{"id":1358,"date":"2024-05-06T05:31:45","date_gmt":"2024-05-06T05:31:45","guid":{"rendered":"https:\/\/www.nicekj.com\/?p=1358"},"modified":"2024-05-06T05:32:01","modified_gmt":"2024-05-06T05:32:01","slug":"langchainxiliejiaochengshiyonglangchaindellmsjinxingduihuajiyi-2","status":"publish","type":"post","link":"https:\/\/www.nicekj.com\/langchainxiliejiaochengshiyonglangchaindellmsjinxingduihuajiyi-2.html","title":{"rendered":"LangChain\u7cfb\u5217\u6559\u7a0b\uff1a\u4f7f\u7528Langchain\u7684LLMs\u8fdb\u884c\u5bf9\u8bdd\u8bb0\u5fc6"},"content":{"rendered":"

\u5bf9\u8bdd\u8bb0\u5fc6\u662f\u6307\u804a\u5929\u673a\u5668\u4eba\u5982\u4f55\u4ee5\u5bf9\u8bdd\u65b9\u5f0f\u54cd\u5e94\u591a\u4e2a\u67e5\u8be2\u3002\u5b83\u4f7f\u5bf9\u8bdd\u8fde\u8d2f\uff0c\u5e76\u4e14\u5982\u679c\u6ca1\u6709\u5b83\uff0c\u6bcf\u4e2a\u67e5\u8be2\u90fd\u5c06\u88ab\u89c6\u4e3a\u5b8c\u5168\u72ec\u7acb\u7684\u8f93\u5165\uff0c\u800c\u4e0d\u8003\u8651\u8fc7\u53bb\u7684\u4ea4\u4e92\u3002<\/p>\n

\n

\"image.png\"<\/figure>\n<\/p>\n

\u8fd9\u79cd\u8bb0\u5fc6\u4f7f\u5f97\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u80fd\u591f\u8bb0\u4f4f\u4e0e\u7528\u6237\u4e4b\u524d\u7684\u4e92\u52a8\u3002\u9ed8\u8ba4\u60c5\u51b5\u4e0b\uff0cLLMs\u662f\u65e0\u72b6\u6001\u7684\uff0c\u8fd9\u610f\u5473\u7740\u6bcf\u4e2a\u4f20\u5165\u7684\u67e5\u8be2\u90fd\u662f\u72ec\u7acb\u5904\u7406\u7684\uff0c\u4e0d\u8003\u8651\u5176\u4ed6\u4e92\u52a8\u3002\u5bf9\u4e8e\u4e00\u4e2a\u65e0\u72b6\u6001\u7684\u4ee3\u7406\uff0c\u552f\u4e00\u5b58\u5728\u7684\u662f\u5f53\u524d\u7684\u8f93\u5165\uff0c\u6ca1\u6709\u5176\u4ed6\u4fe1\u606f\u3002<\/p>\n

\u6709\u8bb8\u591a\u5e94\u7528\u7a0b\u5e8f\u5728\u5176\u4e2d\u8bb0\u4f4f\u4e4b\u524d\u7684\u4e92\u52a8\u975e\u5e38\u91cd\u8981\uff0c\u6bd4\u5982\u804a\u5929\u673a\u5668\u4eba\u3002\u5bf9\u8bdd\u8bb0\u5fc6\u5141\u8bb8\u6211\u4eec\u505a\u5230\u8fd9\u4e00\u70b9\u3002<\/p>\n

\u6709\u51e0\u79cd\u65b9\u6cd5\u53ef\u4ee5\u5b9e\u73b0\u5bf9\u8bdd\u8bb0\u5fc6\u3002\u5728LangChain<\/a>\u7684\u80cc\u666f\u4e0b\uff0c\u5b83\u4eec\u90fd\u662f\u57fa\u4e8eConversationChain\u6784\u5efa\u7684\u3002<\/p>\n

ConversationChain\uff08\u5bf9\u8bdd\u94fe\uff09<\/h1>\n

\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7\u521d\u59cb\u5316ConversationChain\uff08\u5bf9\u8bdd\u94fe\uff09\u5f00\u59cb\u3002\u6211\u4eec\u5c06\u4f7f\u7528OpenAI\u7684text-davinci-003\u4f5c\u4e3aLLM\uff0c\u4f46\u4e5f\u53ef\u4ee5\u4f7f\u7528\u5176\u4ed6\u6a21\u578b\uff0c\u5982gpt-3.5-turbo\u3002<\/p>\n

<\/div>
ini<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>from langchain import OpenAI<\/span>\nfrom langchain.chains import ConversationChain<\/span>\n<\/span>\n# first initialize the large language model<\/span><\/span>\nllm<\/span> = OpenAI(<\/span>\n\ttemperature<\/span>=0<\/span>,<\/span>\n\topenai_api_key<\/span>=\"OPENAI_API_KEY\"<\/span>,<\/span>\n\tmodel_name<\/span>=\"text-davinci-003\"<\/span><\/span>\n)<\/span>\n<\/span>\n# now initialize the conversation chain<\/span><\/span>\nconversation<\/span> = ConversationChain(llm=llm)<\/span>\n<\/code><\/pre>\n
\u6211\u4eec\u53ef\u4ee5\u8fd9\u6837\u67e5\u770bConversationChain\u4f7f\u7528\u7684\u63d0\u793a\u6a21\u677f\uff1a<\/p>\n
In[8]:<\/p>\n
<\/div>
