Invalid or missing API key
RateLimitError	Rate limit exceeded
APIRequestError	General API errors (including routing failures)

from pcompresslr import (
    PcompresslrAPIClient,
    APIKeyError,
    RateLimitError,
    APIRequestError,
)

try:
    result = client.complete(messages=[...])
except APIKeyError as e:
    print("Invalid API key")
except RateLimitError as e:
    print("Rate limited, please retry later")
except APIRequestError as e:
    print(f"API error: {e}")

How It Works

Compress Light Reach uses intelligent algorithms to identify repeated substrings in your prompts and replace them with shorter placeholders.

The library:

1. Identifies repeated substrings using efficient suffix array algorithms
2. Calculates token savings for each potential replacement
3. Selects optimal replacements that reduce total token count
4. Intelligently routes to the best model based on your quality requirements
5. Formats the result for easy LLM consumption
6. Provides perfect decompression

Examples

Example 1: Complete with Compression

from pcompresslr import PcompresslrAPIClient

client = PcompresslrAPIClient(api_key="your-lightreach-api-key")

prompt = """
Write a story about a cat. The cat is very friendly. 
Write a story about a dog. The dog is very friendly.
Write a story about a bird. The bird is very friendly.
"""

result = client.complete(
    messages=[{"role": "user", "content": prompt}],
    desired_hle=30,
)

print(result["decompressed_response"])
print(f"Model used: {result['routing_info']['selected_model']}")
print(f"Token savings: {result['compression_stats']['token_savings']} tokens")
print(f"Compression ratio: {result['compression_stats']['compression_ratio']:.2%}")

Example 2: Multi-turn Conversation

from pcompresslr import PcompresslrAPIClient

client = PcompresslrAPIClient(api_key="your-lightreach-api-key")

result = client.complete(
    messages=[
        {"role": "system", "content": "You are a helpful coding assistant."},
        {"role": "user", "content": "How do I read a file in Python?"},
        {"role": "assistant", "content": "You can use open() with a context manager..."},
        {"role": "user", "content": "How about writing to a file?"},
    ],
    desired_hle=30,
    compression_config={
        "compress_system": False,
        "compress_user": True,
        "compress_assistant": False,
        "compress_only_last_n_user": 2,  # Only compress last 2 user messages
    },
)

Example 3: Provider-Constrained Request

from pcompresslr import PcompresslrAPIClient

client = PcompresslrAPIClient(api_key="your-lightreach-api-key")

# Constrain to Anthropic models only
result = client.complete(
    messages=[
        {"role": "system", "content": "You are a creative writing assistant."},
        {"role": "user", "content": "Write a haiku about coding."},
    ],
    llm_provider="anthropic",  # Only use Anthropic models
    desired_hle=30,
)

print(result["decompressed_response"])
print(f"Model: {result['routing_info']['selected_model']}")  # e.g., claude-3-haiku

Getting an API Key

To use Compress Light Reach, you need an API key from compress.lightreach.io.

1. Visit compress.lightreach.io
2. Sign up for an account
3. Get your API key from the dashboard
4. Set it as an environment variable: export PCOMPRESLR_API_KEY=your-key

Security & Privacy

BYOK model: Provider keys (OpenAI/Anthropic/Google/etc.) are managed in the dashboard and never passed through this SDK. The SDK only uses your LightReach API key for authentication with the service.

BYOK Provider Key Encryption (Required for Dashboard Settings → Provider Keys)

Provider keys are encrypted at rest using Fernet (symmetric authenticated encryption). The backend requires a Fernet key via:

• API_KEY_ENCRYPTION_KEY

Generate a key:

python -c "from cryptography.fernet import Fernet; print(Fernet.generate_key().decode())"

Set it in your runtime environment (examples):

• Docker Compose: set API_KEY_ENCRYPTION_KEY in your shell or .env before running docker compose up
• GitHub Actions: store the value as a GitHub Secret, then map it to the environment variable API_KEY_ENCRYPTION_KEY in your deploy workflow

Requirements

• Python 3.8+
• tiktoken >= 0.5.0
• requests >= 2.31.0
• urllib3 >= 2.0.0
• python-dotenv >= 1.0.0

License

MIT License - see LICENSE file for details.

Support

• Documentation: compress.lightreach.io/docs
• Issues: GitHub Issues
• Email: [e-mail]

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.