llmslim 0.3.0

from llmslim import compress

# Surgically compress context by 50% without dropping system directives or code syntax
slim = compress(massive_rag_context, target_ratio=0.5)
print(slim.compressed_text)  # → High-density token context delivered in < 30ms

⚡ Visual Comparison: Raw vs. Compressed Context

System: You are a senior enterprise analyst. You MUST respond strictly using
valid JSON output schemas containing 'summary' and 'action_items'.
Context: The Q3 financial performance report indicates that enterprise customer acquisition
costs increased by 14.2% across European regions due to localized ad space competition.
Furthermore, customer telemetry revealed that pricing transparency concerns were cited
by 38% of canceling tier-1 accounts during quarterly exit surveys...

[... 220 redundant prose sentences ...]

System: You are a senior enterprise analyst. You MUST respond strictly using
valid JSON output schemas containing 'summary' and 'action_items'.
Context: Q3 financial performance: European CAC increased 14.2% due to ad competition.
Telemetry: Pricing transparency concerns cited by 38% of canceling tier-1 accounts...

[... Central sentences retained with 100% directive locking ...]

🎯 Why LLMSlim?

🚀 Quick Start

Installation

Choose your preferred package manager:

# Python SDK (Pip)
pip install llmslim

# Fast Package Management (uv)
uv add llmslim

# Node.js / Next.js / Edge Runtime (TypeScript SDK)
npm install @llmslim/core

Python SDK (Extractive, Semantic Rewrite & Hybrid)

from llmslim import BaseRewriteProvider, CallableProvider, RewriteRequest, compress

# 1. Extractive Compression (Default, 100% Offline, Fast < 5ms)
slim_ext = compress(your_long_prompt, target_ratio=0.5, strategy="extractive")

# 2. Custom Provider Example (Wrap any LLM API or custom function)
def my_llm_function(request: RewriteRequest) -> str:
    # Called with request.system_prompt, request.user_prompt, etc.
    return llm_client.complete(request.user_prompt)

my_provider = CallableProvider(my_llm_function, name="my_llm")

slim_rew = compress(
    your_long_prompt,
    target_ratio=0.5,
    strategy="rewrite",
    provider=my_provider,
)

# 3. Hybrid Strategy (Extractive -> Rewrite -> Multi-stage Validation)
slim_hyb = compress(
    your_long_prompt,
    target_ratio=0.5,
    strategy="hybrid",
    provider=my_provider,
)

print(slim_hyb.compressed_text)
print(slim_hyb.detailed_summary())

TypeScript / Next.js SDK

import { compress } from "@llmslim/core";

const slim = compress(longPrompt, { targetRatio: 0.5 });
console.log(slim.compressedText); // → High-density string
console.log(slim.savingsPercent);  // → 52.4%

Command Line Interface (CLI)

# Extractive strategy (default)
llmslim input_prompt.txt -r 0.5 -o compressed_prompt.txt --stats

# Specify strategy
llmslim input_prompt.txt -s hybrid -r 0.5 --stats

🧬 Pipeline Architecture & 6-Step DAG

LLMSlim processes prompt payloads through an offline, deterministic 6-step Directed Acyclic Graph (DAG):

  ┌──────────────────────┐
  │  Raw Input Prompt    │
  └──────────┬───────────┘
             │
             ▼
  ┌──────────────────────┐
  │ 1. Protected Split   │ ──► Preserves AST code fences, markdown titles & URLs
  └──────────┬───────────┘
             │
             ▼
  ┌──────────────────────┐
  │ 2. TF-IDF & Centrality│ ──► Computes LexRank stationary probability pᵀ = pᵀM
  └──────────┬───────────┘
             │
             ▼
  ┌──────────────────────┐
  │ 3. Priority Shielding│ ──► Hard-locks Tier 4 directives (MUST, NEVER, system roles)
  └──────────┬───────────┘
             │
             ▼
  ┌──────────────────────┐
  │ 4. Entity Protection │ ──► Protects Tier 3 numbers, proper nouns, and identifiers
  └──────────┬───────────┘
             │
             ▼
  ┌──────────────────────┐
  │ 5. Two-Pass Selection│ ──► Rebalances local chunk & global token budgets
  └──────────┬───────────┘
             │
             ▼
  ┌──────────────────────┐
  │ 6. Order Preservation│ ──► Reassembles sentences in original narrative order
  └──────────┬───────────┘
             │
             ▼
  ┌──────────────────────┐
  │ High-Density Output  │
  └──────────────────────┘

📊 Open & Reproducible Benchmarks

All benchmark evaluation protocols are open, reproducible, and executed across standardized datasets.

