How to Fix Your Context

This repository contains practical implementations of the 6 key context management techniques outlined in Drew Breunig's post "How to Fix Your Context".

🚀 Quickstart

Prerequisites

Python 3.9 or higher
uv package manager

Installation

Clone the repository and activate a virtual environment:

git clone https://github.com/langchain-ai/how_to_fix_your_context
cd how_to_fix_your_context
uv venv
source .venv/bin/activate  # On Windows: .venv\Scripts\activate

Install dependencies:

uv pip install -r requirements.txt

Set up environment variables for the model provider(s) you want to use:

export OPENAI_API_KEY="your-openai-api-key"
export ANTHROPIC_API_KEY="your-anthropic-api-key"

Background

The Context Problem

Chroma's report on Context Rot notes that LLMs do not treat every token in their context window equally. Across 18 models (including GPT‑4.1, Claude 4, Gemini 2.5, Qwen3, etc.), they show that performance on even very simple tasks degrades—often in non‑uniform and surprising ways—as the input length grows.

With this in mind, Drew Breunig outlined four failure modes on why long contexts fail:

Context Poisoning - Hallucinations or errors that enter the context and get repeatedly referenced
Context Distraction - When context grows so large that models focus more on accumulated history than training
Context Confusion - Superfluous content that influences response quality, as models feel compelled to use all available context
Context Clash - Conflicting information within the accumulated context that degrades reasoning

Context Management Techniques

Drew outlined 6 context management techniques to help fix these failure modes. This repository demonstrates each technique using a set of Jupyter notebooks and LangGraph.

LangGraph

LangGraph is a low is a low-level orchestration framework. You can lay out agents and workflows as a set of nodes, define the logic within each one, and define an state object that is passed between them. A StateGraph is LangGraph's primary abstraction for building these stateful workflows and agents with:

Nodes are processing steps that receive current state and return updates
Edges connect nodes to create execution flow (linear, conditional, or cyclical)
State serves as a shared scratchpad between nodes

This low-level control makes it easy to implement each of these techniques.

1. RAG (Retrieval-Augmented Generation)

Notebook: notebooks/01-rag.ipynb

Selectively adding relevant information to improve LLM responses while preventing information overload. RAG prevents context overload by carefully selecting relevant information from vector databases.