ChemCrow is an open-source package designed to accurately solve reasoning-intensive chemistry tasks. It is built on Langchain and integrates a variety of chemistry tools to enhance the reasoning capabilities of large language models in the field of chemistry.
Purpose: ChemCrow is mainly used to handle chemistry problems that require multi-step reasoning, such as retrosynthesis planning, reaction product prediction, and chemistry information retrieval. It assists in completing these tasks by invoking different chemistry tools and databases.
Main capabilities: The chemistry tools used by this package include RDKit (a cheminformatics toolkit), paper-qa (a literature question-answering tool), and related chemistry databases such as PubChem and chem-space. Together, these tools support chemistry reasoning tasks.
How to use: Users first need to set API keys in the environment. The Serp API is optional. By default, ChemCrow relies on the RXN4Chem API for retrosynthesis planning and reaction product prediction, which may be slow and requires the corresponding API key. Users can also choose to self-host these tools by running prebuilt Docker images. After installation, ChemCrow can be used in the usual way.
Applicable scenarios: Suitable for chemistry research, teaching, and scenarios that require automated chemistry reasoning.
Limitations: Due to API usage restrictions, this package does not include all the tools described in the ChemCrow paper, so the results from this repository will not be exactly the same as those in the paper. All experiments have been published as ChemCrow runs. In addition, the project homepage provides an entry point for trying it on HuggingFace.
Citation information: Bran, Andres M., et al. "ChemCrow: Augmenting large-language models with chemistry tools." arXiv preprint arXiv:2304.05376 (2023).
