ner-rag-kg-reranking

name

ner-rag-kg-reranking

description

Operate, test, and diagnose Michael's local medical NER, RAG, pgvector, KG, BGE reranking, and Apache AGE pipeline. Use this project skill whenever the user mentions NER-RAG-KG-reranking, kg_pipeline, GLiNER biomedical NER, pgvector embeddings, BGE reranker, SNOMED KG, Apache AGE medical_graph, edge candidates, ontology validation, or asks to test a clinical query against the local `medical` database.

NER RAG KG Reranking

Scope

Use this skill for the local medical setup connected to /Users/michaelahern/ai-servers.

The pipeline combines:

• GLiNER biomedical NER on port 8091
• embeddings/RAG via pgvector and local endpoint 8084
• BGE reranking on port 8085
• relation tables public.kg_entities and public.kg_relations
• optional Apache AGE promotion into graph medical_graph

Load references/medical-kg-pipeline.md only when commands, SQL, or deeper diagnostic detail are needed.

Workflow

1. Check service state from /Users/michaelahern/ai-servers:
   • ./aictl status
   • ./aictl health
2. Verify GLiNER 8091, embeddings 8084, and reranker 8085 before a pipeline test.
3. Verify PostgreSQL state before drawing conclusions:
   • schema kg_pipeline
   • tables kg_pipeline.runs, entity_mentions, edge_candidates
   • tables public.kg_entities, public.kg_relations
   • AGE graph medical_graph
4. Prefer a targeted kg_pipeline.run_chunk(...) smoke before batch work.
5. For clinical retrieval, separate exact keyword evidence, semantic neighbors, reranker support, candidate edges, promoted KG relations, and AGE state.

Guardrails

• Do not treat semantic proximity as a validated clinical relation.
• Require text evidence and reranking support before creating or promoting an edge.
• Do not run the full SNOMED import unless Michael explicitly asks for it.
• Do not promote to AGE by default; use promotion flags only when intended.
• If NER linking looks over-merged, inspect entity_mentions and edge_candidates before changing linkage behavior.

Expected Output

When answering Michael, include:

• concrete local service and PostgreSQL status
• commands or SQL executed
• useful counters such as mentions, edge candidates, promoted relations, AGE nodes, and AGE edges
• a cautious technical or clinical conclusion separated from local evidence