Chatbot using retrieval-augmented generation

A method, comprising:

• receiving, by one or more processing circuits, from a first computing device, a natural language query for information associated with a first entity;
• generating, by the one or more processing circuits, a vector embedding based on the natural language query;
• searching, by the one or more processing circuits, responsive to generating the vector embedding, a vector database having a plurality of vector embeddings to identify a set of vector embeddings of the plurality of vector embeddings that are associated with the vector embedding, each vector embedding of the set of vector embeddings corresponding to an entity of a plurality of entities;
• identifying, by the one or more processing circuits, for each vector embedding of the set of vector embeddings, metadata that is linked to a corresponding vector embedding of the set of vector embeddings within the vector database;
• selecting, by the one or more processing circuits, based on the metadata for each vector embedding of the set of vector embeddings, a first vector embedding of the set of vector embeddings that corresponds to the first entity, selecting the first vector embedding comprising determining that the first vector embedding corresponds to the first entity using the metadata for the first vector embedding;
• generating, by the one or more processing circuits, an input based on the natural language query using context provided by the first vector embedding and the metadata for the first vector embedding;
• generating, by the one or more processing circuits, by providing the input to a machine learning model, a response to the natural language query;
• and
• providing, by the one or more processing circuits, to the first computing device, an electronic message that includes the response to the natural language query.

One or more non-transitory storage media storing instructions thereon that, when executed by one or more processors, cause the one or more processors to perform operations comprising:

• receiving, from a first computing device, a natural language query for information associated with a first entity;
• generating a vector embedding based on the natural language query;
• searching, responsive to generating the vector embedding, a vector database having a plurality of vector embeddings to identify a set of vector embeddings of the plurality of vector embeddings that are associated with the vector embedding, each vector embedding of the set of vector embeddings corresponding to a respective entity of a plurality of entities;
• identifying, for each vector embedding of the set of vector embeddings, metadata that is linked to a corresponding vector embedding of the set of vector embeddings within the vector database;
• selecting, based on the metadata for each vector embedding of the set of vector embeddings, a first vector embedding of the set of vector embeddings that corresponds to the first entity, selecting the first vector embedding comprising determining that the first vector embedding corresponds to the first entity using the metadata for the first vector embedding;
• generating a prompt based on the natural language query using context provided by the first vector embedding and the metadata for the first vector embedding;
• generating, by providing the prompt to a large language model, a response to the natural language query;
• and
• providing, to the first computing device, an electronic message that includes the response to the natural language query.