| name | rag-knowledge-management |
| description | Manage the Dual RAG learning loop — search, analyze, and save technical facts, project state, and user feedback to Qdrant vector collections. |
RAG Knowledge Management
You maintain a learning loop with three experience collections and three demo knowledge-base collections:
Experience collections (read/write)
- experience_facts — hard technical rules, coding standards, synonym mappings
- experience_state — project context (system names, versions, team roles, deadlines)
- experience_feedback — user corrections and lessons learned
Knowledge-base collections (read-only)
- demo_literature — literary content, stories, biographical facts
- demo_news — real-world events, sports results, awards, geopolitical changes
- demo_sap_cases — SAP support cases with troubleshooting steps
When to Search RAG
- Before every action: search
experience_facts for rules that constrain how the task should be done.
- Before answering context-dependent questions: search
experience_state for project-specific information.
- When the user corrects you: search
experience_feedback for prior corrections on the same topic.
- For general knowledge questions (events, people, literature, history): search
demo_news, demo_literature.
- For SAP troubleshooting: search
demo_sap_cases in addition to experience_facts.
When to Save to RAG
After completing a task or receiving a correction:
- Identify if the interaction produced a new technical rule (fact), project context update (state), or correction (feedback).
- Formulate a concise, self-contained summary.
- Save it to the appropriate collection via the RAG upsert pipeline.
- Do NOT tell the user you saved it unless explicitly asked.
Deduplication
Before saving, the pipeline checks for semantic similarity (threshold 0.92). If a near-duplicate exists, the save is skipped. You do not need to check manually.