scss<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>print<\/span>(conversation.prompt.template)<\/span>\n<\/code><\/pre>\n
Out[8]:<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>The following is a<\/span> friendly conversation between a<\/span> human and an AI. The AI is talkative and provides lots of specific details<\/span> from<\/span> its context. If the AI does not know the answer to<\/span> a<\/span> question, it truthfully says it does not know.<\/span>\n<\/span>\nCurrent conversation:<\/span>\n{history}<\/span>\nHuman: {input<\/span>}<\/span>\nAI:<\/span>\n<\/code><\/pre>\n
\u5728\u8fd9\u91cc\uff0c\u63d0\u793a\u901a\u8fc7\u544a\u8bc9\u6a21\u578b\u4ee5\u4e0b\u5185\u5bb9\u662f\u4e00\u4e2a\u4eba\u7c7b\uff08\u6211\u4eec\uff09\u548c\u4e00\u4e2aAI\uff08text-davinci-003\uff09\u4e4b\u95f4\u7684\u5bf9\u8bdd\u6765\u5f15\u5bfc\u6a21\u578b\u3002\u63d0\u793a\u8bd5\u56fe\u901a\u8fc7\u9648\u8ff0\u4ee5\u4e0b\u5185\u5bb9\u6765\u51cf\u5c11\u5e7b\u89c9\uff08\u6a21\u578b\u7f16\u9020\u4e8b\u7269\u7684\u60c5\u51b5\uff09\uff1a<\/p>\n
\n
“If the AI does not know the answer to a question, it truthfully says it does not know.”<\/p>\n<\/blockquote>\n
\u8fd9\u6709\u52a9\u4e8e\u51cf\u5c11\u5e7b\u89c9\u7684\u53d1\u751f\uff0c\u4f46\u5e76\u4e0d\u80fd\u5b8c\u5168\u89e3\u51b3\u5e7b\u89c9\u7684\u95ee\u9898\uff0c\u4f46\u6211\u4eec\u5c06\u628a\u8fd9\u4e2a\u95ee\u9898\u7559\u5230\u672a\u6765\u7ae0\u8282\u8ba8\u8bba\u3002<\/p>\n
\u5728\u521d\u59cb\u63d0\u793a\u4e4b\u540e\uff0c\u6211\u4eec\u770b\u5230\u4e24\u4e2a\u53c2\u6570\uff1b{history} \u548c {input}\u3002{input} \u662f\u6211\u4eec\u5c06\u653e\u7f6e\u6700\u65b0\u7684\u4eba\u7c7b\u67e5\u8be2\u7684\u5730\u65b9\uff1b\u5b83\u662f\u8f93\u5165\u5230\u804a\u5929\u673a\u5668\u4eba\u6587\u672c\u6846\u7684\u8f93\u5165\u5185\u5bb9\uff1a<\/p>\n
\n
\"image.png\"<\/figure>\n<\/p>\n
{history} \u662f\u7528\u4e8e\u4f7f\u7528\u5bf9\u8bdd\u8bb0\u5fc6\u7684\u5730\u65b9\u3002\u5728\u8fd9\u91cc\uff0c\u6211\u4eec\u63d0\u4f9b\u4e86\u5173\u4e8e\u4eba\u7c7b\u548cAI\u4e4b\u95f4\u5bf9\u8bdd\u5386\u53f2\u7684\u4fe1\u606f\u3002<\/p>\n
\u8fd9\u4e24\u4e2a\u53c2\u6570\u2014\u2014{history} \u548c {input}\u2014\u2014\u4f20\u9012\u7ed9\u4e86\u6211\u4eec\u521a\u521a\u770b\u5230\u7684\u63d0\u793a\u6a21\u677f\u4e2d\u7684LLM\uff0c\u800c\u6211\u4eec\uff08\u5e0c\u671b\u5982\u6b64\uff09\u8fd4\u56de\u7684\u8f93\u51fa\u53ea\u662f\u5bf9\u8bdd\u7684\u9884\u6d4b\u7ee7\u7eed\u90e8\u5206\u3002<\/p>\n
\u5bf9\u8bdd\u8bb0\u5fc6\u7684\u5f62\u5f0f<\/h1>\n
\u6211\u4eec\u53ef\u4ee5\u5728ConversationChain\u4e2d\u4f7f\u7528\u591a\u79cd\u7c7b\u578b\u7684\u5bf9\u8bdd\u8bb0\u5fc6\u3002\u5b83\u4eec\u4fee\u6539\u4f20\u9012\u7ed9{history}\u53c2\u6570\u7684\u6587\u672c\u3002<\/p>\n
\u5bf9\u8bdd\u7f13\u51b2\u8bb0\u5fc6<\/h2>\n
ConversationBufferMemory\u662fLangChain\u4e2d\u6700\u76f4\u63a5\u7684\u5bf9\u8bdd\u8bb0\u5fc6\u5f62\u5f0f\u3002\u6b63\u5982\u6211\u4eec\u4e0a\u9762\u6240\u63cf\u8ff0\u7684\uff0c\u8fc7\u53bb\u4eba\u7c7b\u548cAI\u4e4b\u95f4\u7684\u5bf9\u8bdd\u7684\u539f\u59cb\u8f93\u5165\u4ee5\u5176\u539f\u59cb\u5f62\u5f0f\u4f20\u9012\u5230{history}\u53c2\u6570\u4e2d\u3002<\/p>\n
In[11]:<\/p>\n
<\/div>
ini<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>from langchain.chains.conversation.memory import ConversationBufferMemory<\/span>\n<\/span>\nconversation_buf<\/span> = ConversationChain(<\/span>\n    llm<\/span>=llm,<\/span>\n    memory<\/span>=ConversationBufferMemory()<\/span>\n)<\/span>\n<\/code><\/pre>\n
