What they're describing in the paper goes much further than LangChain. Toolformer is integrating API calls into the training and inference of an LLM.

Short video explainer here: https://youtu.be/ZCdqfIuT81A