The CoD prompt instructs highly powered LLMs such as GPT-4 to produce an initial sparse, verbose summary of an article containing only a few entities. It then iteratively identifies 1–3 missing entities and fuses them into a rewrite of the previous summary in the same number of words.

LangChain excels at chaining LLMs together using agents to delegate actions to models. Its use cases emphasize prompt optimization and context-aware information retrieval/generation; however, with its Pythonic highly modular interface and its huge collection of tools, it is the number-one tool to implement complex business logic.

code-bison generates code from natural language descriptions, with a max input of 4,096 tokens and an output of 2,048 tokens. codechat-bison is a chatbot that is fine-tuned to help with code-related questions. It has an input limit of 4,096 tokens and an output limit of 2,048 tokens. code-gecko suggests code completions.