In[32]:<\/p>\n
<\/div>
scss<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>conversation_buf<\/span>(\"Good morning AI!\")<\/span>\n<\/code><\/pre>\n
Out[32]:<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>{'input<\/span>': 'Good morning AI!'<\/span>,<\/span>\n 'history'<\/span>: ''<\/span>,<\/span>\n 'response'<\/span>: \" Good morning! It's a beautiful day today, isn't it? How can I help you?\"<\/span>}<\/span>\n<\/code><\/pre>\n
\u6211\u4eec\u8fd4\u56de\u4e86\u5bf9\u8bdd\u4ee3\u7406\u7684\u7b2c\u4e00\u4e2a\u54cd\u5e94\u3002\u8ba9\u6211\u4eec\u7ee7\u7eed\u5bf9\u8bdd\uff0c\u7f16\u5199\u53ea\u6709\u5728\u8003\u8651\u5bf9\u8bdd\u5386\u53f2\u65f6LLM\u624d\u80fd\u56de\u7b54\u7684\u63d0\u793a\u3002\u6211\u4eec\u8fd8\u6dfb\u52a0\u4e86\u4e00\u4e2acount_tokens\u51fd\u6570\uff0c\u4ee5\u4fbf\u6211\u4eec\u53ef\u4ee5\u67e5\u770b\u6bcf\u6b21\u4ea4\u4e92\u4f7f\u7528\u4e86\u591a\u5c11\u6807\u8bb0\u3002<\/p>\n
In[6]:<\/p>\n
<\/div>
python<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>from<\/span> langchain.callbacks import<\/span> get_openai_callback<\/span>\n<\/span>\ndef<\/span> count_tokens<\/span>(chain, query<\/span>):<\/span>\n    with<\/span> get_openai_callback() as<\/span> cb:<\/span>\n        result = chain.run(query)<\/span>\n        print<\/span>(f'Spent a total of {cb.total_tokens}<\/span> tokens'<\/span>)<\/span>\n<\/span>\n    return<\/span> result<\/span>\n<\/code><\/pre>\n
In[33]:<\/p>\n
<\/div>
scss<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>count_tokens<\/span>(<\/span>\n    conversation_buf, <\/span>\n    \"My interest here is to explore the potential of integrating Large Language Models with external knowledge\"<\/span>\n)<\/span>\n<\/code><\/pre>\n
Out[33]:<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>Spent a<\/span> total of 179<\/span> tokens<\/span>\n<\/code><\/pre>\n
Out[33]:<\/p>\n
<\/div>
scala<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>' Interesting<\/span>! Large<\/span> Language<\/span> Models<\/span> are a type<\/span> of<\/span> artificial<\/span> intelligence<\/span> that<\/span> can<\/span> process<\/span> natural<\/span> language<\/span> and<\/span> generate<\/span> text<\/span>. They<\/span> can<\/span> be<\/span> used<\/span> to<\/span> generate<\/span> text<\/span> from<\/span> a<\/span> given<\/span> context<\/span>, or<\/span> to<\/span> answer<\/span> questions<\/span> about<\/span> a<\/span> given<\/span> context<\/span>. Integrating<\/span> them<\/span> with<\/span> external<\/span> knowledge<\/span> can<\/span> help<\/span> them<\/span> to<\/span> better<\/span> understand<\/span> the<\/span> context<\/span> and<\/span> generate<\/span> more<\/span> accurate<\/span> results<\/span>. Is<\/span> there<\/span> anything<\/span> else<\/span> I<\/span> can<\/span> help<\/span> you<\/span> with<\/span>?<\/span>'<\/span><\/span>\n<\/code><\/pre>\n
In[34]:<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>count_tokens(<\/span>\n    conversation_buf,<\/span>\n    \"I<\/span> just want to<\/span> analyze the different possibilities. What can you think of?\"<\/span>\n)<\/span>\n<\/code><\/pre>\n
