Enhanced searching using fine-tuned machine learning models

Patent metadata

Field	Value
Publication number	US-12314318-B2
Application number	US-202418444078-A
Country	US
Kind code	B2
Filing date	Feb 16, 2024
Priority date	Feb 17, 2023
Publication date	May 27, 2025
Grant date	May 27, 2025

Abstract

An advanced search system leverages a pre-trained large language model to enhance user query responses. The system, equipped with hardware processors, a search query via an interface and accesses a pre-trained large language model designed to respond to the search query. The system fine-tunes the model to generate a task-specific generative model. The system employs the task-specific generative model to generate a search result to the search query and analyzes the search result based on a performance metric associated with the task-specific generative model. The system refines the task-specific generative model based on the analyzing of the search result.

First claim

What is claimed is:

1. A system comprising:
one or more hardware processors of a machine; and
at least one memory storing instructions that, when executed by the one or more hardware processors, cause the system to perform operations comprising:
receiving, by the one or more hardware processors, a search query via an interface;
accessing a pre-trained large language model designed to respond to the search query; and
performing a plurality of iterations, using the pre-trained large language model, to generate a task-specific generative model, each iteration of the plurality of iterations comprising:
performing domain-specific pre-training on an index to fine tune the pre-trained large language model;
employing the task-specific generative model to generate a search result to the search query;
analyzing the search result based on a performance metric associated with the task-specific generative model; and
refining the task-specific generative model based on the analyzing of the search result.

2. The system of claim 1 , wherein performing the plurality of iterations further comprises:
generating one or more outputs from the task-specific generative model by applying the plurality of iterations on a new search query, the one or more outputs including the search result; and
providing the one or more outputs to a user via the interface, wherein the interface is a browser-based interface.

3. The system of claim 2 , wherein performing the plurality of iterations further comprises:
receiving user feedback based on the one or more outputs; and
utilizing the user feedback to improve accuracy and fluency of the search result generated by the task-specific generative model.

4. The system of claim 1 , wherein performing the plurality of iterations further comprises:
evaluating a quality of the task-specific generative model upon conclusion of each iteration, wherein the quality includes a percentage of correct search results; and
stopping the plurality of iterations after the quality of the task-specific generative model is satisfactory to a user.

5. The system of claim 1 , the operations further comprising:
reducing a size of the task-specific generative model using asymmetric compression techniques including selective pruning of task-specific generative model parameters without identified loss of model performance.

6. The system of claim 1 , wherein performing the domain-specific pre-training on the index to fine tune the pre-trained large language model further comprises:
tailoring the pre-trained large language model using proprietary data, the proprietary data including a curated dataset representative of a plurality of types of queries and content associated with a search system.

7. The system of claim 1 , the operations further comprising:
applying a reward modeling process to the task-specific generative model to align the search result with human preferences; and
improving a quality of the search result based on the reward modeling process, wherein the reward modeling process includes collecting human annotations to define a reward function that approximates human judgments of fluency and relevance associated with the search result.

8. A method comprising:
receiving, by one or more hardware processors, a search query via an interface;
accessing a pre-trained large language model designed to respond to the search query; and
performing a plurality of iterations, using the pre-trained large language model, to generate a task-specific generative model, each iteration comprising:
performing domain-specific pre-training on an index to fine tune the pre-trained large language model;
employing the task-specific generative model to generate a search result to the search query;
analyzing the search result based on a performance metric associated with the task-specific generative model; and
refining the task-specific generative model based on the analyzing of the search result.

9. The method of claim 8 , wherein performing the plurality of iterations further comprises:
generating one or more outputs from the task-specific generative model by applying the plurality of iterations on a new search query, the one or more outputs including the search result; and
providing the one or more outputs to the user.

10. The method of claim 9 , wherein performing the plurality of iterations further comprises:
receiving, from the user, user feedback based on the one or more outputs; and
utilizing the user feedback to literately improve accuracy and fluency of the search result generated by the task-specific generative model.

11. The method of claim 8 , wherein performing the plurality of iterations further comprises:
evaluating a quality of the task-specific generative model at an end of each iteration, wherein the quality includes a percentage of correct search results; and
stopping the plurality of iterations after the quality of the task-specific generative model is satisfactory to the user.

12. The method of claim 8 , further comprising:
reducing a size of the task-specific generative model using asymmetric compression techniques including selective pruning of task-specific generative model parameters without identified loss of model performance.

13. The method of claim 8 , wherein performing the domain-specific pre-training on the index to fine tune the pre-trained large language model further comprises:
tailoring the pre-trained large language model using proprietary data, the proprietary data including a curated dataset representative of a plurality of types of queries and content associated with a search system.

14. The method of claim 8 , further comprising:
applying a reward modeling process to the task-specific generative model to align the search result with human preferences; and
improving a quality of the search result based on the reward modeling process, wherein the reward modeling process includes collecting human annotations to define a reward function that approximates human judgments of fluency and relevance associated with the search result.

15. A machine-storage medium embodying instructions that, when executed by a machine, cause the machine to perform operations comprising:
receiving, by one or more hardware processors, a search query via an interface;
accessing a pre-trained large language model designed to respond to the search query; and
performing a plurality of iterations, using the pre-trained large language model, to generate a task-specific generative model, each iteration comprising:
performing domain-specific pre-training on an index to fine tune the pre-trained large language model;
employing the task-specific generative model to generate a search result to the search query;
analyzing the search result based on a performance metric associated with the task-specific generative model; and
refining the task-specific generative model based on the analyzing of the search result.

16. The machine-storage medium of claim 15 , wherein performing the plurality of iterations further comprises:
generating one or more outputs from the task-specific generative model by applying the plurality of iterations on a new search query, the one or more outputs including the search result; and
providing the one or more outputs to a user device.

17. The machine-storage medium of claim 16 , wherein performing the plurality of iterations further comprises:
receiving user feedback based on the one or more outputs; and
utilizing the user feedback to literately improve accuracy and fluency of the search result generated by the

Assignees

• Snowflake Inc

Inventors

• Bathwal Rahil
• Campos Daniel Fernando
• Devaraj Ashwin
• Li Seth Michael
• Pande Yash
• Raghunathan Vivek
• Samdani Rajhans
• Xu Danmei

Classifications

• G06F16/248

  Presentation of query results · CPC title

• G06F16/24575 Primary

  using context · CPC title

• G06F16/9538

  Presentation of query results · CPC title

• G06F16/9558

  Details of hyperlinks; Management of linked annotations · CPC title

• G06F16/93

  Document management systems · CPC title