Rank search results

This page describes how to rank your search results for applications using the Vertex AI ranking model endpoint.

The Vertex AI ranking API takes a list of documents and ranks those documents based on how relevant the documents are to a given query (a search string). When you use the ai.rank() function, it returns scores for how well a document answers a given query.

To use instructions on this page, you must have an understanding of AlloyDB for PostgreSQL and be familiar with generative AI concepts.

AlloyDB reserves, and creates, the ai schema.

Before you begin

Before you rank search results, do the following:

Enable the Discovery Engine API

Console

Enable the API
In the Confirm project step, click Next to confirm the name of the project you are going to make changes to.
In the Enable APIs step, click Enable to enable the Discovery Engine API. If you already enabled this API, you won't see it listed here.

gcloud

To use ranking models, you must enable the Discovery Engine API.
Replace PROJECT_ID with your Google Cloud project ID and PROJECT_NUMBER with your corresponding project number.

 # 
Enable  
Discovery  
Engine  
API gcloud services enable discoveryengine.googleapis.com --project= PROJECT_ID 
 
 gcloud projects add-iam-policy-binding PROJECT_ID 
\ 
 --member="serviceAccount:service- PROJECT_NUMBER 
@gcp-sa-alloydb.iam.gserviceaccount.com" \ 
 --role="roles/discoveryengine.viewer"

Model registration for ranking isn't required for Vertex AI models. You can use the Vertex AI model name as the model_id , which is shown in the following example.

  
SELECT  
index,  
score  
FROM  
ai.rank ( 
  
 model_id 
  
 = 
>  
 'semantic-ranker-default-003' 
,  
 search_string 
  
 = 
>  
 'Affordable family-friendly vacation spots in Southeast Asia?' 
,  
 documents 
  
 = 
>  
ARRAY [ 
  
 'Luxury resorts in South Korea' 
,  
 'Family vacation packages for Vietnam: Ha Long Bay and Hoi An' 
,  
 'Budget-friendly beaches in Thailand perfect for families' 
,  
 'A backpacker guide to solo travel in India' 
 ])

A common use case for the semantic ranker is to rerank the results returned by vector search for better query ordering. The following example shows how to use the semantic ranking model for this use case. The example retrieves an initial result set for the query personal fitness equipment using vector search. These results are then re-ranked to return the top five results.

  
WITH  
initial_ranking  
AS  
 ( 
  
SELECT  
id,  
description,  
ROW_NUMBER () 
  
OVER  
 () 
  
AS  
ref_number  
FROM  
product  
ORDER  
BY  
embedding  
< = 
>  
google_ml.embedding ( 
  
 'gemini-embedding-001' 
,  
 'personal fitness equipment' 
 ) 
::vector  
LIMIT  
 10 
  
 ) 
,  
reranked_results  
AS  
 ( 
  
SELECT  
index,  
score  
FROM  
ai.rank ( 
  
 model_id 
  
 = 
>  
 'semantic-ranker-default-003' 
,  
 search_string 
  
 = 
>  
 'personal fitness equipment' 
,  
 documents 
  
 = 
>  
 ( 
SELECT  
ARRAY_AGG ( 
description  
ORDER  
BY  
ref_number ) 
  
FROM  
initial_ranking ) 
,  
 top_n 
  
 = 
>  
 5 
 ) 
  
 ) 
  
SELECT  
id,  
description  
FROM  
initial_ranking,  
reranked_results  
WHERE  
initial_ranking.ref_number  
 = 
  
reranked_results.index  
ORDER  
BY  
reranked_results.score  
DESC ;

For a list of available models and use cases, see Supported models .

Integrate with Vertex AI and install the extension

Integrate with Vertex AI .

Ensure that the latest version of google_ml_integration is installed.

To check the installed version, run the following command:

  
SELECT  
extversion  
FROM  
pg_extension  
WHERE  
 extname 
  
 = 
  
 'google_ml_integration' 
 ; 
  
extversion  
  
 1 
.4.3  
 ( 
 1 
  
row )

If the extension isn't installed or if the installed version is earlier than 1.4.3, update the extension by running the following commands:
```
  
CREATE  
EXTENSION  
IF  
NOT  
EXISTS  
google_ml_integration ; 
  
ALTER  
EXTENSION  
google_ml_integration  
UPDATE ; 
  
```
If you experience issues when you run the preceding commands, or if the extension isn't updated to version 1.4.3 after you run the preceding commands, contact AlloyDB support.