Out[34]:<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>Spent a<\/span> total of 268<\/span> tokens<\/span>\n<\/code><\/pre>\n
Out[34]:<\/p>\n
<\/div>
vbnet<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>' Well, integrating Large Language Models with external knowledge can open up a lot of possibilities. For example, you could use them to generate more accurate and detailed summaries of text, or to answer questions about a given context more accurately. You could also use them to generate more accurate translations, or to generate more accurate predictions about future events.'<\/span><\/span>\n<\/code><\/pre>\n
In[35]:<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>count_tokens(<\/span>\n    conversation_buf, <\/span>\n    \"Which data source types could be used to<\/span> give context to<\/span> the model?\"<\/span>\n)<\/span>\n<\/code><\/pre>\n
Out[35]:<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>Spent a<\/span> total of 360<\/span> tokens<\/span>\n<\/code><\/pre>\n
Out[35]:<\/p>\n
<\/div>
vbnet<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>'  There are a variety of data sources that could be used to give context to a Large Language Model. These include structured data sources such as databases, unstructured data sources such as text documents, and even audio and video data sources. Additionally, you could use external knowledge sources such as Wikipedia or other online encyclopedias to provide additional context.'<\/span><\/span>\n<\/code><\/pre>\n
In[36]:<\/p>\n
<\/div>
scss<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>count_tokens<\/span>(<\/span>\n    conversation_buf, <\/span>\n    \"What is my aim again?\"<\/span>\n)<\/span>\n<\/code><\/pre>\n
Out[36]:<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>Spent a<\/span> total of 388<\/span> tokens<\/span>\n<\/code><\/pre>\n
Out[36]:<\/p>\n
<\/div>
vbnet<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>' Your aim is to explore the potential of integrating Large Language Models with external knowledge.'<\/span><\/span>\n<\/code><\/pre>\n
LLM\u663e\u7136\u53ef\u4ee5\u8bb0\u4f4f\u5bf9\u8bdd\u7684\u5386\u53f2\u3002\u8ba9\u6211\u4eec\u770b\u770bConversationBufferMemory\u662f\u5982\u4f55\u5b58\u50a8\u8fd9\u4e2a\u5bf9\u8bdd\u5386\u53f2\u7684\uff1a<\/p>\n
In[37]:<\/p>\n
<\/div>
scss<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>print<\/span>(conversation_buf.memory.buffer)<\/span>\n<\/code><\/pre>\n
Out[37]:<\/p>\n
<\/div>
vbnet<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div><\/span>\nHuman:<\/span> Good morning AI!<\/span>\nAI:<\/span>  Good morning! It's a beautiful day today, isn't it? How can I help you?<\/span><\/span>\nHuman:<\/span> My interest here is<\/span> to<\/span> explore the potential of<\/span> integrating Large Language Models with<\/span> external knowledge<\/span>\nAI:<\/span>  Interesting! Large Language Models are a type of<\/span> artificial intelligence that can process natural language and<\/span> generate text<\/span>. They can be used to<\/span> generate text<\/span> from<\/span> a given context, or<\/span> to<\/span> answer questions about a given context. Integrating them with<\/span> external knowledge can help them to<\/span> better understand the context and<\/span> generate more accurate results. Is<\/span> there anything else<\/span> I can help you with<\/span>?