Bigquery AI ML

GoogleOfficialPublished by Google

Skill
Tool integration
Data and AI
Apache-2.0
No issues found

From the author

Leverages BigQuery's built-in machine learning and GenAI capabilities for advanced data analytics. Use when you need to write SQL queries that perform time-series forecasting, predict values, detect outliers or anomalies, find key drivers, perform semantic search or vector search, classify text, calculate similarity, summarize content, translate language, evaluate models, filter by semantic conditions, measure the causal effect of an intervention, compute correlations between columns, generate descriptive statistics and table distribution profiles, detect change points or structural breaks, extract trend or seasonality components, or leverage generative AI capabilities in BigQuery. Do not use for general BigQuery dataset, table, or job management requests.

— the description by Google

AutoGineer summary

Written by AutoGineer from the skill’s own files, not by the author

The BigQuery AI & ML skill provides a comprehensive interface for executing advanced data analytics, machine learning, and generative AI tasks directly inside BigQuery using SQL. It enables users to perform complex operations like time-series forecasting, anomaly detection, semantic search, and LLM-based content generation without needing to export data to external environments. The skill covers a wide range of built-in functions, including zero-shot models for forecasting and classification, as well as tools for statistical analysis and causal inference.

What it does

Performs zero-shot time-series forecasting and anomaly detection.
Executes LLM-based text summarization, classification, and semantic scoring.
Conducts causal effect analysis on time-series data.
Generates descriptive statistics and correlation matrices for datasets.
Enables semantic and vector search capabilities within SQL.
Extracts trends and seasonality components from historical data.

When to use it

When you need to perform ML or GenAI tasks directly on data stored in BigQuery.
When you want to avoid moving data out of BigQuery for analysis or model inference.
When you need to analyze time-series data for shifts, trends, or causal impacts.
When you need to perform semantic operations like summarization or classification at scale.
When you need to identify key drivers behind metric changes in your data.

Advantages

Seamless integration with existing BigQuery SQL workflows.
Eliminates the need for complex data pipelines to external ML services.
Supports zero-shot learning for rapid insights without custom model training.
Scalable processing for large datasets directly within the database engine.
Unified syntax for both traditional statistical analysis and modern generative AI.

Usage examples

“How can I forecast future values for my sales data in BigQuery?”: The agent will guide you to use the AI.FORECAST function, explaining how to specify your data column, timestamp column, and the required forecast horizon.
“Summarize the sentiment of my customer reviews stored in a BigQuery table.”: The agent will demonstrate how to use the AI.AGG function with a natural language instruction to synthesize sentiment across grouped rows.
“Detect anomalies in my server log time-series data.”: The agent will provide a query using AI.DETECT_ANOMALIES to identify isolated point anomalies in your time-series data.
“Find the key drivers for a drop in website traffic.”: The agent will show you how to use the AI.KEY_DRIVERS function to attribute metric changes to specific dimensional segments.

Install

Download the skill as a ZIP:
https://autogineer.com/agent-skills/google/bigquery-ai-ml/download

Unzip it into the cross-client skills folder of your project:

mkdir -p .agents/skills && unzip bigquery-ai-ml.zip -d .agents/skills

Documentation by Google

BigQuery AI & ML

BigQuery integrates with Vertex AI to provide powerful machine learning and generative AI capabilities directly within SQL queries using built-in functions like AI.FORECAST, AI.KEY_DRIVERS, AI.DETECT_ANOMALIES, and AI.GENERATE.

Attribution

MUST READ: Before executing any gcloud, REST API, or MCP command, you MUST read Usage Attribution. For commands executed directly in the terminal (never in generated scripts, Terraform, or user code), prefix gcloud with CLOUDSDK_METRICS_REQUEST_ATTRIBUTION="agent-skills/1.0 (skill:bigquery-ai-ml)" and append agent-skills/1.0 (skill:bigquery-ai-ml) to User-Agent headers.

