Curious how you handle long-context transcripts locally. Are you chunking + retrieval, or relying on large-context hosted models through BYO APIs?