<\/span>\nHuman:<\/span> I just want to<\/span> analyze the different possibilities. What can you think of<\/span>?<\/span>\nAI:<\/span>  Well, integrating Large Language Models with<\/span> external knowledge can open up a lot of<\/span> possibilities. For<\/span> example, you could use them to<\/span> generate more accurate and<\/span> detailed summaries of<\/span> text<\/span>, or<\/span> to<\/span> answer questions about a given context more accurately. You could also use them to<\/span> generate more accurate translations, or<\/span> to<\/span> generate more accurate predictions about future events.<\/span>\nHuman:<\/span> Which data source types could be used to<\/span> give context to<\/span> the model?<\/span>\nAI:<\/span>   There are a variety of<\/span> data sources that could be used to<\/span> give context to<\/span> a Large Language Model. These include structured data sources such as<\/span> databases, unstructured data sources such as<\/span> text<\/span> documents, and<\/span> even audio and<\/span> video data sources. Additionally, you could use external knowledge sources such as<\/span> Wikipedia or<\/span> other online encyclopedias to<\/span> provide additional context.<\/span>\nHuman:<\/span> What is<\/span> my aim again?<\/span>\nAI:<\/span>  Your aim is<\/span> to<\/span> explore the potential of<\/span> integrating Large Language Models with<\/span> external knowledge.<\/span>\n<\/code><\/pre>\n
\u6211\u4eec\u53ef\u4ee5\u770b\u5230\u7f13\u51b2\u4fdd\u5b58\u4e86\u804a\u5929\u5386\u53f2\u4e2d\u7684\u6bcf\u4e00\u6b21\u4e92\u52a8\u3002\u8fd9\u79cd\u65b9\u6cd5\u6709\u4e00\u4e9b\u4f18\u70b9\u548c\u7f3a\u70b9\u3002\u7b80\u800c\u8a00\u4e4b\uff0c\u5b83\u4eec\u662f\uff1a<\/p>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Pros<\/th>Cons<\/th><\/tr><\/thead>
Storing everything gives the LLM the maximum amount of information<\/td>More tokens mean slowing response times and higher costs<\/td><\/tr>
Storing everything is simple and intuitive<\/td>Long conversations cannot be remembered as we hit the LLM token limit (4096 tokens for text-davinci-003 and gpt-3.5-turbo)<\/td><\/tr><\/tbody><\/table>\n
ConversationBufferMemory\u662f\u4e00\u4e2a\u5f88\u597d\u7684\u5f00\u59cb\u9009\u9879\uff0c\u4f46\u53d7\u5230\u4fdd\u5b58\u6bcf\u6b21\u4e92\u52a8\u7684\u5b58\u50a8\u9650\u5236\u3002\u8ba9\u6211\u4eec\u770b\u770b\u5176\u4ed6\u53ef\u4ee5\u5e2e\u52a9\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u7684\u9009\u9879\u3002<\/p>\n
\u5bf9\u8bdd\u6458\u8981\u8bb0\u5fc6<\/h2>\n
\u4f7f\u7528ConversationBufferMemory\uff0c\u6211\u4eec\u5f88\u5feb\u5c31\u4f7f\u7528\u4e86\u5927\u91cf\u7684\u6807\u8bb0\uff0c\u751a\u81f3\u8d85\u51fa\u4e86\u5f53\u524d\u6700\u5148\u8fdb\u7684LLMs\u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\u9650\u5236\u3002<\/p>\n
\u4e3a\u4e86\u907f\u514d\u8fc7\u591a\u7684\u6807\u8bb0\u4f7f\u7528\uff0c\u6211\u4eec\u53ef\u4ee5\u4f7f\u7528ConversationSummaryMemory\u3002\u6b63\u5982\u5176\u540d\u79f0\u6240\u793a\uff0c\u8fd9\u79cd\u5f62\u5f0f\u7684\u8bb0\u5fc6\u5728\u4f20\u9012\u7ed9{history}\u53c2\u6570\u4e4b\u524d\u5bf9\u5bf9\u8bdd\u5386\u53f2\u8fdb\u884c\u4e86\u603b\u7ed3\u3002<\/p>\n
