LLMs as Chemistry Assistants

Recording

Learning Objectives

• Explain retrieval-augmented generation (RAG) and implement a simple chemistry paper Q&A system
• Describe how tool use (function calling) enables LLMs to interact with RDKit, databases, and simulation codes
• Identify failure modes of LLMs on chemistry tasks: hallucinated reactions, incorrect SMILES, wrong stereochemistry
• Evaluate LLM outputs critically using literature cross-referencing and automated validity checks

Key Takeaways

• Takeaway 1. RAG grounds LLM answers in retrieved documents, dramatically reducing hallucination rates on factual chemistry questions — but retrieval quality is the bottleneck, not the language model.

• Takeaway 2. LLMs are poor chemistry calculators: they hallucinate reaction mechanisms, confidently produce invalid SMILES, and make errors in stoichiometry. Always validate programmatically with RDKit or similar tools.

• Takeaway 3. Tool use transforms LLMs from text generators into orchestrators of computation — they can call a yield predictor, retrieve a crystal structure, or run a retrosynthesis engine without the user writing any integration code.

• Takeaway 4. The bottleneck in LLM-assisted chemistry is not language understanding but knowledge currency: models have training cutoffs, cannot access paywalled journals, and lack lab-specific institutional knowledge unless explicitly provided.

Further Reading & Resources