llm-llm-ctx-mgr 0.1.0

A middleware layer for managing, budgeting, and optimizing LLM context windows.

Project: llm-ctx-mgr - llm context manager (engineering)

1. Background & Problem Statement

Large Language Models (LLMs) are moving from simple "Prompt Engineering" (crafting a single query) to "Context Engineering" (managing a massive ecosystem of retrieved documents, tools, and history).

The current problem is Context Pollution:

1. Overloading: RAG (Retrieval Augmented Generation) pipelines often dump too much data, exceeding token limits.
2. Noise: Duplicate or irrelevant information confuses the model and increases hallucination rates.
3. Formatting Chaos: Different models (Claude vs. Llama vs. GPT) require different formatting (XML vs. Markdown vs. Plain Text), leading to messy, hard-to-maintain string concatenation code.
4. Black Box: Developers rarely see exactly what "context" was sent to the LLM until after a failure occurs.

The Solution: llm-ctx-mgr acts as a middleware layer for the LLM pipeline. It creates a structured, optimized, and budget-aware "context payload" before it reaches the model.

---

2. Architecture: Where It Fits

The package sits strictly between the Retrieval/Agent Layer (e.g., LangChain, LlamaIndex) and the Execution Layer (the LLM API).

Diagram: The "Before" (Standard Pipeline)

Without llm-ctx-mgr, retrieval is messy and often truncated arbitrarily.

graph LR
    A[User Query] --> B[LangChain Retriever]
    B --> C{Result: 15 Docs}
    C -->|Raw Dump| D[LLM Context Window]
    D -->|Token Limit Exceeded!| E[Truncated/Error]

Diagram: The "After" (With llm-ctx-mgr)

With your package, the context is curated, prioritized, and formatted.

graph LR
    A[User Query] --> B[LangChain Retriever]
    B --> C[Raw Data: 15 Docs + History]
    C --> D[**llm-ctx-mgr**]
    
    subgraph "Your Middleware"
    D --> E["1. Token Budgeting"]
    E --> F["2. Semantic Pruning"]
    F --> G["3. Formatting (XML/JSON)"]
    end
    
    G --> H[Optimized Prompt]
    H --> I[LLM API]

---

3. Key Features & Tools

Here is the breakdown of the 4 core modules, the features they provide, and the libraries powering them.

Module A: The Budget Controller (budget)

• Goal: Ensure the context never exceeds the model's limit (e.g., 8192 tokens) while keeping the most important information.
• Feature: PriorityQueue. Users assign a priority (Critical, High, Medium, Low) to every piece of context. If the budget is full, "Low" items are dropped first.
• Supported Providers & Tools:
  • OpenAI (gpt-4, o1, o3, etc.): tiktoken — fast, local token counting.
  • HuggingFace (meta-llama/..., mistralai/...): tokenizers — for open-source models.
  • Google (gemini-2.0-flash, gemma-...): google-genai — API-based count_tokens.
  • Anthropic (claude-sonnet-4-20250514, etc.): anthropic — API-based count_tokens.
• Installation (pick what you need):

  pip install llm-ctx-mgr[openai]       # tiktoken
  pip install llm-ctx-mgr[huggingface]   # tokenizers
  pip install llm-ctx-mgr[google]        # google-genai
  pip install llm-ctx-mgr[anthropic]     # anthropic
  pip install llm-ctx-mgr[all]           # everything
  

Module B: The Semantic Pruner (prune)

• Goal: Remove redundancy. If three retrieved documents say "Python is great," keep only the best one.
• Features:
  • Deduplicator (block-level): Calculates cosine similarity between context blocks and removes duplicate blocks. Among duplicates, the highest-priority block is kept.
  • Deduplicator.deduplicate_chunks() (chunk-level): Splits a single block's content by separator (e.g. \n\n), deduplicates the chunks internally, and reassembles the cleaned content. Ideal for RAG results where multiple retrieved chunks within one block are semantically redundant.
• Tools:
  • FastEmbed: Lightweight embedding generation (CPU-friendly, no heavy PyTorch needed).
  • Numpy: For efficient vector math (dot products).
• Installation:

  pip install llm-ctx-mgr[prune]
  

Module C: Context Distillation (distill)

• Goal: Compress individual blocks by removing non-essential tokens (e.g., reduces a 5000-token document to 2500 tokens) using a small ML model.
• Feature: Compressor. Uses LLMLingua-2 (small BERT-based token classifier) to keep only the most important words.
• Tools:
  • llmlingua: Microsoft's library for prompt compression.
  • onnxruntime / transformers: For running the small BERT model.
• Installation:

  pip install llm-ctx-mgr[distill]
  

Module D: The Formatter (format)

• Goal: Adapt the text structure to the specific LLM being used without changing the data.

