Using Concepts as Contexts for Query Term Substitutions

A novel and efficient method for In-Context Ranking (ICR) using large language models (LLMs) for Information Retrieval (IR). The core problem addressed is the computational cost of standard LLMs for ranking many candidate documents, which scales quadratically with context length due to the attention mechanism. BlockRank solves this by first analyzing LLM attention patterns, revealing inherent inter-document block sparsity and query-document block relevance signals within middle layers

This patent details a system for a large language model (LLM) to respond to user queries by leveraging a custom corpus of documents. The system receives a user query, selects one or more external applications and retrieves relevant documents from the custom corpus based on the query and potentially a context vector or precomputed embeddings. The LLM then generates a response to the user query, conditioned on the retrieved documents, which is subsequently displayed to the user on their client device. Flowcharts and diagrams illustrate the process, including interactions between the client device, the natural language response system, and external applications accessing document embeddings and the custom corpus.

This patent descibes a system and methods for identifying and presenting knowledge elements related to entities, such as musical artists or movies, based on a user's search query and their media consumption history

This US Patent, titled "Searchable Index," describes a system and methods for generating searchable indexes to enhance the efficiency of information retrieval. It outlines how rules derived from a machine-learned model are used to create a token-based index.

This patent application from Google details methods and systems for contextual multimedia search. When a user searches while viewing multimedia content, the system extracts relevant entities from the content and uses them to rewrite the user's query.

This patent publication describes a system for generating search results based on a user's natural language query. The system considers information previously accessed by the user across various devices and applications, such as email and web history.

This patent explores open information extraction, a new extraction paradigm has been developed in which a system makes a single data-driven pass over a corpus of text, extracting a large set of relational tuples without requiring any human input

The Identifying Subjective Attributes by Curation Signal Analysis patent outlines a system and method for automatically identifying and predicting subjective qualities of entities like media or articles.

This patent specification details a system for enhancing search queries by using concepts, which are groups of words with a unified meaning, as contextual clues for suggesting better search terms. The system identifies these concepts within a user's query and then analyses historical search data to find common substitutions for other terms appearing alongside these concepts. By understanding the context provided by a concept like "New York Times", the search engine can more accurately suggest replacing "Puzzle" with "Crossword" compared to a general substitution. This approach allows for more nuanced and relevant query revisions by considering multi-word contexts without excessive computational load, ultimately aiming to improve search result quality.

This patent application details a method for ranking nodes within a linked database, such as the World Wide Web. The described PageRank algorithm assigns importance scores to documents based on the quantity and quality of links pointing to them, essentially a recursive system where a link from a highly-ranked page carries more weight.