Hybrid Keyword-Vector Search for Precise Fact Retrieval
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing information retrieval systems, particularly in the context of chatbots, struggle to accurately and efficiently pinpoint relevant information chunks from large datasets, leading to AI hallucinations and high computational costs.
Innovation Solution
A novel Intelligent Storage and Retrieval (ISAR) system that combines keyword search methods with vector search methods, incorporating ngram searching, entity counts, and hyponym filtering to identify and return precise relevant facts directly, rather than chunks of information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If traditional keyword-based or vector-based search methods are used to retrieve information from large datasets, then the system can handle large volumes of data, but the retrieval accuracy and relevance of returned information deteriorates
Solution Approach 1:
The patent segments the information retrieval process into multiple independent stages: (1) filtering by entity type and count thresholds, (2) n-gram exact matching for precise keyword retrieval, (3) vector embedding similarity search for semantic relevance, and (4) hybrid ranking combining both approaches. This segmentation allows each stage to specialize in specific aspects, improving overall accuracy while handling large datasets efficiently
Solution Approach 2:
The patent introduces an intermediary vector embedding layer that bridges keyword-based and semantic-based search. The intermediary n-gram representations serve as a mediator between exact keyword matching and semantic vector similarity, enabling the system to leverage both precise keyword retrieval and contextual relevance without sacrificing either approach
2Loss of information
If traditional information retrieval systems return information chunks, then they can provide context, but the computational cost and time required increases
Solution Approach 1:
The patent extracts only the essential information needed for the query by returning structured entity data (type, count, n-gram, vector embedding) rather than complete information chunks. This extraction approach preserves necessary context through entity type classifications and n-gram representations while dramatically reducing the amount of data that needs to be processed and displayed to the user
Solution Approach 2:
Instead of retrieving complete documents or chunks and then filtering for relevant information, the patent inverts the approach by pre-processing and indexing entity data in a structured format that enables direct retrieval of only relevant information. The system retrieves processed entity representations rather than raw text, reversing the traditional retrieval workflow to reduce computational overhead
3Productivity
If AI models generate responses without precise information retrieval, then the generation process is faster, but AI hallucinations increase
Solution Approach 1:
The patent performs preliminary action by pre-processing and indexing all documents into structured entity data with vector embeddings before queries are received. This pre-computation of entity representations, n-gram features, and metadata enables the AI model to quickly retrieve accurate information during query time without slowing down the generation process, as the heavy lifting is done in advance
Data Source
AI summary
Intelligent Storage and Retrieval (ISAR) systems and methods are described, which combine keyword search methods with vector search methods. ISAR also includes additional sub-systems and methods such as ngram searching, entity counts, hyponym filtering, and selective synonym expansion. ISAR pinpoints the exact passages that are relevant to a given query and returns the facts that are precisely relevant to a given query. Some embodiments send the relevant facts themselves in lieu of sending any text chunks. Moreover, ISAR is extremely cost effective. It is extremely fast as well.