Structured RAG Ontology Indexing for Precise Semantic Retrieval
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing retrieval-augmented generation (RAG) techniques suffer from imprecision due to the lack of structural and semantic relationship encoding in vector databases, leading to false-positive retrievals and increased computational costs, especially with large datasets, and cosine similarity cutoffs are arbitrary, resulting in irrelevant record selection.
Innovation Solution
Structured RAG utilizes an ontology index constructed dynamically from conversational turns using a generative language model, encoding semantic relationships and enabling precise retrieval through structured queries, which includes domain-specific agents for enhanced accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If vector databases are used to store and retrieve data in RAG systems, then data storage and retrieval functionality is provided, but retrieval precision deteriorates due to lack of structural and semantic relationship encoding
Solution Approach 1:
The patent segments the monolithic vector database into multiple specialized indices (full-text index, vector index, graph index) that each handle specific types of relationships and queries. This segmentation allows each index to be optimized for its specific purpose, improving retrieval precision while managing complexity through modular design.
Solution Approach 2:
The patent adds a new dimension to data organization by introducing graph-based semantic relationships alongside traditional vector embeddings. This multi-dimensional approach (combining full-text search, vector similarity, and graph relationships) enables retrieval across multiple dimensions of data organization, significantly improving precision.
2Ease of operation
If cosine similarity cutoffs are used for record selection in RAG, then retrieval process is simplified, but retrieval precision deteriorates due to arbitrary cutoff values leading to irrelevant record selection
Solution Approach 1:
The patent introduces graph-based semantic relationships as an intermediary layer between vector similarity scoring and final record selection. This intermediary provides additional contextual validation and filtering, allowing the system to move beyond arbitrary cosine similarity cutoffs while maintaining operational simplicity through automated graph-based ranking.
Solution Approach 2:
The patent changes the selection parameters from relying solely on cosine similarity thresholds to incorporating graph-based relationship strength, node importance scores, and multi-dimensional ranking metrics. This parameter transformation enables more precise record selection without significantly complicating the retrieval process.
3Quantity of substance
If large datasets are processed in RAG systems, then data comprehensiveness is improved, but computational costs increase due to lack of structural encoding
Solution Approach 1:
The patent segments large datasets into multiple specialized indices (full-text, vector, graph) that can be queried independently and in parallel. This segmentation reduces the computational burden on any single index and allows the system to process large datasets more efficiently by distributing the computational workload across multiple optimized structures.
Solution Approach 2:
The patent performs preliminary organization of data into structured graph relationships and semantic hierarchies during the indexing phase. This preliminary structuring enables faster, more efficient querying of large datasets by pre-computing relationships and reducing the computational work required during actual retrieval operations.
4Device complexity
If traditional RAG techniques are used, then implementation simplicity is maintained, but retrieval precision deteriorates due to false-positive retrievals
Solution Approach 1:
The patent creates a composite retrieval system that combines multiple indexing technologies (full-text search, vector embeddings, graph databases) into a unified RAG framework. This composite approach leverages the strengths of each technology to reduce false positives while maintaining implementation feasibility through modular integration.
Solution Approach 2:
The patent introduces graph-based semantic relationships as an intermediary validation layer that filters and ranks retrieval results. This intermediary reduces false positives by providing additional contextual verification beyond simple vector similarity, while the system remains implementable through standardized graph database interfaces.
Data Source
AI summary
A computing system including one or more processing devices configured to extract ontology elements from conversational turns. The ontology elements are extracted at least in part by executing a generative language model. The one or more processing devices assign a respective ontology element type to each ontology element and store the ontology elements in an ontology index. The one or more processing devices receive a user input, and, at the generative language model, compute a structured retrieval-augmented generation (RAG) query. The one or more processing devices execute the structured RAG query over the ontology index to obtain one or more retrieved ontology elements. At the generative language model, the one or more processing devices compute and output a generative language model output based at least in part on the user input and the one or more retrieved ontology elements.