A RAG system retrieves 10 documents (each ~1,500 tokens) and includes them all in Claude's context along with a 500-token system prompt and a 200-token user query. Total context is approximately 16,200 tokens. Response quality degrades for questions whose answers appear in documents 7-10 (later in context). What phenomenon explains this and what is the mitigation?

The correct explanation is the lost-in-the-middle effect: transformer-based large language models tend to attend more strongly to content near the beginning and end of the context window, with reduced attention to middle positions. In this setup, the system prompt and early documents sit near the high-attention beginning, the user query sits near the high-attention end, and documents 7-10 fall into lower-attention regions, so answers located there are harder to use. Mitigations include placing the most relevant document first, placing the second most relevant document last, and reducing the number of retrieved documents so all content stays within high-attention zones. The retrieval-quality explanation is wrong because the problem describes degradation based on document position, not on the relevance of the retrieved content; this is a position-based attention effect, not a retrieval quality issue. The Claude-bug explanation is wrong because there is no known bug in Claude that specifically limits correct processing to six documents; the effect applies to large language models generally and depends on context length and relative position. The chunk-size explanation is wrong because reducing chunk size does not address the root cause; smaller chunks might move answers to different positions, but that is indirect and not the position-effect fix.