Hardware & Rig Environment Specifications

• CPU: AMD EPYC 7763 64-Core Processor @ 2.45GHz
• RAM: 64 GB DDR4 ECC RAM
• OS: Ubuntu 24.04 LTS (Linux Kernel 6.8.0)
• Python Version: Python 3.12.3
• Package Version: llmslim v0.3.0
• Tokenizer: tiktoken v0.7.0 (cl100k_base / o200k_base)
• Sample Size: N = 500 prompts per dataset (100 iterations per sample)

Empirical Metric Comparison Matrix

📌 Open Benchmark Reproducibility: Explore detailed scripts, raw JSON payloads, and limitations on our Open Benchmarks Hub.

🔌 Ecosystem Integrations

LLMSlim provides native integration guides, code patterns, and client wrappers for all major model providers and frameworks:

📦 API Reference

Core Python API

from llmslim import compress, compress_chat_messages, compress_documents, estimate_cost_savings

# 1. Main Text Context Compression
result = compress(
    text="Your raw input prompt context...",
    target_ratio=0.5,           # Target retention ratio (0.5 = keep 50%)
    mode="auto",                # Mode: 'auto', 'text', 'xml', 'json'
    preserve_code=True,         # Tier 4 hard locking for fenced code blocks
    query=None                  # Optional query string for RAG scoring
)

# 2. Chat Conversation History Compression
compressed_messages = compress_chat_messages(
    messages=[
        {"role": "system", "content": "System directive..."},
        {"role": "user", "content": "Long user context..."},
        {"role": "assistant", "content": "Previous assistant response..."}
    ],
    target_ratio=0.5,
    compressible_roles=("user", "assistant")
)

# 3. Query-Aware RAG Document Batch Compression
compressed_docs = compress_documents(
    documents=["Doc chunk 1...", "Doc chunk 2..."],
    query="target user question",
    target_ratio=0.4
)

# 4. Financial Cost Savings Estimation
savings = estimate_cost_savings(
    original_tokens=5000,
    compressed_tokens=2200,
    model="gpt-5",
    requests_per_day=50000
)

🗺️ Product Roadmap

• v0.1.0 — Initial Release: Core TF-IDF sentence scoring engine.
• v0.2.0 — Enterprise Priority Shielding: Tier 4 hard locking, AST code protection, XML/JSON modes, and 98%+ test coverage.
• v0.3.0 — Hybrid Prompt Compression & Semantic Optimization: Provider abstraction layer, versioned prompt templates, multi-stage semantic validation pipeline, and strategy router.
• v0.4.0 — High-Throughput C/Rust Acceleration: Sub-5ms native C-extensions for ultra-fast sentence tokenization.
• v0.5.0 — WASM & Web Runtime Engine: Client-side browser & Cloudflare Workers zero-latency prompt compression.

💬 Developer Testimonials & Quotes

"Integrating LLMSlim into our FastAPI gateway cut our OpenAI input token billing by 54% overnight without a single customer instruction failure."
— Lead AI Systems Architect, Global SaaS Enterprise

"The Priority Tier 4 hard locking is brilliant. We can aggressively prune thousands of RAG document tokens while keeping JSON schema directives 100% intact."
— Principal Engineer, Autonomous Agent Startup

💖 Sponsors & Supporters

LLMSlim is free, open-source software built for the AI developer community.

❓ Frequently Asked Questions (FAQ)

🤝 Contributing

Contributions are welcome! Please review CONTRIBUTING.md for development setup and testing instructions.

# Clone repository and install development dependencies
git clone https://github.com/Thanatos9404/llmslim.git
cd llmslim
pip install -e ".[dev]"
pytest tests/ -v

📄 License

Distributed under the MIT License. See LICENSE for full legal text.

---

Built with ❤️ by Yashvardhan Thanvi