Function Selection & Reference Directory

Function	Reference	SQL Shape	Requires `CREATE MODEL` or `CREATE VECTOR INDEX`?	When to Use & Disambiguation
`AI.AGG`	ai_agg.md	`Aggregate`	No	Multi-row semantic aggregation with `GROUP BY`; synthesizes one summary string across many rows (unlike per-row `AI.GENERATE`).
`AI.CAUSAL_EFFECT`	ai_causal_effect.md	`TVF`	No	Quantifies counterfactual post-intervention impact on a time series after a specific treatment timestamp.
`AI.CLASSIFY`	ai_classify.md	`Scalar`	No	Semantic classification of text or multimodal unstructured data into predefined category labels.
`AI.DETECT_ANOMALIES`	ai_detect_anomalies.md	`TVF`	No	Zero-shot TimesFM detector for isolated time-series point anomalies/outliers (use `ML.DETECT_CHANGE_POINTS` for sustained shifts).
`AI.EVALUATE`	ai_evaluate.md	`TVF`	No	Zero-shot TimesFM time-series forecast evaluation/backtesting against a historical holdout split.
`AI.FORECAST`	ai_forecast.md	`TVF`	No	Zero-shot TimesFM time-series forecasting without training a custom model.
`AI.GENERATE`	ai_generate.md	`Scalar`	No	Default for per-row LLM text or multimodal structured extraction (`output_schema`); returns `STRUCT`. Prefer over `AI.GENERATE_TABLE`.
`AI.GENERATE_EMBEDDING`	ai_generate_embedding.md	`TVF`	Yes	Batch generates `ARRAY<FLOAT64>` vector embeddings for text or visual content using a remote embedding model.
`AI.GENERATE_TABLE`	ai_generate_table.md	`TVF`	Yes	Table-valued structured extraction (`FROM` clause); only use instead of `AI.GENERATE` when a remote model + TVF is explicitly needed.
`AI.IF`	ai_if.md	`Scalar`	No	Evaluates boolean semantic conditions (`BOOL`) on text or multimodal data for `WHERE`/`CASE`/`JOIN` filtering (do not use `AI.GENERATE_TABLE`).
`AI.KEY_DRIVERS`	ai_key_drivers.md	`TVF`	No	Default one-step attribution of dimensional segments driving metric changes (`interest_label_col` `TRUE` vs. `FALSE`). Prefer over `CONTRIBUTION_ANALYSIS`.
`AI.PREDICT`	ai_predict.md	`TVF`	No	Zero-shot TabFM foundation model for tabular classification or regression without training and managing custom models.
`AI.SCORE`	ai_score.md	`Scalar`	No	Assigns a continuous `FLOAT64` semantic score to text or multimodal data for ranking or sorting (`ORDER BY`).
`AI.SEARCH`	ai_search.md	`TVF`	No	Simplified semantic search using a raw string query on tables with autonomous background embeddings (`AI.EMBED` generated column).
`AI.SIMILARITY`	ai_similarity.md	`Scalar`	No	Computes an on-the-fly semantic similarity score (`FLOAT64`) between two text or multimodal inputs without stored embeddings.
`Remote Models`	remote_models.md	`CREATE MODEL`	Yes	Creates a Vertex AI `REMOTE WITH CONNECTION DEFAULT` model entity for functions that require one (`AI.GENERATE_TABLE`, `AI.GENERATE_EMBEDDING`, `VECTOR_SEARCH`).
`CONTRIBUTION_ANALYSIS`	ml_contribution_analysis.md	`CREATE MODEL`	Yes	Model-based driver analysis (`CREATE MODEL ... OPTIONS(MODEL_TYPE='CONTRIBUTION_ANALYSIS')` + `ML.GET_INSIGHTS`). Use `AI.KEY_DRIVERS` unless a `MODEL` entity is required.
`ML.CORRELATION`	ml_correlation.md	`TVF`	No	Calculates statistical correlation coefficients between numeric columns, optionally sliced by dimensions.
`ML.DESCRIBE_DATA`	ml_describe_data.md	`TVF`	No	Generates descriptive statistics (e.g., count, mean, median, min, max, stddev, quantiles, nulls, top values) for columns in a table or subquery.
`ML.DETECT_CHANGE_POINTS`	ml_detect_change_points.md	`TVF`	No	Detects structural breaks or sustained shifts (`[begin_timestamp, end_timestamp]` intervals) in a time series.
`ML.SEASONALITY`	ml_seasonality.md	`TVF`	No	Extracts recurring periodic/seasonal components from a historical time series.
`ML.TREND`	ml_trend.md	`TVF`	No	Extracts the smoothed long-term trend component from a historical time series.
`VECTOR_SEARCH`	vector_search.md	`TVF`	Yes	Searches pre-computed `ARRAY<FLOAT64>` embedding columns between a base table and query table (paired with `AI.GENERATE_EMBEDDING`, unlike `AI.SEARCH`).