• Feature: ModelAdapter.

• Claude Mode: Wraps data in XML tags (<doc id="1">...</doc>).

• Llama Mode: Uses specific Markdown headers or [INST] tags.

• Tools:

• Jinja2: For powerful, logic-based string templates.

• Pydantic: To enforce strict schema validation on the input data.

Module E: Observability (inspect)

• Goal: Let the developer see exactly what is happening.
• Feature: ContextVisualizer and Snapshot. Prints a colored bar chart of token usage to the terminal and saves the final prompt to a JSON file for debugging.
• Tools:
• Rich: For beautiful terminal output and progress bars.

---

4. Installation & Usage Guide

Installation

pip install llm-ctx-mgr[all]

Feature A: Budgeting & Priority Pruning

Ensure your context fits the token limit by prioritizing critical information.

from context_manager import ContextEngine, ContextBlock
from context_manager.strategies import PriorityPruning

# 1. Initialize Engine with a token limit
engine = ContextEngine(
    model="gpt-4",
    token_limit=4000,
    pruning_strategy=PriorityPruning()
)

# 2. Add Critical Context (System Prompts) - NEVER dropped
engine.add(ContextBlock(
    content="You are a helpful AI assistant.",
    role="system",
    priority="critical"
))

# 3. Add High Priority Context (User History) - Dropped only if critical fills budget
engine.add(ContextBlock(
    content="User: Explain quantum computing.",
    role="history",
    priority="high"
))

# 4. Add Medium/Low Priority (RAG Docs) - Dropped first
docs = ["Quantum computing uses qubits...", "Quantum mechanics is...", "Cake recipes..."]
for doc in docs:
    engine.add(ContextBlock(
        content=doc,
        role="rag_context",
        priority="medium"
    ))

# 5. Compile - Triggers budgeting and pruning
final_prompt = engine.compile()
print(f"Final token count: {engine.compiled_tokens}")

Feature B: Semantic Pruning (Deduplication)

Remove duplicate or highly similar content to save space and reduce noise.

from context_manager import ContextEngine, ContextBlock
from context_manager.prune import Deduplicator

# 1. Initialize Deduplicator (uses FastEmbed by default)
dedup = Deduplicator(threshold=0.85)

# 2. Initialize Engine with Deduplicator
engine = ContextEngine(
    model="gpt-4",
    token_limit=4000,
    deduplicator=dedup
)

# 3. Add duplicate content (simulating RAG retrieval)
# The second block will be detected as a duplicate and removed/merged
engine.add(ContextBlock(
    content="Python was created by Guido van Rossum.",
    role="rag_context",
    priority="medium"
))
engine.add(ContextBlock(
    content="Guido van Rossum created the Python language.",
    role="rag_context",
    priority="low"  # Lower priority duplicate is dropped
))

# 4. Compile - Deduplication happens before budgeting
final_prompt = engine.compile()

Feature C: Context Distillation (Compression)

Compress long documents using LLMLingua to keep essential information within budget.

from context_manager import ContextEngine, ContextBlock, Priority
from context_manager.distill import LLMLinguaCompressor

# 1. Initialize Compressor (loads small local model)
compressor = LLMLinguaCompressor(
    model_name="microsoft/llmlingua-2-xlm-roberta-large-meetingbank",
    device_map="cpu"
)

# 2. Initialize Engine
engine = ContextEngine(
    model="gpt-4",
    token_limit=2000,
    compressor=compressor
)

# 3. Add a long document marked for compression
long_text = "..." * 1000  # Very long text
engine.add(ContextBlock(
    content=long_text,
    role="rag_context",
    priority=Priority.HIGH,
    can_compress=True  # <--- Triggers compression for this block
))

# 4. Compile - Compression happens first, then deduplication, then budgeting
final_prompt = engine.compile()

5. Roadmap for Development

1. v0.1 (MVP): tiktoken counting and PriorityPruning. (Done)
2. v0.2 (Structure): Jinja2 templates for formatting. (Done)
3. v0.3 (Smarts): FastEmbed for semantic deduplication. (Done)
4. v0.4 (Vis): Rich terminal visualization. (Done)
5. v0.5 (Distill): LLMLingua integration for context compression. (Done)
6. v0.6 (Next): Streaming support and advanced caching strategies.

This design gives you a clear path to building a high-value tool that solves a specific, painful problem for AI engineers.