Context Retrieval Sorting

Expert

Order context documents to avoid model recall degradation.

01 The Concept

LLMs suffer from the 'lost in the middle' phenomenon, where they recall information placed at the beginning or end of the context window better than information in the middle.

Sorting context documents places the most critical references first and last to ensure recall, while placing general or less important references in the middle.

02 Weak vs. Strong

EX 01Context Positioning Setup

# Sort documents to prevent 'lost in the middle' decay docs = [doc1, doc2, doc3, doc4, doc5] # Place most critical docs at indices 0 and -1 sorted_docs = [critical_doc_A, doc2, doc3, doc4, critical_doc_B] prompt = "Context:\n" + "\n".join(sorted_docs) + "\nTask: ..."

→ Why it works

Combats attention decay. Placing load-bearing documents at the beginning and end of the context block ensures the model retains them.

03 Key Points

01Lost in the Middle: The statistical dip in model attention weights for items in the center of prompts.

02Positional Sorting: Placing high-value reference files at the absolute beginning or end of context.

03Relevance Ranking: Sorting documents by similarity search scores before compilation.

04Telemetry Auditing: Testing query accuracy across different document positions.

05Context Packing: Grouping static resources together to hit prompt caching points.

04 Model-Specific Notes

Claude has excellent long-context recall, but placing critical guidelines at the very end of the prompt ensures compliance.

05 For Your Role

It's like studying for a test: you usually remember the very first page and the very last page of the notes best, and forget the middle.

Next: Directed Acyclic Graph Prompts →