Skill: Rebuild the Gold Index

Use when BDPM/ANSM source data has changed, Silver models were updated, or gold.embedding is missing or stale.

Steps

1. Ensure Postgres is running (pgvector/pgvector:pg16):

   docker compose up -d
   docker compose ps
   

2. Ensure Silver is up to date (run if source data changed):

   uv run dotenv -f .env run -- dbt run --project-dir dbt --profiles-dir dbt
   uv run dotenv -f .env run -- dbt test --project-dir dbt --profiles-dir dbt
   

3. Materialize the Gold asset:

   uv run dotenv -f .env run -- uv run dagster asset materialize --select gold_embeddings
   

4. Verify row counts:

   uv run dotenv -f .env run -- python -c "
   from sqlalchemy import create_engine, text
   from medox.pipeline.config_pipeline import PipelineSettings
   engine = create_engine(PipelineSettings().postgres_dsn)
   with engine.connect() as conn:
       rows = conn.execute(text('SELECT collection, count(*) FROM gold.embedding GROUP BY 1')).fetchall()
       print(rows)
   "
   

5. Smoke-test the agent tools:

   uv run dotenv -f .env run -- python -c "
   from medox.agent.tools.tool_search_drug import search_drug
   from medox.agent.tools.tool_check_interactions import check_interactions
   print(search_drug('doliprane'))
   print(check_interactions('warfarine', 'ibuprofene'))
   "
   

Collections

Collection	Content	Rebuilt from
idx_bdpm_medicament_v1	Drug denominations + metadata	silver_bdpm__medicament
idx_ansm_interaction_v1	ANSM interaction pairs	silver_ansm__interaction

Rules

• Run dbt tests before reindexing when Silver changed — bad Silver data produces bad embeddings
• A rebuild replaces one collection inside a transaction — expect several minutes of embedding
• If gold_embeddings fails, check Postgres is up and the image is pgvector/pgvector:pg16