Neural Network Embeddings for Multi-Tenant Database Search Relevance

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

In multi-tenant database systems, users face inefficiencies when searching for data due to high resource consumption and long processing times, often resulting in irrelevant search results, as the system struggles to accurately predict the most relevant entity type for a user's query, especially across different organizations with varying interests.

Innovation Solution

A neural network model is implemented to capture organization-specificities, allowing it to predict the most relevant entities for search queries in a multi-tenant database system, taking into account the specific search patterns and interests of users from different organizations, thereby enhancing user experience by providing more accurate and relevant search results.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a database system processes search queries by identifying relevant objects from many records, then the user can access desired information, but the process consumes large amounts of system resources and takes a long period of time

Engineering Contradiction:

Improvesearch result relevanceVSAvoidquery processing time

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

The system generates embedding representations for all database objects in advance and stores them in an index structure before queries are submitted. This preliminary action allows the search system to quickly compare query embeddings against pre-computed object embeddings without performing heavy computational analysis during query processing, thereby reducing query processing time while maintaining search result relevance.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent replaces traditional keyword-based mechanical search mechanisms with a neural network-based embedding system. The neural network automatically learns and extracts semantic features from object data, transforming the search process from manual keyword matching to intelligent semantic comparison. This substitution enables faster and more accurate identification of relevant objects without consuming excessive system resources during query processing.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

2Measurement precision

Engineering Contradiction:

Improvesearch result relevanceVSAvoidsystem resource consumption

Core Design Contradiction:

Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The system generates and stores embedding representations for all database objects in advance, during off-peak periods or as data is added to the database. This preliminary computation shifts the resource-intensive work from query-time processing to batch processing, dramatically reducing the system resources required during actual search operations while maintaining high search result relevance.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent creates compressed embedding representations (vectors) of database objects that capture essential semantic information in a compact form. These embedding copies allow the system to perform efficient similarity comparisons without accessing or processing the full original objects, reducing memory access patterns and computational resources required during query processing.

Inventive Principle:

Principle #26Copying

3Adaptability or versatility

If a database system returns search results containing many false hits, then the user can see multiple potential matches, but the most relevant information is buried or obscured in the returned results

Engineering Contradiction:

Improvesearch result coverageVSAvoidsearch result relevance

Core Design Contradiction:

Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The patent replaces traditional keyword-matching search mechanisms with a neural network-based embedding system that understands semantic meaning. The neural network learns to distinguish between truly relevant objects and false hits by analyzing contextual relationships in the data. This substitution enables the system to return search results that are both comprehensive and precisely ranked by relevance, eliminating the problem of buried relevant information among false hits.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The system uses feedback from user interactions with search results to continuously improve the quality of search rankings. By monitoring which results users click on and which they ignore, the system refines its embedding generation and comparison processes to better prioritize truly relevant information. This feedback mechanism ensures that the most relevant information remains prominent in search results while filtering out false hits.

Inventive Principle:

Principle #23Feedback

Data Source

PatentUS11328203B2Capturing organization specificities with embeddings in a model for a multi-tenant database system

Publication Date: 2022.05.10 SALESFORCE INC

AI summary

For a multi-tenant database accessible by a plurality of separate organizations, a system is provided for capturing organization specificities in a model for the multi-tenant database. The system includes a neural network. The system is configured to: receive an organization encoding for one or more separate organizations making previous search queries into the multi-tenant database; generate a vector matrix from the organization encoding to embed organization specificities for training a model of the neural network; and using the vector matrix, train the model of the neural network for processing a present search query into the multi-tenant database. In some embodiments, the model of the neural network is global across the separate organizations accessing the database.