PRINCE: Bayer's Agentic RAG System for Preclinical Drug Discovery

Title

PRINCE: Bayer's Agentic RAG System for Preclinical Drug Discovery

Content

Bayer deployed PRINCE — the Preclinical Information Center — as an agentic RAG system for its preclinical drug-discovery researchers. Available since early 2024 (with agentic integration added later that year), the platform progressed through three phases: Search (unified gateway over structured study metadata), Ask (RAG over unstructured PDF study reports), and Do (multi-agent orchestration that drafts regulatory documents). The article states that "PRINCE has already shown promising results in enhancing data accessibility and research efficiency at Bayer, transforming how scientists interact with preclinical information," and that "building a production-ready LLM application in a complex enterprise environment like preclinical drug discovery is a journey marked by significant technical and engineering challenges." PRINCE was motivated by three structural problems: data fragmented across silos accumulated through decades of system migrations; keyword search struggling with nuanced scientific terminology; and manual extraction of insights across documents consuming significant researcher time. Because the authoritative "gold standard" knowledge lived inside approved PDF study reports — even where structured metadata was incomplete or incorrect — RAG became the natural bridge from filter-based search to natural-language "ask." The system is built on LangGraph and served via FastAPI with a React UI. Vectors of study reports live in OpenSearch; curated structured metadata is queried through Amazon Athena (Text-to-SQL); agent state is persisted in PostgreSQL via a LangGraph checkpointer; app state sits in DynamoDB. Models from OpenAI, Anthropic, Google, and open-source providers are exposed through an internal GenAI platform with a unified OpenAI-compatible endpoint. Observability runs on CloudWatch and Langfuse; RAGAS powers evaluation — live-traffic evaluation runs daily, dataset evaluations run when workflow, prompts, or models change. The agentic workflow chains five stages: Clarify User Intent disambiguates vague queries ("fail-fast"); Think & Plan performs process reflection (inspired by Anthropic's Think tool) to assess trajectory and choose among tools; the Researcher Agent runs hybrid retrieval (RAG over PDFs plus Text-to-SQL over Athena, with Snowflake Cortex Analyst used for pharmacology queries); the Reflection Agent validates evidence sufficiency; the Writer Agent synthesizes the final answer. Each agent receives a deliberately scoped slice of context — a discipline the authors call "context discipline." Every sentence links to a citation pointing to the exact page and quote in the source PDF. Resilience is layered: retries at LLM-call and LangGraph-node level, automatic fallback to alternate models, and user-initiated retries that resume from the persisted failed node. The authors are evolving the Researcher into a hierarchy of domain-specific sub-agents (toxicology, pharmacology, etc.), each owning its own toolset and prompt instructions. They are also integrating a Named Entity Recognition utility that reads study PDFs and backfills missing metadata into Athena — high-confidence extractions auto-applied, lower-confidence fields quarantined for human review. A companion paper in Frontiers in Artificial Intelligence (DOI 10.3389/frai.2025.1636809) covers product evolution and business impact in more depth.