Long Transcripts and Duplicates

chunking, cost, dedup

Part of: Meeting Notes to Action Items

A five-minute standup fits in one call. An hour-long planning meeting is thousands of words, and a full day of transcripts can blow past what you want to send in a single request, both for the context limit and for the bill. This lesson handles scale: split the transcript, and clean up the duplicates that splitting creates. Chunk the transcript You break a long transcript into smaller pieces and extract from each, then combine the results. The simplest, most predictable split is by a fixed number of lines: A chunk of, say, 40 lines is small enough to send cheaply and get a focused extraction. You run the extractor on each chunk and concatenate the item lists. The tradeoff: an action item mentioned across a chunk boundary might get caught twice, once in each chunk, which is exactly the problem the next step solves. Estimate the cost before you spend it Every chunk is a call, and every call is tokens you pay for. A rough gauge, about four characters per token, lets you predict a bill before running: Summing that over your chunks tells you roughly what one pass costs. If a transcript is 8,000 characters, that's ~2,000 input tokens per full pass, useful for deciding whether to chunk fi

Challenge: Deduplicate the Action Items