Your CTO mandates fixed-size chunking for all indexes,...

Acknowledge the mandate’s utility for standard knowledge bases. For transcripts, argue that the high-delta nature of drifting topics between speakers makes them a unique case. The best approach is to use the existing structural metadata (speaker turns and session IDs) to split chunks, which is superior to both fixed-size and cosine-based semantic chunking. Be prepared to defend the operational cost of maintaining a separate ingestion path, length distribution, and evaluation set for this specific policy.