← All questions
System DesignHard· policy exception· chat transcript characteristics· operational cost

Your CTO reads the negative result and mandates fixed‑s...

Concede that the mandate is appropriate for documentation, policy, and knowledge‑base corpora, which dominate the company. Then explain that customer‑support chat transcripts consist of concatenated speaker turns on drifting topics—a high‑Δ regime not covered by the negative result. Offer the Δ measurement as the arbiter rather than asking for a blanket exemption. Argue that transcripts already contain exact structural boundaries (speaker turns, session IDs), so the proper chunker is a metadata splitter, not a cosine detector. Highlight the operational cost of a per‑corpus exception: two chunking policies mean two ingestion pipelines, two length distributions, and a separate evaluation set for each policy.

Code playground (Python)