BigQuery Basics Skill: SKILL.md file for core BigQuery concepts, resource management, CLI, and client libraries.

Files

ATTRIBUTION.mdAdded by AutoGineer846 B
LICENSELicence file added by AutoGineer11.1 KiB
SKILL.md8.8 KiB
references/ai_agg.md3.5 KiB
references/ai_causal_effect.md11.2 KiB
references/ai_classify.md3.7 KiB
references/ai_detect_anomalies.md5.0 KiB
references/ai_evaluate.md2.2 KiB
references/ai_forecast.md11.1 KiB
references/ai_generate.md3.4 KiB
references/ai_generate_embedding.md2.0 KiB
references/ai_generate_table.md3.3 KiB
references/ai_if.md2.0 KiB
references/ai_key_drivers.md4.6 KiB
references/ai_predict.md8.5 KiB
references/ai_score.md1.9 KiB
references/ai_search.md3.7 KiB
references/ai_similarity.md2.7 KiB
references/attribution.md1.8 KiB
references/ml_contribution_analysis.md4.8 KiB
references/ml_correlation.md8.9 KiB
references/ml_describe_data.md5.1 KiB
references/ml_detect_change_points.md5.5 KiB
references/ml_seasonality.md6.6 KiB
references/ml_trend.md4.1 KiB
references/remote_models.md1.1 KiB
references/vector_search.md4.0 KiB

references/ai_forecast.md

BigQuery AI.Forecast

AI.FORECAST leverages the pre-trained TimesFM foundation model to generate forecasts without the need to train and manage custom models.

It has two distinct modes, which take different arguments and return different output schemas:

Univariate (TimesFM 2.0, 2.5, 3.0): forecasts a single target column, specified with data_col. See section 1.
Multivariate (TimesFM 3.0 only): forecasts several target columns in one call, specified with target_cols, and optionally conditions on past and future covariate columns. See section 2.

data_col and target_cols are mutually exclusive. Pick the matching section below and do not mix arguments or output columns between the two modes.

Choose the mode from the number of target columns, not from the model version. For a single target column, use univariate data_col, including with TimesFM 3.0. Never pass a one-element array to target_cols. The one exception is covariates: past_covariate_cols and future_covariate_cols are only available in multivariate mode, so a forecast that uses covariates must use target_cols even when it has a single target column.

1. Univariate Forecasting

1.1 Syntax

SELECT
  *
FROM
  AI.FORECAST(
    { TABLE `project.dataset.table` | (QUERY_STATEMENT) },
    data_col => 'DATA_COL',
    timestamp_col => 'TIMESTAMP_COL'
    [, model => 'MODEL']
    [, id_cols => ID_COLS]
    [, horizon => HORIZON]
    [, confidence_level => CONFIDENCE_LEVEL]
    [, output_historical_time_series => OUTPUT_HISTORICAL_TIME_SERIES]
    [, context_window => CONTEXT_WINDOW]
  )