\u6211\u4eec\u53ef\u4ee5\u8fd9\u6837\u521d\u59cb\u5316ConversationChain\uff0c\u4f7f\u7528\u603b\u7ed3\u8bb0\u5fc6\uff1a<\/p>\n
<\/div>
ini<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>from langchain.chains.conversation.memory import ConversationSummaryMemory<\/span>\n<\/span>\nconversation<\/span> = ConversationChain(<\/span>\n\tllm<\/span>=llm,<\/span>\n\tmemory<\/span>=ConversationSummaryMemory(llm=llm)<\/span>\n)<\/span>\n<\/code><\/pre>\n
\u5f53\u4f7f\u7528ConversationSummaryMemory\u65f6\uff0c\u6211\u4eec\u9700\u8981\u5c06\u4e00\u4e2aLLM\u4f20\u9012\u7ed9\u8be5\u5bf9\u8c61\uff0c\u56e0\u4e3a\u603b\u7ed3\u662f\u7531\u4e00\u4e2aLLM\u63d0\u4f9b\u652f\u6301\u7684\u3002\u6211\u4eec\u53ef\u4ee5\u5728\u8fd9\u91cc\u770b\u5230\u7528\u4e8e\u6b64\u64cd\u4f5c\u7684\u63d0\u793a\uff1a<\/p>\n
In[19]:<\/p>\n
<\/div>
scss<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>print<\/span>(conversation_sum.memory.prompt.template)<\/span>\n<\/code><\/pre>\n
Out[19]:<\/p>\n
<\/div>
vbnet<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>Progressively summarize the lines of<\/span> conversation provided, adding onto the previous summary returning a new<\/span> summary.<\/span>\n<\/span>\nEXAMPLE<\/span>\nCurrent summary:<\/span>\nThe human asks what the AI thinks of<\/span> artificial intelligence. The AI thinks artificial intelligence is<\/span> a force for<\/span> good.<\/span>\n<\/span>\nNew<\/span> lines of<\/span> conversation:<\/span>\nHuman:<\/span> Why do<\/span> you think artificial intelligence is<\/span> a force for<\/span> good?<\/span>\nAI:<\/span> Because artificial intelligence will help humans reach their full potential.<\/span>\n<\/span>\nNew<\/span> summary:<\/span>\nThe human asks what the AI thinks of<\/span> artificial intelligence. The AI thinks artificial intelligence is<\/span> a force for<\/span> good because it will help humans reach their full potential.<\/span>\nEND<\/span> OF<\/span> EXAMPLE<\/span>\n<\/span>\nCurrent summary:<\/span>\n{summary}<\/span>\n<\/span>\nNew<\/span> lines of<\/span> conversation:<\/span>\n{new_lines}<\/span>\n<\/span>\nNew<\/span> summary:<\/span>\n<\/code><\/pre>\n
\u4f7f\u7528\u8fd9\u4e2a\u65b9\u6cd5\uff0c\u6211\u4eec\u53ef\u4ee5\u603b\u7ed3\u6bcf\u4e00\u6b21\u65b0\u7684\u4e92\u52a8\uff0c\u5e76\u5c06\u5b83\u6dfb\u52a0\u5230\u6240\u6709\u8fc7\u53bb\u4e92\u52a8\u7684\u201c\u8fd0\u884c\u6458\u8981\u201d\u4e2d\u3002\u8ba9\u6211\u4eec\u518d\u6b21\u8fdb\u884c\u4e00\u6b21\u5229\u7528\u8fd9\u79cd\u65b9\u6cd5\u7684\u5bf9\u8bdd\u3002<\/p>\n
In[40]:<\/p>\n
<\/div>
rust<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div># without count_tokens we'd<\/span> call `conversation_sum<\/span>(\"Good morning AI!\"<\/span>)`<\/span>\n# but let<\/span>'s<\/span> keep track of our tokens:<\/span>\ncount_tokens<\/span>(<\/span>\n    conversation_sum, <\/span>\n    \"Good morning AI!\"<\/span><\/span>\n)<\/span>\n<\/code><\/pre>\n
Out[40]:<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>Spent a<\/span> total of 290<\/span> tokens<\/span>\n<\/code><\/pre>\n
Out[40]:<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>\" Good morning! It's a<\/span> beautiful day today, isn't it? How can I<\/span> help you?\"<\/span>\n<\/code><\/pre>\n
