Sample Page Title

Sponsored Content material

Google Cloud

Introduction

Enterprises handle a mixture of structured knowledge in organized tables and a rising quantity of unstructured knowledge like photos, audio, and paperwork. Analyzing these numerous knowledge varieties collectively is historically advanced, as they usually require separate instruments. Unstructured media usually requires exports to specialised companies for processing (e.g. a pc imaginative and prescient service for picture evaluation, or a speech-to-text engine for audio), which creates knowledge silos and hinders a holistic analytical view.

Think about a fictional e-commerce help system: structured ticket particulars reside in a BigQuery desk, whereas corresponding help name recordings or images of broken merchandise reside in cloud object shops. And not using a direct hyperlink, answering a context-rich query like “determine all help tickets for a selected laptop computer mannequin the place name audio signifies excessive buyer frustration and the photograph reveals a cracked display screen“ is a cumbersome, multi-step course of.

This text is a sensible, technical information to ObjectRef in BigQuery, a function designed to unify this evaluation. We’ll discover learn how to construct, question, and govern multimodal datasets, enabling complete insights utilizing acquainted SQL and Python interfaces.

Half 1: ObjectRef – The Key to Unifying Multimodal Knowledge

ObjectRef Construction and Perform

To deal with the problem of siloed knowledge, BigQuery introduces ObjectRef, a specialised STRUCT knowledge kind. An ObjectRef acts as a direct reference to an unstructured knowledge object saved in Google Cloud Storage (GCS). It doesn’t include the unstructured knowledge itself (e.g. a base64 encoded picture in a database, or a transcribed audio); as a substitute, it factors to the placement of that knowledge, permitting BigQuery to entry and incorporate it into queries for evaluation.

The ObjectRef STRUCT consists of a number of key fields:

• uri (STRING): a GCS path to an object
• authorizer (STRING): permits BigQuery to securely entry GCS objects
• model (STRING): shops the particular Era ID of a GCS object, locking the reference to a exact model for reproducible evaluation
• particulars (JSON): a JSON aspect that usually comprises GCS metadata like contentType or measurement

Here’s a JSON illustration of an ObjectRef worth:

JSON

{
  "uri": "gs://cymbal-support/calls/ticket-83729.mp3",
  "model": 1742790939895861,
  "authorizer": "my-project.us-central1.conn",
  "particulars": {
    "gcs_metadata": {
      "content_type": "audio/mp3",
      "md5_hash": "a1b2c3d5g5f67890a1b2c3d4e5e47890",
      "measurement": 5120000,
      "up to date": 1742790939903000
    }
  }
}

By encapsulating this data, an ObjectRef gives BigQuery with all the mandatory particulars to find, securely entry, and perceive the essential properties of an unstructured file in GCS. This types the inspiration for constructing multimodal tables and dataframes, permitting structured knowledge to reside side-by-side with references to unstructured content material.

Create Multimodal Tables

A multimodal desk is a normal BigQuery desk that features a number of ObjectRef columns. This part covers learn how to create these tables and populate them with SQL.

You possibly can outline ObjectRef columns when creating a brand new desk or add them to current tables. This flexibility means that you can adapt your present knowledge fashions to reap the benefits of multimodal capabilities.

Creating an ObjectRef Column with Object Tables

When you’ve got many recordsdata saved in a GCS bucket, an object desk is an environment friendly technique to generate ObjectRefs. An object desk is a read-only desk that shows the contents of a GCS listing and mechanically features a column named ref, of kind ObjectRef.

SQL

CREATE EXTERNAL TABLE `project_id.dataset_id.my_table`
WITH CONNECTION `project_id.area.connection_id`
OPTIONS(
  object_metadata="SIMPLE",
  uris = ['gs://bucket-name/path/*.jpg']
);

The output is a brand new desk containing a ref column. You should utilize the ref column with features like AI.GENERATE or be part of it to different tables.

Programmatically Establishing ObjectRefs

For extra dynamic workflows, you possibly can create ObjectRefs programmatically utilizing the OBJ.MAKE_REF() perform. It’s widespread to wrap this perform in OBJ.FETCH_METADATA() to populate the particulars aspect with GCS metadata. The next code additionally works when you change the gs:// path with a URI area in an current desk.