After you ensure that the version is current, install the preview functionality by running the upgrade_to_preview_version procedure:

  
CALL  
google_ml.upgrade_to_preview_version () 
 ; 
  
SELECT  
extversion  
FROM  
pg_extension  
WHERE  
 extname 
  
 = 
  
 'google_ml_integration' 
 ; 
  
extversion  
  
 1 
.4.4  
 ( 
 1 
  
row )

Required roles

To get the permissions that you need to use ranking models from Discovery Engine, ask your administrator to grant you the Discovery Engine Viewer ( roles/discoveryengine.viewer ) Identity and Access Management (IAM) role on your project . For more information about granting roles, see Manage access to projects, folders, and organizations .

You might also be able to get the required permissions through custom roles or other predefined roles .

Rank your search results

The following SQL query shows how to rank your search results :

  SELECT 
  
 ai 
 . 
 rank 
 ( 
  
 model_id 
  
 = 
>  
 ' MODEL_ID 
' 
 , 
  
 search_string 
  
 = 
>  
 ' SEARCH_STRING 
' 
 , 
  
 documents 
  
 = 
>  
 ARRAY 
 [ 
 ' DOCUMENT_1 
' 
 , 
  
 ' DOCUMENT_2 
' 
 , 
  
 ' DOCUMENT_3 
' 
 ]);

Replace the following:

Parameter	Description
`MODEL_ID`	A unique ID for the model endpoint that you define.
`SEARCH_STRING`	A search string against which the records are ranked.
`DOCUMENTS`	A unique string that identifies the record.

For a list of the supported Vertex AI ranking models, see Supported models .

Examples

To rank search results using a Vertex AI ranking model, run the following query:

  SELECT 
  
 index 
 , 
  
 score 
 FROM 
  
 ai 
 . 
 rank 
 ( 
  
 model_id 
  
 = 
>  
 'semantic-ranker-default-003' 
 , 
  
 search_string 
  
 = 
>  
 'AlloyDB is a PostgreSQL compatible AI database that is ready for production.' 
 , 
  
 documents 
  
 = 
>  
 ARRAY 
 [ 
  
 'Alloys are made from combination of metals' 
 , 
  
 'The best enterprise-ready PostgreSQL database.' 
 , 
  
 'You can feel the heat in Alloy apartments.' 
 ]);

The response is a table that shows each document and the score based on relevance to the search query. The following is the sample response:

   
 index 
  
 | 
  
 score 
 -------+------------ 
  
 2 
  
 | 
  
 0 
 . 
 33 
  
 1 
  
 | 
  
 0 
 . 
 28 
  
 3 
  
 | 
  
 0 
 . 
 16 
 ( 
 3 
  
 rows 
 )

Consider an example AlloyDB database with a list of review descriptions that are converted to embeddings. The following sample code snippet shows how to use the ranking model to retrieve the name of the top-ranked products based on their review descriptions' semantic similarity to a query.

  WITH 
  
 initial_ranking 
  
 AS 
  
 ( 
  
 SELECT 
  
 product_id 
 , 
  
 name 
 , 
  
 review 
 , 
  
 review_id 
 , 
  
 ROW_NUMBER 
 () 
  
 OVER 
  
 () 
  
 AS 
  
 ref_number 
  
 FROM 
  
 user_reviews 
  
 ORDER 
  
 BY 
  
 review_desc_embedding 
  
< = 
>  
 google_ml 
 . 
 embedding 
 ( 
  
 'gemini-embedding-001' 
 , 
  
 'good desserts' 
 ):: 
 vector 
  
 LIMIT 
  
 10 
  
 ), 
  
 reranked_results 
  
 AS 
  
 ( 
  
 SELECT 
  
 index 
 , 
  
 score 
  
 FROM 
  
 ai 
 . 
 rank 
 ( 
  
 model_id 
  
 = 
>  
 'semantic-ranker-512' 
 , 
  
 search_string 
  
 = 
>  
 'good desserts' 
 , 
  
 documents 
  
 = 
>  
 ( 
 SELECT 
  
 ARRAY_AGG 
 ( 
 review 
  
 ORDER 
  
 BY 
  
 ref_number 
 ) 
  
 FROM 
  
 initial_ranking 
 ), 
  
 top_n 
  
 = 
>  
 5 
 ) 
  
 ) 
 SELECT 
  
 product_id 
 , 
  
 name 
 FROM 
  
 initial_ranking 
 , 
  
 reranked_results 
 WHERE 
  
 initial_ranking 
 . 
 ref_number 
  
 = 
  
 reranked_results 
 . 
 index 
 ORDER 
  
 BY 
  
 reranked_results 
 . 
 score 
  
 DESC 
 ;