Drug Discovery Acceleration with In-House RAG
Business challenge: R&D teams were spending too much time managing a disorganized knowledge base consisting of over 500,000 documents. Project objectives: Speed up in silico hypothesis testing and compound analysis. Client requirement: Create an AI-driven research assistant that provides accurate, real-time responses to scientific inquiries using both internal and external data sources. Key performance indicator: Achieve a 50% reduction in analytical workload. Summary: We implemented a RAG assistant utilizing vector search across internal reports, PubChem, and scientific literature. The solution featured a fine-tuned large language model (LLM) specifically designed for pharmaceutical terminology. The web-based chat interface presents answers along with complete citations and explanations. Interface display: A chat window showcasing responses, citations, and links to internal documents. Main features: A RAG-based assistant that utilizes semantic search, provides citations, and conducts scientific Q&A through natural language processing. Administrative tools are included for managing data, models, and users, as well as for monitoring, feedback, and compliance. Technologies utilized: Python, LangChain, LangGraph, Weaviate, FastAPI, React, self-hosted LLM (Qwen 2.5 32b), LangFuse, and a Grafana stack (Prometheus, Loki, Tempo). Team composition: 1 Solution Architect, 2 ML Engineers, 1 Backend Engineer, and 1 QA. Results: Reduced hypothesis validation time from 4 hours to 1.5 hours, leading to increased confidence in outcomes. Client feedback: “The assistant comprehends our data more effectively than most new hires.”