RAG Recommendation Retrieval Using Vector Search for Accurate Personalization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing recommender systems fail to provide accurate and timely recommendations due to reliance on outdated large language models (LLMs) and lack of personalization, often generating incorrect results and requiring resource-intensive retraining to update item information.
Innovation Solution
A retrieval-augmented generation approach that uses a vector database to supplement an LLM, allowing for the extraction of query tags and generation of query vectors to identify candidate items, ensuring recommendations are based on up-to-date and accurate item information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If an LLM is used to generate recommendations based on its default knowledge base, then the system can provide conversational natural language interface, but the recommendations may be inaccurate due to hallucinations and outdated information
Solution Approach 1:
The patent introduces a retrieval-augmented generation system where an LLM serves as an intermediary between the user's natural language query and a structured recommendation database. The LLM generates queries based on user preferences, these queries are executed against the database to retrieve accurate item information, and the results are returned to the user. This intermediary approach allows the system to maintain conversational ease while ensuring recommendation accuracy through database verification.
2Reliability
If an LLM is trained with item information, then it can generate recommendations, but re-training is resource-intensive and time-consuming, causing the data to become outdated
Solution Approach 1:
The patent segments the recommendation system into two independent components: an LLM for natural language processing and a separate structured database for item information storage. This segmentation allows the database to be updated independently without re-training the LLM. The LLM remains responsible for understanding user preferences and generating queries, while the database stores and provides access to current item information, enabling timely updates without resource-intensive re-training.
3Productivity
If existing recommender systems use item selections from another system as preference information, then they can generate recommendations, but they do not provide engaging experiences or recommendations particularly similar to personal preferences
Solution Approach 1:
The patent transforms the recommendation approach by changing the parameters used to represent user preferences. Instead of relying on simple item selection history, the system uses the LLM to generate detailed preference descriptions from natural language inputs. These preference descriptions are then translated into structured queries that can be executed against the database, allowing for highly personalized recommendations that capture nuanced user preferences rather than just categorical similarities.
Data Source
AI summary
In various examples, a technique for recommending items is disclosed that includes receiving a request to recommend an item. The technique further includes generating, based on the request, one or more query tags, the one or more query tags including one or more preferred tags, at least one of the preferred tags specifying an item of interest. The technique also includes generating one or more subqueries based on the one or more query tags. The technique still further includes generating, based on the one or more subqueries, a query vector in an embedding space of the vector database. The technique further includes identifying, in the vector database and using the query vector, a plurality of candidate items. The technique still further includes selecting a recommended item from the plurality of candidate items based on the request.