In[41]:<\/p>\n
<\/div>
scss<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>count_tokens<\/span>(<\/span>\n    conversation_sum, <\/span>\n    \"My interest here is to explore the potential of integrating Large Language Models with external knowledge\"<\/span>\n)<\/span>\n<\/code><\/pre>\n
Out[41]:<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>Spent a<\/span> total of 440<\/span> tokens<\/span>\n<\/code><\/pre>\n
Out[41]:<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>\" That sounds like an interesting project! I<\/span>'m familiar with Large Language Models, but I<\/span>'m not sure how they could be integrated with external knowledge. Could you tell me more about what you have in mind?\"<\/span>\n<\/code><\/pre>\n
In[42]:<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>count_tokens(<\/span>\n    conversation_sum, <\/span>\n    \"I<\/span> just want to<\/span> analyze the different possibilities. What can you think of?\"<\/span>\n)<\/span>\n<\/code><\/pre>\n
Out[42]:<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>Spent a<\/span> total of 664<\/span> tokens<\/span>\n<\/code><\/pre>\n
Out[42]:<\/p>\n
<\/div>
vbnet<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>' I can think of a few possibilities. One option is to use a large language model to generate a set of candidate answers to a given query, and then use external knowledge to filter out the most relevant answers. Another option is to use the large language model to generate a set of candidate answers, and then use external knowledge to score and rank the answers. Finally, you could use the large language model to generate a set of candidate answers, and then use external knowledge to refine the answers.'<\/span><\/span>\n<\/code><\/pre>\n
In[43]:<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>count_tokens(<\/span>\n    conversation_sum, <\/span>\n    \"Which data source types could be used to<\/span> give context to<\/span> the model?\"<\/span>\n)<\/span>\n<\/code><\/pre>\n
Out[43]:<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>Spent a<\/span> total of 799<\/span> tokens<\/span>\n<\/code><\/pre>\n
Out[43]:<\/p>\n
<\/div>
vbnet<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>' There are many different types of data sources that could be used to give context to the model. These could include structured data sources such as databases, unstructured data sources such as text documents, or even external APIs that provide access to external knowledge. Additionally, the model could be trained on a combination of these data sources to provide a more comprehensive understanding of the context.'<\/span><\/span>\n<\/code><\/pre>\n
In[44]:<\/p>\n
<\/div>
scss<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>count_tokens<\/span>(<\/span>\n    conversation_sum, <\/span>\n    \"What is my aim again?\"<\/span>\n)<\/span>\n<\/code><\/pre>\n
Out[44]:<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>Spent a<\/span> total of 853<\/span> tokens<\/span>\n<\/code><\/pre>\n
Out[44]:<\/p>\n
<\/div>
vbnet<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>' Your aim is to explore the potential of integrating Large Language Models with external knowledge.'<\/span><\/span>\n<\/code><\/pre>\n
