To open your data in AI Platform: productive multimodal analytics AI

Supported content

Traditional data platforms have long-term approaches in tabar data – think that West Region sold last quarter? “This basic basis is stronger. But with the increasing volume and significance of multimodal data (eg photographs, noises, Semantic Extes questions for relying on a significant bottle.

Consider the general form of e-commerce: “Identify the electronic products with higher rates connected to customer pictures showing signs of damage to the damage when you arrive.” Historically, this meant using SQL to get a formal product data, sending photos into a separate ML pipeline to analyze, and finally try to combse different results. A number of steps, time-consuming time when AI is divided into the dataflow rather than combined with nature in the analysis.

Think of coping with this work – to compile formal data with understanding from the random checkpoint – using one SQL statement. This jump is possible by combinating Airative Ai directly in the mind of today's data speaker. Introduce the new era in which a complex analyzement, multimodal analysis can be done by a regular SQL.

Let us examine how AI produced can recycle data platforms and allow doctors to bring multilontal understanding with SQL proatility.

Related algebra meets Generative Ai

Traditional information for traditional information gets their power from the related algebra. This provides a statistical and consistent framework for informed, tabar, magnificent information when schemas are properly defined.

But multimodal data contains richest decorative content algebra, only one, cannot directly translate. AI of AI of AI as a Semantic bridge. This enables the questions that affect the power of AI recommend complex signals included in multimodal information, allowing to think as well as people, thus passing issues of traditional data and SQL operations.

To fully grasp this, evolution, let us first examine the skills that give these skills.

Action Ai is active

Modern data on AI platforms allow businesses to participate with the data by motivating ai generating AI products. Instead of ETL pipes in external services, tasks such as Bombury's AI.GENERATE including AI.GENERATE_TABLE Allow users to update large powerful languages ​​(LLMS) using the standard SQL. These activities include data on the table available, and the Prompt described by the user, in the LLM, returns the answer.

Average analysis

Think of the e-commerce business with a table containing millions of product reviews in all thousands of things. The analysis of the hands and volume is to understand the customer opinion of time prevents time. Instead, AI services can automatically remove the themes of the key to each review and produce short summaries. These summaries can provide potential and discreet customers.

Multimodal analysis

And these tasks are easier beyond the non-tabar data. Modern llms can disagree from multimodal data. This data usually lasts in the cloud stores such as Google Cloud Greed (GCS). BriIquery makes access to these items with ObjectRef. ObjectRef Columns live within common tables that begin and safely to the GCS for analysis.

Consider opportunities to combine formal and mysterious data for e-commerce:

• Picture all calls for sale by 2024 for general customer complaints of “Bluetooth Manging Apperations” and the user reference to the product brochure (PDF) to see that resolutions are lost.
• List the Western-Western district carriers by analyzing photos moved by customers to display travel related injury.

Dealing with the details depending on the external file analysis next to the formal table information, the use of the Greesery ObjectRef. Let's see how ObjectRef Improves the standard BioBquery table. Consider a basic product table:

We can easily add an ObjectRef the column named manuals In this example, referring to the PDF official PDF is stored in GCS. This allows ObjectRef Living along the formal data:

This meeting of integration of multimodal complexity. Let's look at the example where we produced by pairs using customer reviews and product documents (PDF):

SQL 

SELECT
product_id,
product_name,
question_answer
FROM
  AI.GENERATE_TABLE(
    MODEL `my_dataset.gemini`,
    (SELECT product_id, product_name,
    ('Use reviews and product manual PDF to generate common question/answers',
    customer_reviews, 
    manuals
    ) AS prompt, 
    FROM `my_dataset.reviews_multimodal`
    ),
  STRUCT("question_answer ARRAY" AS output_schema)
);

Quick argument for AI.GENERATE_TABLE In this question, the top three inputs:

• Textual instruction in the model to generate regular questions
• This page customer_reviews Column (string with a compound text of text)
• This page manuals ObjectRef Column, to link directly to product PDF

Work using a fixed text column including The sub-PDF is stored in GCS to perform AI function. To exit the Q & A set of pairs who help can be better meaning product:

Expanding WESSREF USE

We can easily put multimodal deries easily by adding more ObjectRef columns on our table. To continue the E-commerce status, we include an ObjectRef Column is called product_imagewhich means a form of the official product set out in the website.

