0004 — V2 LLM runs inside dbt as SQL (Pattern C)¶

Status: accepted (for V2)

LLM extraction/scoring will run as dbt SQL models using BigQuery AI functions (AI.GENERATE, AI.GENERATE_INT, AI.EMBED). Rejected: a Python pre-step (sees bronze, before silver's dedup/filter — would label everything) and dbt Python models (diverge across DuckDB/BigQuery). Reading from silver means the LLM only sees survivors; in-SQL inference bills at the batch rate.