使用 LangChain 中的 KnowledgeGraphMemory，检测其处理对话信息的能力。 from langchain.memory import ConversationKGMemory # 知识图谱记忆 llm = OpenAI(temperature=0) memory=ConversationKGMemory(llm=llm) template = """The following is a conversation between a human and an AI. The AI provides lots of specific details from its context. AI uses all the information contained in "Related Information", "Current Conversation" and "Context" If the AI does not know the answer to a question, it truthfully says it does not know，and does not hallucinate. answer in 20 word. 对话如下 第一轮 Q1 (喂给了AI 这些信息，牵扯到几个方向，试AI 如何提取并处理这些信息) input="I am James and I'm helping Will. He's an engineer working in IBM's electronics department. He encountered a problem in the 'Knowledge Graph Memory' project" A1 > Finished chain. " Hi James, I understand that Will is having difficulty with the 'Knowledge Graph Memory' project. I'm not familiar with this project, but I can help you find the resources you need to solve the problem." (回答中AI 分出了James 和 Will，其它的大差不差。) 第二轮对话 Q2 input="What do you know about Will?" A2 > Finished chain. " Will is an engineer working in IBM's electronics department. He is currently working on a Knowledge Graph Memory project and has encountered a problem." (AI 充分利用了在第一轮对话中得到的有关Will 的信息。) 第三轮对话 A3 input="What do you know about me?" Q3 > Finished chain. " I know that you are talking to me right now and that you asked me a question about Will. I also know that Will is an engineer working in IBM's electronics department and is working on the Knowledge Graph Memory project." (AI 没有捕捉到 me, I, James 之间的关系，尽管它用了Hi James, ) 第四轮对话 A4 (继续追问上轮中出现的漏洞) input="What is my name?" Q4 > Finished chain. "Your name is John Smith." (AI 彻底错了) 结果有点意外，是算法有错还是参数设置有待改进? 共指消解(coreference/entity resolution)在知识图谱记忆中是最基本的方法，整合同一实体的不同称谓。