Introducing AI Features: Integrating Giant Language Fashions with Databricks SQL

With all of the unbelievable progress being made within the area of Giant Language Fashions, prospects have requested us how they will allow their SQL analysts to leverage this highly effective know-how of their day-to-day workflows.

Immediately, we’re excited to announce the general public preview of AI Features. AI Features is a built-in DB SQL perform, permitting you to entry Giant Language Fashions (LLMs) immediately from SQL.

With this launch, now you can rapidly experiment with LLMs in your firm’s knowledge from inside a well-known SQL interface. After getting developed the right LLM immediate, you possibly can rapidly flip that right into a manufacturing pipeline utilizing present Databricks instruments resembling Delta Reside Tables or scheduled Jobs. This enormously simplifies each the event and productionization workflow for LLMs.

AI Features abstracts away the technical complexities of calling LLMs, enabling analysts and knowledge scientists to start out utilizing these fashions with out worrying in regards to the underlying infrastructure.

Utilizing AI Features

To indicate how AI Features works, think about that you just’re an analyst and have been given a historic record of hundreds of name transcripts with the duty of offering a report that breaks down all the calls into considered one of 4 classes [Frustrated, Happy, Neutral, Satisfied]. Usually, this might require you to request the information science workforce create a classification mannequin. As an alternative, with AI Features, you possibly can immediate a Giant Language Mannequin, resembling OpenAI’s ChatGPT mannequin, immediately from SQL. An instance LLM immediate may seem like the next:

Immediate: Classify the next textual content into one of 4 classes 
[Frustrated, Happy, Neutral, Satisfied]: 

Thanks a lot for serving to me as we speak, you have got resolved my concern with the additional invoice.

Response: Happy

With AI Features, you possibly can flip this immediate in a customized SQL perform. This allows you to do the next, fairly than a sophisticated multi-step pipeline:

SELECT 
transcript_line, 
CLASSIFY_TRANSCRIPT(transcript_line) as classification
FROM
your_dataset

Let’s use this instance to stroll by the steps wanted to take action. We are going to use the Azure OpenAI service as our giant language mannequin, though you can additionally use OpenAI. In future releases, we are going to allow different Giant Language Fashions, together with open supply LLMs resembling Dolly.

We’ve beforehand saved an Azure OpenAI API key as a Databricks Secret so we will reference it with the SECRET perform. With that saved, let’s check out the brand new AI_GENERATE_TEXT perform to see how we make it work for our particular objective:

AI_GENERATE_TEXT(
  immediate,
 'azure_openai/gpt-35-turbo',
 'apiKey', SECRET('tokens', 'azure-openai'),
 "deploymentName", "llmbricks",
 "apiVersion", "2023-03-15-preview", 
 "resourceName", "lakehouserules",
 "temperature", CAST(0.0 as DOUBLE)
);

We recommend that you just wrap the AI_GENERATE_TEXT perform with one other perform. This makes it simpler to cross alongside enter knowledge resembling transcripts, and to call the perform to make it extra descriptive of the meant objective:

CREATE OR REPLACE FUNCTION CLASSIFY_TRANSCRIPT(transcript STRING)
RETURNS STRING
RETURN AI_GENERATE_TEXT(
 CONCAT(‘Classify the next textual content into considered one of 4 classes [Frustrated, Happy, Neutral,  
  Satisfied]’, 
  transcript),
 'azure_openai/gpt-35-turbo',
 'apiKey', SECRET('tokens', 'azure-openai'),
 "deploymentName", "llmbricks",
 "apiVersion", "2023-03-15-preview", 
 "resourceName", "llmbricks",
 "temperature", CAST(0.0 as DOUBLE)
);

That is all it takes to attain the performance that we set out at the beginning, which results in our question being so simple as:

SELECT 
transcript_line, 
CLASSIFY_TRANSCRIPT(transcript_line) as classification
FROM
your_dataset

No difficult pipelines, no submitting tickets for knowledge engineers to create new processes, or tickets for knowledge scientists to create new fashions – all it takes is a little bit of your creativity to develop the immediate and easy SQL to carry the unbelievable energy of LLMs proper to your knowledge.

AI Features allows you to harness the facility of Giant Language Fashions – from translating from one language to a different, summarizing textual content, suggesting subsequent steps for help groups, and even utilizing a number of perform requires multi-shot prompts.

AI Features are simply the beginning on our journey to empower customers to simply customise LLMs to your online business and use LLMs along with your knowledge. We will’t wait to see what you construct!

Join the Public Preview of AI Features right here and ensure to take a look at our Webinar masking construct your individual LLM like Dolly right here! For extra particulars you can even learn the docs right here.