AI news

BigQuery AI Functions: A New Chapter in the AI Era of SQL

Honghao Wang

13 Nov 2025 — 3 min read

Introduction

For decades, SQL has been the universal language for data analysis, especially for structured data.

Now, Large Language Models (LLMs) such as Gemini bring nuanced insights to unstructured data — text, images, and videos.

Historically, integrating LLMs into SQL workflows meant moving data, tuning prompts/parameters, and incurring high scaling costs — limiting adoption for many data practitioners.

---

BigQuery-Managed AI Functions (Public Preview)

We’re excited to announce the public preview of BigQuery-managed AI functions — a set of SQL-native capabilities for the AI era.

Available functions:

`AI.IF` – Semantic filtering and joining
`AI.CLASSIFY` – Categorization of text or images
`AI.SCORE` – Natural language ranking

Key advantages:

Direct integration into SQL queries — no complex prompt engineering
Optimized per use case — no manual model selection or parameter tuning
Lower costs through intelligent prompt processing and query execution

---

What You Can Do

Semantic filter/join:
Use `AI.IF` in `WHERE` or `ON` clauses
Categorize data:
Use `AI.CLASSIFY` in `GROUP BY` clauses
Rank by criteria:
Use `AI.SCORE` in `ORDER BY` clauses

These functions make AI-powered workflows seamless for analysts and engineers.

---

Example Use Case: Content Creation and Publishing

Platforms like AiToEarn官网 complement BigQuery AI functions by combining:

AI generation
Analytics
Multi-platform publishing
(e.g., Douyin, Kwai, WeChat, Bilibili, Rednote/Xiaohongshu, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, X/Twitter)

This pairing lets users convert AI insights into shareable, monetizable content.

---

Function Deep Dive

AI.IF — Semantic Filtering & Joining

Purpose: Filter or join data based on conditions written in natural language.

Use cases:

Detect negative customer reviews
Find images with certain attributes
Extract relevant document content

Optimization: BigQuery evaluates non-AI filters first to reduce LLM calls.

Example — Semantic Filtering:

SELECT title, body 
FROM bigquery-public-data.bbc_news.fulltext 
WHERE AI.IF(("The news is related to Google, news: ", body),   
             connection_id => "us.test_connection")
      AND category = "tech"  -- Non-AI filter evaluated first

Example — Semantic Join for Entity Resolution:

WITH product_catalog_A AS (
    SELECT "Veridia AquaSource Hydrating Shampoo" AS product
    UNION ALL 
    SELECT "Veridia Full-Lift Volumizing Shampoo"
),
product_catalog_B AS (
    SELECT "Veridia Shampoo, AquaSource Hydration" AS product
)
SELECT *
FROM product_catalog_A a 
JOIN product_catalog_B b
ON AI.IF((a.product, " is the same product as ", b.product),
    connection_id => "us.test_connection")

---

AI.CLASSIFY — Data Classification

Purpose: Categorize text or images into user-specified labels.

Use cases:

Route support tickets by topic
Classify images by style
Categorize articles by subject

Example — Content Categorization:

SELECT
  AI.CLASSIFY(
    body, 
    categories => ['tech', 'sport', 'business', 'politics', 'entertainment'],
    connection_id => 'us.test_connection') AS category,
  COUNT(*) num_articles
FROM bigquery-public-data.bbc_news.fulltext
GROUP BY category;

---

AI.SCORE — Semantic Ranking

Purpose: Rank rows using natural language criteria.

Automated benefit:

BigQuery refines prompts into a structured scoring rubric for higher quality results.

Example — Ranking Top Positive Reviews:

SELECT
  review,
  AI.SCORE(("From 1 to 10, rate how much does the reviewer like the movie :", review),
           connection_id => 'us.test_connection') AS ai_rating,
  reviewer_rating AS human_rating
FROM bigquery-public-data.imdb.reviews
WHERE title = 'Movie'
ORDER BY ai_rating DESC
LIMIT 10;

---

Built-In Optimizations

BigQuery-managed AI functions include:

Prompt Optimization
Transforms prompts for Gemini to boost quality and consistency
Improves cache hit rates
Query Plan Optimization
Reduces LLM calls by reordering filters and joins
Model Endpoint & Parameter Tuning
Improves quality and consistency across executions

---

Getting Started

These functions complement general-purpose Gemini inference functions like `AI.GENERATE`.

Early Access: Sign up here

Documentation:

Overview
Python usage in notebook
and API reference

For feedback: bqml-feedback@google.com

---

Choosing the Right Function

Use Managed AI Functions (`AI.IF`, `AI.CLASSIFY`, `AI.SCORE`) when:
You want optimized cost + quality
You don’t need custom model selection
Use `AI.GENERATE` when:
You require custom prompts or additional control
You want to choose from a range of LLMs

---

Complementary Tools for Publishing

If you want to build scalable AI-powered workflows and publish across platforms:

Try AiToEarn官网 — open-source AI content monetization platform
Integrates AI generation, multi-platform publishing, analytics, and AI model ranking (AI模型排名)
Supports Douyin, Kwai, WeChat, Bilibili, Rednote/Xiaohongshu, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, X/Twitter

---

Would you like me to create a side-by-side visual cheat sheet showing when to use each BigQuery AI function (`AI.IF`, `AI.CLASSIFY`, `AI.SCORE`, `AI.GENERATE`) so it’s quick to decide in real workflows? That could make this even easier to use.