thin-adapter-turn-classification

name	thin-adapter-turn-classification
description	Guidance for classifying user turns into directive, upload, or question flows.

Thin Adapter Turn Classification

Turn signals

• has_directive: one or more ingestion directives are present
• has_files: one or more file uploads are present
• has_substantive_question: remaining user text contains a meaningful natural-language request after stripping directives and upload-only artifacts

Routing rules

1. directive only -> synthetic directive acknowledgement
2. upload only -> synthetic upload acknowledgement
3. directive + upload, no question -> synthetic combined acknowledgement
4. any turn with a substantive question -> process directives/uploads first, then forward to chat LLM

Guardrails

• directive parsing must not leak into user-visible LLM prompts
• acknowledgements must be deterministic and generated locally