| name | knowledge_management |
| description | Use when a user wants to build, search, inspect, delete, or reindex a namespace-scoped knowledge base through the runtime knowledge tool. |
Knowledge Management
Use the builtin knowledge tool for actual operations. This skill only describes the chat workflow.
Workflow
- Missing namespace: ask for the target
namespace when the user has not provided one and the current task context does not imply one.
- Small text: call
knowledge with action=ingest_text, namespace, and source.text.
- Large local file: call
knowledge with action=ingest_file; return the job_id and current status from the tool result.
- Progress check: call
knowledge with action=ingest_status, namespace, and job_id.
- Search: call
knowledge with action=search, then cite returned source_id and chunk_id in the answer.
- Inspect source text: call
knowledge with action=get_chunk for the cited chunk_id.
- Delete: use
action=delete_source only when the user clearly authorizes deleting that source; destructive operations need 用户确认.
- Reindex: use
action=reindex only when the user clearly authorizes rebuilding vectors and gives confirmation for the namespace, for example knowledge.reindex --namespace fanqie.
- Model status: call
knowledge with action=model_status when the user asks whether local embedding/reranker models are loaded.
- Model unload: call
knowledge with action=unload_models when the user asks to release knowledge model memory.
Output rules
- For ingest jobs, tell the user the knowledge write is running and include
job_id.
- For search answers, include concise citations using
source_id and chunk_id; 中文回答可写“引用:source_id / chunk_id”。
- For missing model or config mismatch diagnostics, surface the tool message and the suggested
knowledge.reindex --namespace ... command when provided.
- For model_status/unload_models, report only loaded/not_loaded, model_id, and idle_ttl_seconds.