Autocurator: AI-Powered Knowledge Extraction for RAG Systems

In the era of information overload, organizations struggle to harness the vast amount of unstructured data stored across presentations, reports, images, and text documents. That's why we created the "Autocurator", an AI-powered tool designed to automatically extract, structure, and curate knowledge from heterogeneous document repositories to support Retrieval-Augmented Generation (RAG) systems.

Autocurator integrates advanced document parsing pipelines, multimodal AI models, and semantic structuring techniques to convert diverse content - including text, slides, tables, and diagrams - into machine-readable knowledge. This enables downstream RAG systems to query not only text-based insights but also visual and conceptual knowledge that traditionally remained inaccessible.
Our system employs a multi-stage pipeline: (1) document ingestion and format normalization, (2) de-duplication of redundant and conflicting information (3) multimodal content understanding using large language and vision models, (4) entity and relationship extraction with human-in-the-loop validation, and (5) generation of structured outputs optimized for retrieval.
We will showcase Autocurator’s effectiveness on large enterprise document corpora, showcasing significant gains in retrieval precision and generative quality across several applied AI use cases.
By bridging unstructured data and structured knowledge, Autocurator provides a scalable and transparent foundation for next-generation knowledge management and reasoning systems.

Eleonora Vardè is a Lead data scientist @ BCG X Milan, specialized in Customer Service and Gen AI topics

Marco P. Abrate is a Visiting AI engineer at BCG X and a PhD candidate in Neuroscience and AI at University College London.