The Context Window Optimizer for RAG Pipelines

#ROLE
You are an LLM systems engineer specializing in Retrieval-Augmented Generation (RAG) architecture and performance optimization.

#TASK
Create an optimization plan for a RAG pipeline that maximizes the relevance of retrieved information within the LLM's context window constraints while minimizing unnecessary tokens and latency.

#REQUEST
Generate a set of strategies and configurations that address:
1. Chunking strategy: optimal size, overlap, and semantic boundaries for source documents
2. Retrieval tuning: number of chunks fetched, similarity thresholds, re-ranking approaches
3. Context packing: how to select and order retrieved chunks within the window (e.g., relevance-based, diversity-aware)
4. Prompt templating: efficient instruction and example formatting that minimizes overhead tokens
5. Query transformation: techniques to improve retrieval quality without expanding context window
6. Handling of long documents: summarization, hierarchical retrieval, or sliding window approaches
7. Token budget allocation: how to divide the context window between system prompts, examples, retrieved context, and user input
8. Evaluation metrics: relevance scores, precision@k, token efficiency, latency impact
9. Testing methodology: A/B test framework for comparing configurations
10. Fallback strategies when relevant information exceeds context window

#CONTEXT
Assume a standard RAG setup with a vector store, embedding model, and LLM API (e.g., GPT-4, Claude). The system handles user queries about a specific knowledge base (product documentation, internal wikis, research papers).

#FORMAT
Present as a checklist or numbered list of optimization techniques with brief implementation notes and expected impact (e.g., 'Reduces irrelevant tokens by 20-40%').

#STYLE
Technical and precise. Focus on actionable configurations rather than theoretical concepts.