Categories:: Aggregate functions (General) String & binary functions (Large Language Model)

AI_AGG¶

Reduces a column of text data using a natural language instruction.

For example, AI_AGG(reviews, 'Describe the most common complaints mentioned in the book reviews') will return a summary of user feedback.

Unlike COMPLETE (SNOWFLAKE.CORTEX) and SUMMARIZE (SNOWFLAKE.CORTEX), this function supports datasets larger than the maximum language model context window.

See also:: AI_SUMMARIZE_AGG

Syntax¶

AI_AGG( <expr>, <instruction> )

Copy

Arguments¶

Required:

expr: This is an expression that contains text on which an aggregation operation is to be performed, such as restaurant reviews or phone transcripts.
instruction: A string containing a natural language specification of the aggregation to perform, for example “Summarize the reviews” or “Identify all people mentioned and write a short biography for each of them”.

Returns¶

Returns a string containing the result of the aggregation.

The function may indicate that the data you’ve provided doesn’t contain the answer if:

You don’t provide a clear instruction specifying how to aggregate the data
The data doesn’t have the information necessary to complete your instruction

Usage notes¶

For optimal performance, follow these guidelines:

Use plain English text for the instruction.
Provide a declarative instruction instead of asking a question. For example, instead of a question like “Can you summarize this?”, use “Summarize the phone call transcripts”.
Describe the text provided in the instruction. For example, instead of an instruction like “summarize”, use “Summarize the phone call transcripts”.
Describe the intended use case. For example, instead of “find the best review”, use “Find the most positive and well-written restaurant review to highlight on the restaurant website”.
Consider breaking the instruction into multiple steps. For example, instead of “Summarize the new articles”, use “You will be provided with news articles from various publishers presenting events from different points of view. Please create a concise and elaborative summary of source texts without missing any crucial information.”.

Examples¶

AI_AGG can be used as a simple scalar function on string constants. In the following example, AI_AGG is used to summarize product ratings, which are provided as a single string.

SELECT AI_AGG('[Excellent, Excellent, Great, Mediocre]',
              'Summarize the product ratings for a blog post targeting consumers');

Copy

Overall, the product has received overwhelmingly positive reviews, with the majority of users rating it as 'Excellent' or 'Great'. Only a small percentage of users had a mediocre experience with the product. This suggests that the product is well-liked by most consumers and is a great option for those looking for a reliable choice.

AI_AGG can also be used on a column of data. In the following example, the product ratings from the above example are provided as a column in a table using a Common Table Expression.

WITH reviews AS (
            SELECT 'The restaurant was excellent.' AS review
  UNION ALL SELECT 'Excellent! I loved the pizza!'
  UNION ALL SELECT 'It was great, but the service was meh.'
  UNION ALL SELECT 'Mediocre food and mediocre service'
)
SELECT AI_AGG(review, 'Summarize the restaurant reviews for potential consumers')
  FROM reviews;

Copy

Reviews for this restaurant are mixed. Some customers had a very positive experience, describing the restaurant as "excellent" and loving the pizza. However, others had a more neutral or negative experience, citing mediocre food and service.

AI_AGG can also be used in combination with GROUP BY. The following example summarizes product ratings for two products (identified by the column product_id) in a table of reviews.

WITH reviews AS (
            SELECT 1 AS restaurant_id, 'The restaurant was excellent.' AS review
  UNION ALL SELECT 1, 'Excellent! I loved the pizza!'
  UNION ALL SELECT 1, 'It was great, but the service was meh.'
  UNION ALL SELECT 1, 'Mediocre food and mediocre service'
  UNION ALL SELECT 2, 'Terrible quality ingredients, I should have eaten at home.'
  UNION ALL SELECT 2, 'Bad restaurant, I would avoid this place.'
)
SELECT restaurant_id,
       AI_AGG(review, 'Summarize the restaurant reviews for potential consumers')
  FROM reviews
 GROUP BY 1;

Copy

+---------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| RESTAURANT_ID | SUMMARIZED_REVIEW                                                                                                                                                                                                                                 |
|---------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| 1             | Reviews for this restaurant are mixed. Some customers had a very positive experience, describing the restaurant as "excellent" and loving the pizza. However, others had a more neutral or negative experience, citing mediocre food and service. |
+---------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| 2             | Two reviewers had extremely negative experiences at this restaurant, citing poor quality ingredients and advising others to avoid it.                                                                                                             |
+---------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+

The instruction can be used for various aggregation tasks and to configure the style and tone of the response. The following example uses an instruction to find the most positive rating for each product and provide French and Polish translations of the rating.

WITH reviews AS (
            SELECT 1 AS product_id, 'Excellent' AS review
  UNION ALL SELECT 1, 'Excellent'
  UNION ALL SELECT 1, 'Great'
  UNION ALL SELECT 1, 'Mediocre'
  UNION ALL SELECT 2, 'Terrible'
  UNION ALL SELECT 2, 'Bad'
  UNION ALL SELECT 2, 'Average'
)
SELECT product_id,
       AI_AGG(review, 'Identify the most positive rating and translate it into French and Polish, one word only') AS summarized_review
  FROM reviews
 GROUP BY 1;

Copy

+------------+--------------------+
| PRODUCT_ID | SUMMARIZED_REVIEW  |
|------------+--------------------+
| 1          | French: Excellent  |
|            | Polish: Doskonały  |
+------------+--------------------+
| 2          | French: Moyen      |
|            | Polish: Przeciętny |
+------------+--------------------+