Multi-Modal RAG Knowledge Hub for Pharmaceutical Research

A research laboratory inside Kalbe, one of Indonesia's largest pharmaceutical groups, sits on a deep and growing library of biochemistry and pharmaceutical journals. The catch: the most valuable findings are locked inside charts, figures, and experimental tables, exactly the content ordinary text search cannot read. Specialists were spending weeks per initiative just to surface what was already known.

Knowledge was abundant but hard to reach:

• Journals packed with charts, graphs, and experimental tables that text search ignores
• Manual review of hundreds of papers taking weeks per research initiative
• Slow, inefficient retrieval of specific experimental evidence
• Knowledge siloed across documents, limiting reuse and cross-project learning

VentureSEA built a multi-modal Retrieval-Augmented Generation knowledge hub purpose-built for scientific research.

• Built a multi-modal ingestion pipeline for PDFs and image-heavy scientific documents
• Used GPT-4 Vision and GPT-4o to interpret charts, figures, and tables alongside text
• Embedded content with text-embedding-3-large into Pinecone for fast semantic retrieval
• Shipped a RAG assistant on FastAPI and React that answers queries with grounded, reference-backed responses, with documents on Google Cloud Storage and metadata in PostgreSQL
• Attached original pages, charts, and tables as citations, and added an admin console for document management and data integrity
• Ran a validation phase where lab researchers blind-scored answer quality and citation accuracy against source papers before rollout