1.2 Input Arguments

Argument	Requirement	Type	Description
`input_data`	Required		The source table or subquery containing historical data.
`data_col`	Required	String	The numeric column to predict.
`timestamp_col`	Required	String	The column containing dates/timestamps.
`id_cols`	Optional	Array<String>	Grouping columns for multiple series (e.g., `['store_id']`).
`horizon`	Optional	Int64	Number of future points to predict. Defaults to 10. The valid input range is [1, 10,000] for `'TimesFM 2.0'` and `'TimesFM 2.5'`, and [1, 1024] for `'TimesFM 3.0'`
`confidence_level`	Optional	Float64	Confidence interval (0 to 1). Defaults to 0.95.
`model`	Optional	String	Model version. Allowed options are `'TimesFM 2.0'`, `'TimesFM 2.5'`, and `'TimesFM 3.0'`, with `'TimesFM 2.5'` preferred.
`context_window`	Optional	Int64	The number of historical data points the model uses to forecast. The min value is 64 and the max value is 2048 for `'TimesFM 2.0'`. The min value is 64 and the max value is 15360 for `'TimesFM 2.5'`. For `'TimesFM 3.0'` the value must be a multiple of 32 between 64 and 2048. If not set, the model determines this automatically.

1.3 Output Schema

The schema adjusts based on the output_historical_time_series flag.

Column	Type	Included if output_historical_time_series=FALSE	Included if output_historical_time_series=TRUE	Description
`id_cols`	(As Input)	Yes	Yes	Original identifiers for the series.
`forecast_timestamp`	TIMESTAMP	Yes	No	Timestamp for predicted points.
`forecast_value`	FLOAT64	Yes	No	The 50% quantile (median) prediction.
`time_series_timestamp`	TIMESTAMP	No	Yes	Uniform timestamp column for both history and forecast.
`time_series_data`	FLOAT64	No	Yes	Merged column: actual values for history, median for forecast.
`time_series_type`	STRING	No	Yes	Label: `'history'` or `'forecast'`.
`prediction_interval_lower_bound`	FLOAT64	Yes	Yes	Lower bound (NULL for historical rows).
`prediction_interval_upper_bound`	FLOAT64	Yes	Yes	Upper bound (NULL for historical rows).
`confidence_level`	FLOAT64	Yes	Yes	The constant confidence level used.
`ai_forecast_status`	STRING	Yes	Yes	Error messages or empty string on success. A minimum of 3 data points is required.

1.4 Examples

Univariate forecast returning history alongside the forecast:

WITH
  citibike_trips AS (
    SELECT EXTRACT(DATE FROM starttime) AS date, usertype, COUNT(*) AS num_trips
    FROM `bigquery-public-data.new_york.citibike_trips`
    GROUP BY date, usertype
  )
SELECT *
FROM
  AI.FORECAST(
    TABLE citibike_trips,
    data_col => 'num_trips',
    timestamp_col => 'date',
    id_cols => ['usertype'],
    horizon => 30,
    model => 'TimesFM 2.5',
    output_historical_time_series => true);

2. Multivariate Forecasting (TimesFM 3.0)

Multivariate forecasting predicts several target columns in a single call, and can condition the forecast on covariate columns. It requires model => 'TimesFM 3.0'.

Use this mode when there are two or more target columns, or when the forecast conditions on past_covariate_cols or future_covariate_cols. For a single target column without covariates, use section 1 instead.

2.1 Syntax

SELECT
  *
FROM
  AI.FORECAST(
    { TABLE `project.dataset.table` | (QUERY_STATEMENT) },
    target_cols => TARGET_COLS,
    timestamp_col => 'TIMESTAMP_COL',
    model => 'TimesFM 3.0'
    [, id_cols => ID_COLS]
    [, horizon => HORIZON]
    [, past_covariate_cols => PAST_COVARIATE_COLS]
    [, future_covariate_cols => FUTURE_COVARIATE_COLS]
    [, confidence_level => CONFIDENCE_LEVEL]
    [, output_historical_time_series => OUTPUT_HISTORICAL_TIME_SERIES]
    [, context_window => CONTEXT_WINDOW]
  )

