Chaining Model Tasks with AI Functions: A Customer Review Pipeline

1.Task-specific function or ai_query?

After you have named the model task, there is one more choice: a task-specific function or the general-purpose one. Databricks offers both kinds. Task-specific AI Functions are "Purpose-built functions optimized for a specific task, such as document parsing, entity extraction, classification, and sentiment analysis." The other kind is "ai_query — The general-purpose function for task and model flexibility. Provide a prompt and choose any supported Foundation Model API."

The rule follows from what each one gives you. If the task is in the catalog (sentiment, classification, extraction, summarization, translation, masking), start with the task-specific function. You don't write a prompt, you don't parse output, and Databricks maintains the technique behind it. Use ai_query when the requirement needs a custom prompt or a particular model. The catalog says so itself next to its generation function: "For custom prompts or a specific model, see Use ai_query."

All of these run where your data lives. They "can be run from Databricks SQL, notebooks, Lakeflow pipelines, and Workflows", with one exception: they are not available on Classic SQL warehouses.

Sources1

2.One requirement, four chained tasks

The Databricks customer-review example turns one vague requirement into a chain of tasks: deal with unhappy customers and improve the product descriptions. The pipeline determines the sentiment of each review. For negative reviews, it extracts information to classify the cause. It identifies whether the customer needs a response, and it generates a reply that mentions alternative products.

Each step narrows the input for the next one. Sentiment is a cheap filter. Only negative reviews go on to cause classification, and the WHERE clause reuses the sentiment function to do the filtering:

SELECT
  review,
  ai_classify(
    review,
    '["Arrives too late", "Wrong size", "Wrong color", "Dislike the style"]'
  ) AS reason
FROM
  product_reviews
WHERE
  ai_analyze_sentiment(review) = "negative"

The labels come from the business. Logistics and merchandising care about different failure reasons, so they write the label list. That is what makes this classification rather than open-ended generation: the documentation notes that ai_classify() "is able to correctly categorize the negative reviews based on custom labels to allow for further analysis."

To improve product descriptions, the team needs a specific fact from each review, such as the customer's usual size. That fact is a field, so the task is extraction. The same query also classifies fit, which shows that you can mix tasks in one SELECT.

SELECT
  review,
   ? (review, '["usual_size"]') AS usual_size,
  ai_classify(review, '["Size is wrong", "Size is right"]') AS fit
FROM
  product_reviews

Sources2

3.Closing the loop: the generation task

The last step is the only one that creates text the customer will read. That makes it generation, and the catalog's function for it is ai_gen, which will "Answer a user-provided prompt using a state-of-the-art AI model." Here, unlike the earlier steps, you write the prompt yourself. In the example, the prompt sets the length (60 words), the content (the customer's opinions are valued and a 30% discount coupon has been sent to their email), and the input (the review text joined onto the instruction). It runs only on rows where ai_analyze_sentiment(review) = "negative", the same filter the classification step used.

The progression is worth remembering. Sentiment, classification and extraction each produce a narrow output. A label or a field can be checked, so those steps need no prompt. Generation produces open-ended text, so the business rules have to go into the prompt. If those rules ever require a model other than the one Databricks manages, the requirement has outgrown ai_gen. The catalog points you to ai_query for custom prompts or a specific model.

Sources12

Exam traps

Each one states something that sounds right. Open it to see what is actually true.

1. 1.Sentiment or label classification over text always needs ai_query with a carefully engineered prompt and output parsing.Why is that wrong?

   Task-specific functions such as ai_analyze_sentiment return the result directly. Databricks recommends starting with them because they need no customization.

   Covered in Task-specific function or ai_query?

Practise it for real

Turn the requirement "respond to unhappy customers and learn why they are unhappy" into chained AI Function calls over a product_reviews table

1. 1.Run SELECT review, ai_analyze_sentiment(review) AS sentiment FROM product_reviews; from a notebook or a non-Classic SQL warehouse

   Why: Sentiment is the cheap first filter, and it decides which rows the later tasks process

   You should see: Each review is labeled positive, negative, neutral or mixed, with no prompt written

2. 2.Run the ai_classify query with the labels Arrives too late, Wrong size, Wrong color and Dislike the style, filtered on negative sentiment

   Why: Cause classification uses labels the business defines, so the results can be grouped and counted

   You should see: A reason column holding one of your labels for each negative review

3. 3.Run the ai_extract query for usual_size alongside ai_classify for fit

   Why: Extraction pulls out a named field that can drive changes to product descriptions

   You should see: usual_size and fit columns next to each review

4. 4.Run ai_gen on negative reviews with a prompt for a 60-word reply that mentions the discount coupon

   Why: Generation is the only step whose output customers read, so its rules go into the prompt

   You should see: A reply column with a short, tailored response for each negative review