Is it just me, or do agents get worse at the end of a long document?
Show notes for a two-hour episode: the first hour is summarised well, the second hour gets vague. Same model, same prompt. Is this a known thing and what do people do about it?
Known. Chunk by segment with a running summary, then summarise the summaries. Also score each chunk separately in your evaluation set, or the average hides exactly this.
We cut long recordings at natural breaks before anything else touches them. The pipeline treats a two-hour file as four thirty-minute ones.
Chunking by segment fixed it in one afternoon. Should have asked a week ago.