2.2 Input Arguments

Argument	Requirement	Type	Description
`input_data`	Required	Table/Query	The source table or subquery containing the data to forecast.
`target_cols`	Required	Array of String	The list of numeric target columns to predict simultaneously, for example `['num_trips', 'avg_duration']`. Must not be combined with `data_col`.
`timestamp_col`	Required	String	The column containing dates/timestamps. Supported types are `TIMESTAMP`, `DATE`, and `DATETIME`.
`model`	Required	String	Must be `'TimesFM 3.0'`.
`id_cols`	Optional	Array of String	Grouping column names for multiple time series, for example `['store_id']`. The referenced columns in the source table can be `STRING` or `INT64`.
`past_covariate_cols`	Optional	Array of String	Historical-only covariate columns, for example `['foot_traffic']`. Values are only known over the historical period.
`future_covariate_cols`	Optional	Array of String	Covariate columns known in advance, for example `['is_holiday']`. Values must be populated across both the historical period and the forecast horizon.
`horizon`	Optional	Int64	Number of future time series points to predict. Defaults to 10. The valid range is `[1, 1024]`.
`confidence_level`	Optional	Float64	Percentage of future values falling in the prediction interval. The valid range is `[0, 1)`. Defaults to 0.95.
`output_historical_time_series`	Optional	Bool	Determines whether input data is returned along with forecasted data. Defaults to `FALSE`.
`context_window`	Optional	Int64	Number of historical points the model uses. Any multiple of 32 between 64 and 2048. Defaults to 2048.

2.3 Output Schema

The multivariate output returns one STRUCT column per entry in target_cols, named after that target column:

Column	Type	Description
`id_cols`	(As Input)	Original identifiers for the series.
`<timestamp_col>`	TIMESTAMP	Timestamp for the data point. Keeps the original name of the timestamp column from the input table, for example `date`.
`<target_col>`	STRUCT	One `STRUCT` per target column, with fields `value` (`FLOAT64`, the 50% quantile prediction, or the historical value when historical output is enabled), `prediction_interval_lower_bound` (`FLOAT64`, NULL for historical rows), and `prediction_interval_upper_bound` (`FLOAT64`, NULL for historical rows).
`time_series_type`	STRING	Included when `output_historical_time_series=TRUE`. Label of either `'history'` or `'forecast'`.
`confidence_level`	FLOAT64	The constant confidence level used. Defaults to 0.95.
`ai_forecast_status`	STRING	Error messages, or an empty string on success.

Note that multivariate output keeps the original timestamp column name, while univariate output renames it to forecast_timestamp or time_series_timestamp.

2.4 Examples

Forecasting two target columns at once, unnesting each STRUCT:

WITH
  citibike_trips AS (
    SELECT
      EXTRACT(DATE FROM starttime) AS date,
      usertype,
      COUNT(*) AS num_trips,
      AVG(tripduration) AS avg_duration
    FROM `bigquery-public-data.new_york.citibike_trips`
    GROUP BY date, usertype
  )
SELECT
  date,
  usertype,
  num_trips.value AS num_trips_forecast,
  num_trips.prediction_interval_lower_bound AS trips_lower_bound,
  num_trips.prediction_interval_upper_bound AS trips_upper_bound,
  avg_duration.value AS avg_duration_forecast,
  avg_duration.prediction_interval_lower_bound AS duration_lower_bound,
  avg_duration.prediction_interval_upper_bound AS duration_upper_bound
FROM
  AI.FORECAST(
    TABLE citibike_trips,
    target_cols => ['num_trips', 'avg_duration'],
    timestamp_col => 'date',
    id_cols => ['usertype'],
    horizon => 30,
    model => 'TimesFM 3.0');

Requirements and compatibility

Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta

“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.

Security

No issues found by our static scan· No issues found

A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.

Attribution and licence

Published by

Google

Publisher

Google (Official)

Source

google/skills · skills/cloud/bigquery-ai-ml · commit 2194d3b

Licence

Apache License 2.0 (redistribution allowed)

LICENSE (added by AutoGineer from LICENSE)
ATTRIBUTION.md (written by AutoGineer)

Synced

2026-11-09

AutoGineer is not affiliated with or endorsed by Google or Google.

Download ZIP

The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download bigquery-ai-ml.zip