Combining vector search and llms for entity matching on large scale

Abstract

Certain aspects provide a method including: receiving a user request from a first user to create a local contact for an entity; providing the user request to a first language model configured to access to a database of network contacts; receiving a list of candidate entities from the first language model; and in response to determining that a similarity score for at least one candidate entity falls between a minimum similarity score threshold and a maximum similarity score threshold, providing the user request, the list of candidate entities, and the prompt to a second language model; receiving a response from the second language model comprising a predicted entity; and generating for display a user recommendation to replace the local contact of the entity indicated in the user request with a network contact corresponding to the predicted entity.