Email Attachment Search Prioritization for Faster Retrieval
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing email search systems face challenges in efficiently searching through large corpuses of emails, particularly when content is in attachments, due to prohibitive processing times and increased power consumption, and existing AI methods struggle with large volumes of text conversion from attachments.
Innovation Solution
A method utilizing multiple large language models (LLMs) to process and score emails and attachments based on user likelihood of searching, with stages to filter and vectorize relevant data, including stop-word removal and OCR for attachments, to enhance retrieval efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If all attachments are converted to text and processed through LLM for search, then search completeness is improved, but processing time and power consumption increase significantly
Solution Approach 1:
The patent applies preliminary action by pre-processing emails during ingestion, extracting metadata, keywords, and summaries before search operations. This allows the system to prepare search-ready data in advance, so when a search occurs, the LLM only needs to process pre-extracted features rather than converting all attachments from scratch, significantly reducing search processing time while maintaining completeness
Solution Approach 2:
The patent segments the email corpus into processed metadata/keywords and attachment content. The LLM processes only the segmented metadata and extracted keywords rather than all raw attachment text, reducing the token volume while preserving search capability for both email content and attachment content
2Measurement precision
If all attachments are converted to text for LLM processing, then search accuracy is improved, but the volume of text tokens increases significantly
Solution Approach 1:
The patent extracts only the most search-relevant information from attachments using OCR and text extraction, then feeds only this extracted content to the LLM rather than all attachment text. This extraction process isolates key searchable elements while discarding redundant information, reducing token volume while preserving search accuracy
Solution Approach 2:
The patent applies partial action by processing only a subset of attachment content (extracted text and metadata) through the LLM rather than all attachment text. This selective processing reduces token volume while maintaining sufficient search accuracy for most use cases
3Reliability
If keyword search is performed across large email volumes, then search coverage is improved, but false positive results increase
Solution Approach 1:
The patent implements feedback mechanisms where the LLM evaluates search results and ranks them by relevance, using the email metadata, extracted keywords, and attachment content to score results. This feedback loop ensures comprehensive coverage while filtering out irrelevant results, improving result relevance through intelligent ranking rather than simple keyword matching
Data Source
Figure 1
Figure 1
AI summary
The present disclosure is directed towards improving the speed and reducing the power consumption of a search for an attachment in a corpus of emails. The system provides a computer-implemented method of storing an email to facilitate its retrieval. The method comprises: processing at least one of the subject line of an email, the body of the email, or a file name of one or more attachments to the email with a first LLM to determine a first score for the email indicative of how likely a user is to search for the email; and storing the email based on the first score, whereby: an email which is more likely to be searched for is stored such that it more easily retrieved than an email that is not likely to be searched for.