Location: /opt/mcp-servers/shared/training_helper.py
Core Functions:
- export_kb_training(kb_id)
- Walk KB tree, create Q&A for each node
- 'What is [title]?' → '[content]'
-
'What are the sections of [parent]?' → '[children list]'
-
export_context_training()
- Convert Context entries to training pairs
-
'What is [key]?' → '[value]'
-
export_documents_training(doc_ids)
- Parse documents into Q&A pairs
-
Extract key facts, procedures, policies
-
export_track_training()
- Convert workflows into instruction pairs
-
'How do I [task type]?' → '[workflow pattern]'
-
generate_dataset(output_path)
- Combine all sources into JSONL
-
Deduplicate, validate, quality check
-
validate_dataset(path)
- Check format correctness
- Identify duplicates
- Score quality