---

name: knowledge-rag description: "Operate on BiSheng knowledge libraries, knowledge spaces, document ingestion, parser providers, RAG pipeline, Milvus/Elasticsearch recall, and knowledge workers." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

knowledge-rag

Use this sub-skill when a task touches BiSheng knowledge libraries, knowledge spaces, QA libraries, file upload/parse states, document parser providers, the Load → Transform → Ingest pipeline, Milvus + Elasticsearch dual recall, MinIO file artifacts, or knowledge_celery worker behavior.

Start here

Run bundled helper commands from this sub-skill directory, or adjust the script path to this directory after import.

1. Inspect config-shaped text before debugging parser/storage problems:

   python scripts/check_knowledge_config.py --config <bisheng-checkout>/docker/bisheng/config/config.yaml
   

2. Read references/workflows.md for ingestion, parser, retrieval, and worker workflows.
3. Read references/troubleshooting.md for parser, object storage, vector index, queue, and permission symptoms.

Owned responsibilities

• Knowledge API/domain services under src/backend/bisheng/knowledge/.
• RAG pipeline files under knowledge/rag/, especially file pipelines, loader/provider selection, transformers, Milvus and Elasticsearch factories.
• bisheng_langchain/rag/ retrievers, RAG pipeline, scoring, and rerank integration when used by BiSheng knowledge workflows.
• Knowledge Celery workers under src/backend/bisheng/worker/knowledge/.
• Tests under src/backend/test/knowledge/, test/knowledge/rag/, and safe config/service tests.
• Frontend knowledge-space or knowledge-library callers only at routing depth; UI rendering details belong to frontend-apps.

Route sibling areas instead of duplicating them

• Use identity-permissions-tenancy for space visibility, OpenFGA/ReBAC, approval, tenant filtering, and cursor permission scans.
• Use workflow-engine for workflow RAG node execution outside the knowledge pipeline itself.
• Use backend-core for generic FastAPI router/envelope/service conventions.
• Use frontend-apps for Platform/Client knowledge pages and request wrappers.
• Use deployment-maintenance for Docker storage service startup and production config rollout.

Non-negotiables

• Treat Load, Transform, and Ingest as separate stages when debugging: a loader success does not prove vector writes or preview cache.
• Preserve Milvus dense and Elasticsearch sparse/BM25 behavior when changing retrieval.
• Do not run external parser services, OCR downloads, or large document parsing as a smoke check unless the environment is approved.
• Use small fixtures for tests; do not require production MinIO/Milvus/ES for unit-level parser or service logic.
• Keep parser-provider credentials, endpoints, and object-storage secrets out of skill output and committed test fixtures.