version: "1.0.0" name: bigquery-ai description: Use when calling Gemini, TimesFM, TabFM, or pre-trained Cloud AI services from BigQuery SQL or BigFrames — generation (AI.GENERATE*, AI.COUNT_TOKENS), managed scoring/classification (AI.IF, AI.SCORE, AI.CLASSIFY, AI.AGG), embeddings and vector/hybrid search (AI.EMBED, VECTOR_SEARCH, AI.SEARCH), forecasting/prediction (AI.FORECAST, AI.PREDICT, AI.KEY_DRIVERS), and Cloud AI service models (ML.PROCESS_DOCUMENT, ML.TRANSLATE, ML.TRANSCRIBE). metadata: version: "0.8.0"

BigQuery AI (bigquery-ai)

BigQuery AI brings foundation models (Gemini, TimesFM, TabFM, EmbeddingGemma, Text/Multimodal Embeddings) and pre-trained Google Cloud AI services directly into SQL and Python (bigframes.bigquery.ai) with zero model training.

This skill packages a verified, field-tested 4-Lens reference distilled from live-executing all 29 functions and 14 composed workflows in bigquery-ai/.

Decision Tree (Lens 1 → Lens 2)

1. Do you want free-form or structured schema output per row (summaries, multi-field JSON/STRUCT extraction, typed scalars, token counting)? → see `reference/general-purpose-generation.md`
2. Do you want a managed, zero-config judgment per row or group (boolean filter AI.IF, numeric rating AI.SCORE, fixed-category label AI.CLASSIFY, or group synthesis AI.AGG)? → see `reference/managed-functions.md`
3. Do you need to embed text/multimodal content, compute pairwise similarity, or run vector/hybrid RRF search? → see `reference/embeddings-and-search.md`
4. Do you need zero-training predictions — forecast a time series (AI.FORECAST on TimesFM 2.5 / 3.0), detect anomalies (AI.DETECT_ANOMALIES), or predict tabular targets without training (AI.PREDICT + AI.EVALUATE on TabFM)? → see `reference/predictive-ai.md`
5. Do you need to explain why a summable metric moved (`AI.KEY_DRIVERS`), or estimate the causal impact of an intervention (`AI.CAUSAL_EFFECT`)? → see `reference/augmented-analytics.md` (which also routes AI.CAUSAL_EFFECT to `../bigquery-ml/reference/workflows-and-pipelines.md`)
6. Are you extracting structured records from PDFs/images or passing `ObjectRef` media into `AI.*` functions? → see `reference/document-processing.md`
7. Do you need a pre-trained Cloud AI service (`REMOTE_SERVICE_TYPE`) with a fixed API schema and per-row status column (ML.TRANSLATE, ML.UNDERSTAND_TEXT, ML.ANNOTATE_IMAGE, ML.TRANSCRIBE, ML.PROCESS_DOCUMENT)? → see `reference/cloud-ai-service-models.md`
8. Are you composing multiple capabilities into an end-to-end application (RAG, hybrid catalog search, document intelligence, log triage, content moderation, metric diagnostics)? → see `reference/workflows.md`

If the task is ambiguous between zero-training foundation functions (bigquery-ai) and trained models (bigquery-ml), consult `../bigquery-data-science/SKILL.md` first.

Tool Wrapper & Generator Rules (SQL + bigframes.bigquery.ai)

• Dual Surface & Pinned Models: Always pin the explicit endpoint or model parameter ('gemini-2.5-flash', 'TimesFM 2.5', 'TimesFM 3.0', 'embeddinggemma-300m', 'text-embedding-005', 'gemini-embedding-001') from skills/pinned-models.json so silent default upgrades never shift outputs. Pair SQL calls with bigframes.bigquery.ai / bigframes.ml.llm wrappers when generating Python notebooks.
• Per-Statement Syntax Lookup (Lens 3): Before emitting AI.*, VECTOR_SEARCH, or OBJ.* calls, inspect `syntax/INDEX.md` (or run bigquery-data-science lookup <SYMBOL>) to read the exact < 25 KB syntax card in syntax/*.md.
• Direct `ObjectRef` Input: Pass ObjectRef columns (from an Object Table's ref column, OBJ.LIST(...), or OBJ.MAKE_REF(uri, connection)) directly into AI.GENERATE*, AI.CLASSIFY, AI.SCORE, AI.IF, AI.AGG, and AI.EMBED without wrapping in OBJ.GET_ACCESS_URL.

