Email Ingestion Through Pattern Registries and Selective AI
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current email systems struggle to efficiently extract useful information from large volumes of emails, leading to the underutilization and loss of important data due to the inefficiency of keyword-based searches and the high computational cost of using generative AI for individual email interpretation.
Innovation Solution
A system utilizing a pattern registry and generative AI to interpret emails, where pattern matching with vector embeddings is used to identify similar emails, reducing the need for full generative AI processing on common messages, and leveraging machine learning to update the pattern registry for new patterns.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If keyword-based search is used to locate useful information in emails, then the search process is simple and fast, but the accuracy and completeness of information retrieval deteriorates when users are unaware of or have forgotten about the presence of useful information
Solution Approach 1:
The patent replaces manual keyword-based search with an automated machine learning system that generates vector embeddings for emails and performs pattern matching. This substitution transforms the mechanical keyword search process into an intelligent system that can automatically identify and retrieve relevant information based on semantic understanding rather than exact keyword matches, thereby improving both speed and completeness of information retrieval
Solution Approach 2:
The patent introduces vector embeddings as an intermediary representation between raw email content and search queries. By converting emails into vector embeddings and performing pattern matching in this intermediate space, the system enables more accurate and comprehensive information retrieval compared to direct keyword matching, while maintaining computational efficiency
2Measurement precision
If generative AI is used for individual email interpretation, then information extraction accuracy is improved, but computational cost and processing time increase significantly
Solution Approach 1:
The patent segments the email processing task into two distinct stages: (1) pattern matching stage using vector embeddings for quick comparison against the pattern registry, and (2) generative AI interpretation stage applied only when pattern matching fails to find sufficient matches. This segmentation allows the system to achieve high information extraction accuracy while significantly reducing computational cost by avoiding full generative AI processing for all emails
Solution Approach 2:
The patent applies partial action by using generative AI only when necessary (i.e., when no sufficient pattern match is found) rather than applying it to every email. This selective application of the computationally expensive generative AI model maintains high information extraction accuracy for complex cases while reducing overall computational cost by using the lighter pattern matching approach for the majority of emails
3Productivity
If pattern matching with vector embeddings is used to identify similar emails, then processing efficiency is improved for common messages, but the system complexity increases due to the need for pattern registry and embedding generation
Solution Approach 1:
The patent performs preliminary action by pre-processing emails into vector embeddings and storing them in a pattern registry before the actual search and matching operations. This preliminary transformation of email content into a standardized vector representation enables efficient pattern matching and similarity comparison, improving processing efficiency while the complexity is amortized over time as the pattern registry is built
Solution Approach 2:
The patent changes the parameter representation of emails from raw text to vector embeddings, transforming the data into a format that enables efficient mathematical operations for similarity comparison. This parameter transformation allows the system to achieve high processing efficiency through vector space operations while managing complexity through the use of established machine learning techniques
Data Source
AI summary
Systems and methods are described for ingesting and interpretation email messages. A pattern registry is generated for emails associated with a plurality of users based on a machine learning model interpretation of the emails. Email messages are received by a server computer, which performs pattern matching for each email message with the pattern registry. For example, embeddings for the emails may be extracted and compared to embeddings stored in a pattern database. Information for the emails is extracted based on the results of the pattern matching. For example, if a matching pattern is present in the pattern registry, the information for the email may be extracted based on the matching pattern. On the other hand, if no matching pattern is present in the pattern registry, the information for the email may be extracted using a machine learning model, such as a large language model, and the pattern registry is updated.