EMBL AI Librarian: A Natural-Language Gateway to Biomedical Literature

AI agents increasingly need to search scientific literature, but existing databases like Europe PMC—which holds over 40 million records—were designed for human keyword searches, not machines. An agent must learn complex query syntax, issue multiple searches, and parse entire papers to extract relevant evidence. The authors introduce EMBL AI Librarian, a knowledge layer that translates natural-language questions into targeted, evidence-backed answers by orchestrating searches and reading papers automatically.

How It Works

A single large language model (LLM) sits between the agent and Europe PMC, acting as an intelligent middleman. When an agent asks a question in plain English, the LLM first plans a set of complementary subqueries—breaking the question into searches that will find relevant papers from different angles. These subqueries are executed against the live Europe PMC search engine. The LLM then reads the returned papers, identifies passages that actually answer the original question, and surfaces that evidence to the agent.

The key insight is that an LLM is better suited than humans to orchestrate this multi-step retrieval-and-reading process. Rather than forcing every agent to learn Europe PMC's syntax or implement search heuristics independently, Librarian handles the complexity once, in a way that's reusable across many downstream tasks.

Empirical Results

The authors tested Librarian on four evaluation benchmarks spanning different real-world needs:

• Literature synthesis and claim verification: On ScholarQABench, Librarian improved Citation F1 (a metric measuring how well cited papers match ground-truth evidence) by over 16 points compared to recent published baselines. When integrated into an existing claim-verification pipeline, it increased agreement with expert consensus.

• Open-domain question answering: On the LitQA2 benchmark, a GPT-4 agent scored approximately 8 points higher when using Librarian as its retrieval layer versus standard web search.

• Downstream biology tasks: The system was also evaluated on practical life-sciences questions such as protocol recommendations and sequence manipulation, demonstrating utility beyond academic Q&A.

Why It Matters

Life-sciences AI agents—tools that automate research synthesis, hypothesis generation, and experimental design—are growing rapidly. These agents need reliable access to evidence. Librarian removes friction by eliminating the need for agents to learn bespoke query syntax or implement domain-specific search logic themselves. By centralizing the translation from natural language to evidence retrieval, it lowers the barrier for building better scientific agents.

The improvements are concrete: 16-point gains in citation matching and 8-point gains in open-domain QA are meaningful margins in competitive benchmarks. Crucially, the method works with an existing, live database (Europe PMC) rather than requiring expensive retraining or proprietary indexes, making it immediately deployable.

The authors have released code publicly, enabling researchers to build on this approach. As AI increasingly mediates access to scientific knowledge, ensuring that mediation is accurate and efficient—rather than forcing every agent to reinvent search—is both practical and foundational for the field.