SQL

SELECT 
OBJ.FETCH_METADATA(OBJ.MAKE_REF('gs://my-bucket/path/picture.jpg', 'us-central1.conn')) AS customer_image_ref,
OBJ.FETCH_METADATA(OBJ.MAKE_REF('gs://my-bucket/path/name.mp3', 'us-central1.conn')) AS support_call_ref

By utilizing both Object Tables or OBJ.MAKE_REF, you possibly can construct and preserve multimodal tables, setting the stage for built-in analytics.

Half 2: Multimodal Tables with SQL

Safe and Ruled Entry

ObjectRef integrates with BigQuery’s native security measures, enabling governance over your multimodal knowledge. Entry to underlying GCS objects just isn’t granted to the end-user straight. As an alternative, it’s delegated by a BigQuery connection useful resource specified within the ObjectRef’s authorizer area. This mannequin permits for a number of layers of safety.

Think about the next multimodal desk, which shops details about product photos for our e-commerce retailer. The desk contains an ObjectRef column named picture.

Column-level safety: limit entry to whole columns. For a set of customers who ought to solely analyze product names and rankings, an administrator can apply column-level safety to the picture column. This disallows these analysts from choosing the picture column whereas nonetheless permitting evaluation of different structured fields.

Row-level safety: BigQuery permits for filtering which rows a consumer can see based mostly on outlined guidelines. A row-level coverage might limit entry based mostly on a consumer’s position. For instance, a coverage may state “Don’t permit customers to question merchandise associated to canine”, which filters out these rows from question outcomes as in the event that they don’t exist.

A number of Authorizers: this desk makes use of two completely different connections within the picture.authorizer aspect (conn1 and conn2).

This permits an administrator to handle GCS permissions centrally by connections. For example, conn1 may entry a public picture bucket, whereas conn2 accesses a restricted bucket with new product designs. Even when a consumer can see all rows, their means to question the underlying file for the “Fowl Seed” product relies upon fully on whether or not they have permission to make use of the extra privileged conn2 connection.

AI-Pushed Inference with SQL

The AI.GENERATE_TABLE perform creates a brand new, structured desk by making use of a generative AI mannequin to your multimodal knowledge. That is best for knowledge enrichment duties at scale. Let’s use our e-commerce instance to create website positioning key phrases and a brief advertising and marketing description for every product, utilizing its identify and picture as supply materials.

The next question processes the merchandise desk, taking the product_name and picture ObjectRef as inputs. It generates a brand new desk containing the unique product_id, an inventory of website positioning key phrases, and a product description.

SQL 

SELECT
  product_id,
  seo_keywords,
  product_description
FROM AI.GENERATE_TABLE(
  MODEL `dataset_id.gemini`, (
    SELECT (
		'For the picture of a pet product, generate:'
            '1) 5 website positioning search key phrases and' 
            '2) A one sentence product description', 
            product_name, image_ref) AS immediate,
            product_id
    FROM `dataset_id.products_multimodal_table`
  ),
  STRUCT(
     "seo_keywords ARRAY, product_description STRING" AS output_schema
  )
);

The result’s a brand new structured desk with the columns product_id, seo_keywords, and product_description. This automates a time-consuming advertising and marketing activity and produces ready-to-use knowledge that may be loaded straight right into a content material administration system or used for additional evaluation.

Half 3: Multimodal DataFrames with Python

Bridging Python and BigQuery for Multimodal Inference

Python is the language of selection for a lot of knowledge scientists and knowledge analysts. However practitioners generally run into points when their knowledge is just too giant to suit into the reminiscence of an area machine.

BigQuery DataFrames gives an answer. It affords a pandas-like API to work together with knowledge saved in BigQuery with out ever pulling it into native reminiscence. The library interprets Python code into SQL that’s pushed down and executed on BigQuery’s extremely scalable engine. This gives the acquainted syntax of a preferred Python library mixed with the ability of BigQuery.

This naturally extends to multimodal analytics. A BigQuery DataFrame can signify each your structured knowledge and references to unstructured recordsdata, collectively in a single multimodal dataframe. This lets you load, rework, and analyze dataframes containing each your structured metadata and tips to unstructured recordsdata, inside a single Python surroundings.

Create Multimodal DataFrames

After getting the bigframes library put in, you possibly can start working with multimodal knowledge. The important thing idea is the blob column: a particular column that holds references to unstructured recordsdata in GCS. Consider a blob column because the Python illustration of an ObjectRef – it doesn’t maintain the file itself, however factors to it and gives strategies to work together with it.

