Optimized retrieval-augmented generation for complex documents

Abstract

The subject technology relates to optimized retrieval-augmented generation (RAG) for complex documents. An example method includes generating partial vector embeddings from a document, the partial vector embeddings being representative of respective defined sections of the document; assigning a data block, including a representation of tabular data present in the document, to a partial vector embedding, of the partial vector embeddings and corresponding to a section, of the defined sections of the document, in which the tabular data appears; and generating, in response to an input query to a machine learning model, a query output based on selected output data, where the selected output data includes selected vector embeddings, of a group of vector embeddings that includes the partial vector embeddings, and the data block, and where the selected vector embeddings are selected based on a relevancy criterion and a data size constraint of the machine learning model.