Reviewer Verification Checklist & Rubric

Before returning any bigquery-ai SQL or Python code, verify against this checklist:

• [ ] No Retired `AI.PARSE_DOCUMENT`: Reject any query using AI.PARSE_DOCUMENT (offline since 2026-06-01); replace with ML.PROCESS_DOCUMENT or AI.GENERATE + output_schema.
• [ ] No Non-Existent `HYBRID_SEARCH` TVF: Use VECTOR_SEARCH(..., options => '{"lexical_search_columns": [...]}') or AI.SEARCH(..., mode => 'HYBRID') and size top_k >= R/10 for target semantic rank R.
• [ ] `output_schema` Field Access: When output_schema is passed to AI.GENERATE or AI.GENERATE_TABLE, select the schema field names directly (never .result, which disappears when output_schema is specified).
• [ ] Subquery Sampling Before `LIMIT`: Materialize or subquery-limit input rows before invoking per-row AI.* scalar functions so LIMIT does not evaluate and bill the entire table.
• [ ] Per-Row Status Check on Cloud AI TVFs: Queries calling ML.PROCESS_DOCUMENT, ML.TRANSLATE, ML.UNDERSTAND_TEXT, ML.ANNOTATE_IMAGE, or ML.TRANSCRIBE inspect <function>_status because per-row failures do not fail the BigQuery job.

Cross-Cutting Gotchas (Apply Across Functions)

• `AI.PARSE_DOCUMENT` is offline and its documentation has been withdrawn (since 2026-06-01). Always use ML.PROCESS_DOCUMENT (via a CLOUD_AI_DOCUMENT_V1 remote model) or AI.GENERATE over ObjectRef.
• `output_schema` replaces the `result` field entirely in AI.GENERATE and AI.GENERATE_TABLE.
• `LIMIT` / `OFFSET` on an outer `SELECT` does not reduce billed LLM calls — filter or LIMIT inside a CTE/subquery before calling scalar AI.* functions.
• Pass `ObjectRef` directly to multimodal `AI.*` functions — OBJ.GET_ACCESS_URL is only needed when generating short-lived signed URLs (<= 6 hours; never persist signed URLs).
• There is no standalone `HYBRID_SEARCH` function — hybrid RRF retrieval (distance = 1 - (1/(60 + rank_vector) + 1/(61 + rank_lexical))) is enabled via lexical_search_columns on VECTOR_SEARCH or mode => 'HYBRID' on AI.SEARCH over a 10 * top_k semantic pool.
• Pin `TimesFM 2.5` or `TimesFM 3.0` on `AI.FORECAST`, `AI.DETECT_ANOMALIES`, and `AI.EVALUATE`; materialize AI.PREDICT (TabFM) outputs because TabFM has no version pin and is non-deterministic across runs.
• Use `AI.COUNT_TOKENS` as a free pre-flight check before running large paid generative batches.
• Cloud AI service models (`REMOTE_SERVICE_TYPE`) are handles, not trained models: CREATE MODEL does not validate IAM permissions (roles/serviceusage.serviceUsageConsumer on the connection's service account is mandatory), ML.EVALUATE is rejected, and errors surface per row in <function>_status.

4-Lens Index

• Lens 2 — Use-Case Decision Guides (`reference/`):
• `reference/general-purpose-generation.md` — AI.GENERATE, AI.GENERATE_TEXT, AI.GENERATE_TABLE, AI.GENERATE_BOOL, AI.GENERATE_DOUBLE, AI.GENERATE_INT, ML.GENERATE_TEXT, AI.COUNT_TOKENS
• `reference/managed-functions.md` — AI.IF, AI.SCORE, AI.CLASSIFY, AI.AGG
• `reference/embeddings-and-search.md` — AI.EMBED, AI.GENERATE_EMBEDDING, ML.GENERATE_EMBEDDING, AI.SIMILARITY, VECTOR_SEARCH, AI.SEARCH, and hybrid RRF search
• `reference/predictive-ai.md` — AI.FORECAST, AI.DETECT_ANOMALIES, AI.PREDICT, AI.EVALUATE (TimesFM & TabFM)
• `reference/augmented-analytics.md` — AI.KEY_DRIVERS and routing for AI.CAUSAL_EFFECT
• `reference/document-processing.md` — ML.PROCESS_DOCUMENT, AI.PARSE_DOCUMENT (retired), Object Tables, OBJ.LIST, OBJ.MAKE_REF, OBJ.FETCH_METADATA, OBJ.GET_ACCESS_URL
• `reference/cloud-ai-service-models.md` — ML.TRANSLATE, ML.UNDERSTAND_TEXT, ML.ANNOTATE_IMAGE, ML.TRANSCRIBE, ML.PROCESS_DOCUMENT
• `reference/workflows.md` — 14 composed end-to-end workflows (RAG, hybrid catalog search, document intelligence, tabular prediction, etc.)
• Lens 3 — Per-Statement Syntax Cards (`syntax/`, < 25 KB each): `syntax/INDEX.md` (44 H2-sliced syntax cards)
• Lens 4 — Worked Walkthroughs (`narrative/`): 43 extracted notebook narratives in narrative/*.md (e.g., `narrative/ai_forecast.md`, `narrative/ai_generate.md`, `narrative/catalog_search.md`)