There are three widespread methods to create or designate a blob column:

PYTHON

import bigframes
import bigframes.pandas as bpd

# 1. Create blob columns from a GCS location
df = bpd.from_glob_path(  "gs://cloud-samples-data/bigquery/tutorials/cymbal-pets/photos/*", identify="picture")

# 2. From an current object desk
df = bpd.read_gbq_object_table("", identify="blob_col")

# 3. From a dataframe with a URI area
df["blob_col"] = df["uri"].str.to_blob()

To clarify the approaches above:

1. A GCS location: Use from_glob_path to scan a GCS bucket. Behind the scenes, this operation creates a brief BigQuery object desk, and presents it as a DataFrame with a ready-to-use blob column.
2. An current object desk: if you have already got a BigQuery object desk, use the read_gbq_object_table perform to load it. This reads the prevailing desk with no need to re-scan GCS.
3. An current dataframe: you probably have a BigQuery DataFrame that comprises a column of STRING GCS URIs, merely use the .str.to_blob() technique on that column to “improve” it to a blob column.

AI-Pushed Inference with Python

The first profit of making a multimodal dataframe is to carry out AI-driven evaluation straight in your unstructured knowledge at scale. BigQuery DataFrames means that you can apply giant language fashions (LLMs) to your knowledge, together with any blob columns.

The final workflow entails three steps:

1. Create a multimodal dataframe with a blob column pointing to unstructured recordsdata
2. Load a pre-existing BigQuery ML mannequin right into a BigFrames mannequin object
3. Name the .predict() technique on the mannequin object, passing your multimodal dataframe as enter.

Let’s proceed with the e-commerce instance. We’ll use the gemini-2.5-flash mannequin to generate a quick description for every pet product picture.

PYTHON

import bigframes.pandas as bpd

# 1. Create the multimodal dataframe from a GCS location
df = bpd.from_glob_path(
"gs://cloud-samples-data/bigquery/tutorials/cymbal-pets/photos/*", identify="image_blob")


# Restrict to 2 photos for simplicity
df = df.head(2)

# 2. Specify a big language mannequin
from bigframes.ml import llm


mannequin = llm.GeminiTextGenerator(model_name="gemini-2.5-flash-preview-05-20")

# 3. Ask the LLM to explain what's within the image

reply = mannequin.predict(df_image, immediate=["Write a 1 sentence product description for the image.", df_image["image"]])

reply[["ml_generate_text_llm_result", "image"]]

While you name mannequin.predict(df_image), BigQuery DataFrames constructs and executes a SQL question utilizing the ML.GENERATE_TEXT perform, mechanically passing file references from the blob column and the textual content immediate as inputs. The BigQuery engine processes this request, sends the info to a Gemini mannequin, and returns the generated textual content descriptions to a brand new column within the ensuing DataFrame.

This highly effective integration means that you can carry out multimodal evaluation throughout hundreds or thousands and thousands of recordsdata utilizing just some strains of Python code.

Going Deeper with Multimodal DataFrames

Along with utilizing LLMs for technology, the bigframes library affords a rising set of instruments designed to course of and analyze unstructured knowledge. Key capabilities out there with the blob column and its associated strategies embody:

• Constructed-in Transformations: put together photos for modeling with native transformations for widespread operations like blurring, normalizing, and resizing at scale.
• Embedding Era: allow semantic search by producing embeddings from multimodal knowledge, utilizing Vertex AI-hosted fashions to transform knowledge into embeddings in a single perform name.
• PDF Chunking: streamline RAG workflows by programmatically splitting doc content material into smaller, significant segments – a typical pre-processing step.

These options sign that BigQuery DataFrames is being constructed as an end-to-end software for multimodal analytics and AI with Python. As improvement continues, you possibly can anticipate to see extra instruments historically present in separate, specialised libraries straight built-in into bigframes.

Conclusion:

Multimodal tables and dataframes signify a shift in how organizations can strategy knowledge analytics. By making a direct, safe hyperlink between tabular knowledge and unstructured recordsdata in GCS, BigQuery dismantles the info silos which have lengthy difficult multimodal evaluation.

This information demonstrates that whether or not you’re an information analyst writing SQL, or an information scientist utilizing Python, you now have the flexibility to elegantly analyze arbitrary multimodal recordsdata alongside relational knowledge with ease.

To start constructing your personal multimodal analytics options, discover the next assets:

1. Official documentation: learn an summary on learn how to analyze multimodal knowledge in BigQuery
2. Python Pocket book: get hands-on with a BigQuery DataFrames instance pocket book
3. Step-by-step tutorials:

Writer: Jeff Nelson, Google Cloud – Developer Relations Engineer