Identifying search results in a history repository

Patent metadata

Field	Value
Publication number	US-12579191-B2
Application number	US-202519041779-A
Country	US
Kind code	B2
Filing date	Jan 30, 2025
Priority date	Jan 31, 2024
Publication date	Mar 17, 2026
Grant date	Mar 17, 2026

Abstract

The present disclosure provides a system and methods for providing responses to user provided questions that are grounded in the user's browsing history and enables searching of the history for resources previously viewed based on the content of the resources. In one example embodiment, a portion of a query is received. A query suggestion relevant to the portion of the query is obtained. A resource from a history repository relevant to the portion of the query is identified by obtaining a semantic representation of the portion of the query, and identifying a semantic representation of content associated with the resource from the history repository. The semantic representation of the content includes a similarity score with the semantic representation of the query that satisfies a threshold. A resource suggestion is generated for the resource. The query suggestion and the resource suggestion is provided as selectable query completions.

First claim

What is claimed is:

1 . A method comprising:
receiving a portion of a query from a user interface; identifying a resource from a history repository relevant to the portion of the query by:
processing the portion of the query through a generative model to generate a semantic representation of the portion of the query, and
identifying a matching quantized embedding representing a corresponding passage from the resource by comparing the semantic representation of the portion of the query against a plurality of quantized embeddings stored in the history repository, wherein:
a measure of similarity between the matching quantized embedding representing the corresponding passage and the semantic representation of the portion of the query satisfies a threshold,
the plurality of quantized embeddings were previously generated by quantizing a semantic representation of content to reduce its memory footprint for storage in the history repository, and
a quantity of quantized embeddings generated for a resource is determined based on an amount of memory available for the history repository;

processing the portion of the query and the passage through the generative model to generate a response to the portion of the query based on the passage; generating a resource suggestion corresponding to the resource; and providing, to the user interface, the response and the resource suggestion as a citation for the response.

2 . The method of claim 1 , wherein the matching quantized embedding is a quantized numerical representation, generated by the generative model, that approximates the passage.

3 . A non-transitory computer-readable medium including instructions that, when executed by an electronic processor, cause a computing system to perform a method comprising:
receiving a portion of a query from a user interface; identifying a resource from a history repository relevant to the portion of the query by:
processing the portion of the query through a generative model to generate a semantic representation of the portion of the query, and
identifying a matching quantized embedding representing a corresponding passage from the resource by comparing the semantic representation of the portion of the query against a plurality of quantized embeddings stored in the history repository, wherein:
a measure of similarity between the matching quantized embedding representing the corresponding passage and the semantic representation of the portion of the query satisfies a threshold,
the plurality of quantized embeddings were previously generated by quantizing a semantic representation of content to reduce its memory footprint for storage in the history repository, and
a quantity of quantized embeddings generated for a resource is determined based on an amount of memory available for the history repository;

processing the portion of the query and the passage through the generative model to generate a response to the portion of the query based on the passage; generating a resource suggestion corresponding to the resource; and providing, to the user interface, the response and the resource suggestion as a citation for the response.

4 . The non-transitory computer-readable medium of claim 3 , wherein the passage from the resource is included in a webpage, and
the generative model is configured to generate embeddings for webpage content by training using training sets of webpage content.

5 . The non-transitory computer-readable medium of claim 3 , wherein processing the portion of the query and the passage through the generative model to generate a response to the portion of the query includes:
retrieving the passage from the resource, indexed content, or a cache; and providing the portion of the query and the retrieved passage to the generative model.

6 . A system comprising:
a display providing a user interface; a repository; and an electronic processor communicably coupled to the display and the repository, the electronic processor configured to:
receive a portion of a query from the user interface;
identify a resource from the repository relevant to the portion of the query by:
processing the portion of the query through a generative model to generate a semantic representation of the portion of the query, and

identifying a matching quantized embedding representing a corresponding passage from the resource by comparing the semantic representation of the portion of the query against a plurality of quantized embeddings stored in the repository, wherein:
a measure of similarity between the matching quantized embedding representing the corresponding passage and the semantic representation of the portion of the query satisfies a threshold,
the plurality of quantized embeddings were previously generated by quantizing a semantic representation of content to reduce its memory footprint for storage in the repository, and
a quantity of quantized embeddings generated for a resource is determined based on an amount of memory available for the repository;

process the portion of the query and the passage through the generative model to generate a response to the portion of the query based on the passage;
generate a resource suggestion corresponding to the resource; and
provide, to the user interface, the response and the resource suggestion as a citation for the response.

7 . The method of claim 1 , wherein the generative model is trained using a training set of webpage content to generate the semantic representation of the portion of the query and the plurality of quantized embeddings.

8 . The method of claim 7 , wherein the generative model is further trained based on a quantity of passages to be stored per resource.

9 . The method of claim 1 , wherein the resource suggestion is provided as a selectable query completion within a navigation text box of a browser.

10 . The method of claim 1 , wherein the plurality of quantized embeddings stored in the history repository includes multiple quantized embeddings that correspond to multiple respective passages from a single resource.

11 . The method of claim 10 , further comprising generating the multiple quantized embeddings by:
breaking content of the single resource into the multiple respective passages; and generating respective quantized embeddings for the multiple respective passages.

12 . The method of claim 1 , further comprising ranking a plurality of resources from the history repository based on a measure of similarity between their corresponding quantized embeddings and the semantic representation of the portion of the query, wherein the ranking limits a quantity of resources identified from a single domain.

13 . The method of claim 1 , wherein the history repository is within a memory-constrained computing environment, and wherein the identifying of the resource is performed without communication with a server.

14 . The non-transitory computer-readable medium of claim 3 , wherein the generative model is trained using a training set of webpage content to generate the semantic representation of the portion of the query and the plurality of quantized embeddings.

15 . The non-transitory computer-readable medium of claim 3 , wherein the plurality of quantized embeddings stored in the history repository includes multiple quantized embeddings that correspond to multiple respective passages from a single resource.

16 . The system of claim 6 , wherein the generative model is trained using a training set of webpage content to generate the semantic representation of the portion of the query and the plurality of q

Assignees

• Google Llc

Inventors

• Yushkina Yana
• Ahmadi Mohamad Hasan
• Chang Sophie
• Crouse Michael Blair
• Donnelly Justin Gabriel
• Por John Oliver
• Chang Mark
• Meng Shuo
• Li Tommy Chendong
• Pathak Manas
• Brinzarei Denis
• Jaworski Orin Gabriel
• Zou Jun

Classifications

• G06F16/90328 Primary

  using search space presentation or visualization, e.g. category or range presentation and selection · CPC title