Embedding Cache for Search Feed Services
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge lies in efficiently determining similarities between a vast number of web documents and entities within a search and feed service, which is computationally intensive and memory-intensive due to the large corpus of data, especially when dealing with long-tail entities that are less frequently requested by users.
Innovation Solution
The implementation of a search and feed service that pre-computes and maintains embeddings for popular entities, using techniques like cosine similarity, dot product, or Euclidean distance to quickly identify similar web documents, while generating embeddings on the fly for long-tail entities to reduce computational and memory burdens.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If embeddings are pre-computed and maintained for all entities in the large corpus, then similarity determination speed is improved, but memory usage and computational costs increase significantly
Solution Approach 1:
The patent applies local quality by differentiating between popular entities and long-tail entities, pre-computing and storing embeddings only for popular entities that are frequently queried. This selective approach ensures fast similarity determination for common queries while avoiding the memory burden of storing embeddings for all entities in the corpus.
Solution Approach 2:
The patent implements preliminary action by pre-computing and maintaining embeddings for popular entities in advance. This allows the system to quickly retrieve and compare embeddings when similarity determination is needed for frequently accessed entities, improving response time without preparing all possible embeddings beforehand.
2Quantity of substance
If embeddings are pre-computed for popular entities only, then memory usage is reduced, but similarity determination accuracy may deteriorate for long-tail entities
Solution Approach 1:
The patent uses an intermediary approach by introducing a fallback mechanism for long-tail entities. When an embedding is not found in the pre-computed popular entities cache, the system generates an approximate embedding on-demand using techniques like random projection or by leveraging related popular entities. This mediator approach maintains acceptable accuracy for long-tail entities while preserving memory efficiency.
3Loss of energy
If embeddings are generated on the fly for long-tail entities, then computational costs are reduced, but processing time increases for each query
Solution Approach 1:
The patent applies partial action by generating embeddings only when necessary for long-tail entities that are not in the pre-computed cache. For the majority of queries involving popular entities, the system uses the pre-stored embeddings, avoiding redundant computation. This selective generation minimizes both computational cost and processing time overhead.
Data Source
AI summary
An embedding associated with a query comprising one or more query terms is determined to be not stored in memory. An embedding is generated for the one or more query terms. One or more web documents that are similar to the generated embedding are determined. One or more content cards associated with the one or more web documents that are determined to be similar to the generated embedding are provided in a content feed.