Conversational Retrieval Accuracy via Vector Embeddings and Reverse Indices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large language models, such as those powering ChatGPT, often 'hallucinate' when processing natural language inputs in e-commerce contexts, leading to inaccurate product information retrieval and inclusion of irrelevant data from different manufacturers, which hampers precise product search and purchase processes.
Innovation Solution
A system that encodes product data using reverse text indices or vector embeddings, stores it in a database, and processes end-user inputs to retrieve relevant product information, ensuring responses include only accurate and contextually relevant products from the original catalog, thereby reducing hallucinations and improving search accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If large language models are used to process natural language inputs for product information retrieval, then the system can generate context-relevant responses, but the models may hallucinate and return inaccurate results
Solution Approach 1:
The patent introduces an intermediary retrieval system between the user's natural language query and the LLM. This intermediary layer searches the product catalog database first, retrieves actual product information, and then feeds this verified data to the LLM for response generation. This mediator prevents the LLM from hallucinating by grounding its responses in actual product data rather than relying solely on its training data.
Solution Approach 2:
The system performs preliminary retrieval of product information from the database before generating a response with the LLM. The retrieval operation executes first to gather accurate product details, which then serve as the foundation for the LLM's response generation. This preliminary action ensures that the LLM works with verified information rather than potentially incorrect data.
2Adaptability or versatility
If LLMs are used to provide product information, then the system can generate natural language responses, but the responses may include information from manufacturers other than the original catalog
Solution Approach 1:
The system implements feedback mechanisms where the retrieved product information is fed back to the LLM as context for response generation. The LLM processes this feedback information and generates responses that are grounded in the actual product catalog data. This feedback loop ensures that the LLM maintains contextual relevance and does not introduce information from external sources.
Solution Approach 2:
The patent applies local quality by tailoring the LLM's response generation to specific local context from the product catalog. Instead of generating generic responses, the system customizes the LLM's output based on the specific product details retrieved from the catalog, ensuring that the response is locally optimized for the actual product being discussed rather than being a general knowledge response.
3Productivity
If the system retrieves product information using traditional search methods, then the process is simple, but the accuracy and relevance of retrieved products is insufficient
Solution Approach 1:
The patent replaces traditional mechanical search methods with a hybrid system that combines vector embeddings, reverse text indices, and LLM-based understanding. Instead of relying solely on keyword matching, the system uses semantic embeddings to capture meaning, reverse indices for efficient lookup, and LLM reasoning for accurate interpretation, thereby improving both speed and precision of product retrieval.
Data Source
AI summary
Systems, methods, and apparatuses for customer engagement that receive a product catalog including information associated with a plurality of products; encode product data by at least one of generating a reverse text index associated with a plurality of products in the product catalog or vectorizing embeddings of the information associated with the plurality of products in the product catalog; store the encoded product data in a product catalog database; receive input from an end user; at least one of convert the end user input to a text query or create input vectors by vectorizing embeddings associated with the input; retrieve a list of products from the product catalog database based on at least one of the text query or the input vectors associated with the input; and output a response to the end user, wherein the response includes a link to information of products in the list of products.