And from ObjectRefS Types of buildings, supporting houses. This is a powerful force in conditions where one main record is related to many unplanned things. For example, a customer_images Column can be a type of ObjectRefSix, then, identifies the image of the secure product for customers stored in GCS.

This skill has modulated the relationships to one and to-to-to-to-to-a formal records and various random materials (within a biobquery and using the previous external instruments that require the previous international tools.

The functions of a type

AI.GENERATE Jobs provide flexibility in describing Output Schems, but with regular analytical activities that require stronger-painted results, Bechery provides specialized AI services. These activities can analyze the text or ObjectRefS with a llm and restore feedback as a adful directly to Bombenery.

Here are a few examples:

• AI.Generae_Bool: Applying input (text or back items) and returns the bool value, which is useful in emotional analysis or any true / false determination.
• Ai.genise_int: Returns the full amount, it is useful for issuing the number of numbers, ratings, or basic qualifications based on the characters from the data.
• AI.Genaerae_DUBLE: Returns a floating point number, useful in producing scores, ratings, or financial rates.

The primary benefit of these relevant activities is the validity of the types of data discharge, confirms the results of the scalar (eg

Building in our e-commerce study, imagine that we want to quickly study the product update that is moving or packaging issues. We can use AI.GENERATE_BOOL For this binary separation:

SQL

SELECT *
FROM `my_dataset.reviews_table`
AI.GENERATE_BOOL(
   prompt => ("The review mentions a shipping or packaging problem", customer_reviews),
   connection_id => "us-central1.conn");

Question-sorting records and returns lines talking about issues by sending or packaging. Note that we do – You must specify keywords (eg.

To bring everything together: the joint question of multimodal

We examined how AI produced to raise data platform. Now, let's go back to the e-commerce challenge in the introduction: “Identify Electronic products with the highest return prices connected to customer pictures showing the damage symptoms when you arrive.” Historically, this requires different pipes and often see multiple persons (data scientist, data analyst, data engineer).

With the combined skills of AI, a good SQL question can now answer this question:

This combined question shows the significant appearance of how data platform works. Instead of maintaining and returning a variety of data, the platform becomes an active location where users can ask the business questions and return the answers directly of formal and random data, using the normal SQL unemployment. This combined provides a specific way to understand what is most needed technology and tool.

Semantic Reasoning and AI engine question (coming soon)

While jobs like AI.GENERATE_TABLE They have the ability to process the unkind AI (to advise each of the records or produce new data to them), Brieted and intended to combine complete, the thinking of ai (aqa).

The purpose of Aijate is to enable detailed commentators, even those who do not have ai-deep technology for AI, to make semantic complex thoughts all datasets. The AQA has accessed this by giving trouble such as speedy engineering and allows users to focus on Business Logic.

Sample sample activities can include:

• Ai.if: With semantic filtering. The llm checked if the line data matches the environmental condition during the fastest time (eg “Return a product review that suggests frustration by excessive heat”).
• Ai.join: Join the tables based on the same size in Semantic or in a relationship shown in natural language – not just the main equity (eg
• Ai.score: Estimates or orders lines how well they match the Semantic state, useful in the “Top-K” (eg Top Customer Support Calls.

Conclusion: Data Platform

Data platforms live in the ongoing environment of evolution. From the Source focuses on the management of formal, relationships, now accept the opportunities that are random, multimodal opportunities. A direct combination of SQL operators and index support to the controversial shopping files such as ways as these ObjectRef You represent a basic change in the way we work with the data.

As the lines between data management and AI continue to change, the repository represents the Central Hub of business data – now included in the ability to understand the rich, additional person. Multimodal questions that once required Disparate tools and a thick AI technology can be referred for the greatest. This appears on the skilled database platforms of demismics and allows the broad range of SQL users who receive deep understanding.

Testing these skills and start working with Multimodal data from Benery:

Author: Jeff Nelson, Engineering Engineering, Google Cloud