\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u603b\u7ed3\u5305\u542b\u8db3\u591f\u7684\u4fe1\u606f\uff0c\u4ee5\u4f7fLLM\u201c\u8bb0\u4f4f\u201d\u6211\u4eec\u6700\u521d\u7684\u76ee\u6807\u3002\u6211\u4eec\u53ef\u4ee5\u50cf\u8fd9\u6837\u67e5\u770b\u603b\u7ed3\u7684\u539f\u59cb\u5f62\u5f0f\uff1a<\/p>\n
In[45]:<\/p>\n
<\/div>
scss<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>print<\/span>(conversation_sum.memory.buffer)<\/span>\n<\/code><\/pre>\n
Out[45]:<\/p>\n
<\/div>
sql<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div><\/span>\nThe human greeted the AI with<\/span> a good morning, to<\/span> which the AI responded with<\/span> a good morning and<\/span> asked how it could help. The human expressed interest in<\/span> exploring the potential of<\/span> integrating Large<\/span> Language<\/span> Models with<\/span> external<\/span> knowledge, to<\/span> which the AI responded positively and<\/span> asked for<\/span> more information. The human asked the AI to<\/span> think of<\/span> different possibilities, and<\/span> the AI suggested three options: using<\/span> the large<\/span> language<\/span> model to<\/span> generate a set<\/span> of<\/span> candidate answers and<\/span> then<\/span> using<\/span> external<\/span> knowledge to<\/span> filter<\/span> out<\/span> the most relevant answers, score and<\/span> rank the answers, or<\/span> refine the answers. The human then<\/span> asked which data source types could be used to<\/span> give context to<\/span> the model, to<\/span> which the AI responded that there are<\/span> many different types of<\/span> data sources that could be used, such as<\/span> structured data sources, unstructured data sources, or<\/span> external<\/span> APIs. Additionally, the model could be trained on<\/span> a combination of<\/span> these data sources to<\/span> provide a more comprehensive understanding of<\/span> the context. The human then<\/span> asked what their aim was again, to<\/span> which the AI responded that their aim was to<\/span> explore the potential of<\/span> integrating Large<\/span> Language<\/span> Models with<\/span> external<\/span> knowledge.<\/span>\n<\/code><\/pre>\n
\u8fd9\u6b21\u5bf9\u8bdd\u4e2d\u4f7f\u7528\u7684\u6807\u8bb0\u6570\u91cf\u6bd4\u4f7f\u7528ConversationBufferMemory\u591a\uff0c\u6240\u4ee5\u4f7f\u7528ConversationSummaryMemory\u662f\u5426\u6709\u4efb\u4f55\u4f18\u52bf\u5462\uff1f<\/p>\n
\n
\"image.png\"<\/figure>\n<\/p>\n
\u5bf9\u4e8e\u66f4\u957f\u7684\u5bf9\u8bdd\u6765\u8bf4\uff0c\u662f\u7684\u3002\u5728\u8fd9\u91cc\uff0c\u6211\u4eec\u6709\u4e00\u4e2a\u8f83\u957f\u7684\u5bf9\u8bdd\u3002\u5982\u4e0a\u6240\u793a\uff0c\u6458\u8981\u5185\u5b58\u6700\u521d\u4f7f\u7528\u7684\u6807\u8bb0\u8981\u591a\u5f97\u591a\u3002\u7136\u800c\uff0c\u968f\u7740\u5bf9\u8bdd\u7684\u8fdb\u884c\uff0c\u603b\u7ed3\u7684\u65b9\u6cd5\u589e\u957f\u901f\u5ea6\u8f83\u6162\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u7f13\u51b2\u5185\u5b58\u4f1a\u968f\u7740\u804a\u5929\u4e2d\u7684\u6807\u8bb0\u6570\u91cf\u7ebf\u6027\u589e\u957f\u3002<\/p>\n
\u6211\u4eec\u53ef\u4ee5\u603b\u7ed3ConversationSummaryMemory\u7684\u4f18\u7f3a\u70b9\u5982\u4e0b\uff1a<\/p>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Pros<\/th>Cons<\/th><\/tr><\/thead>
Shortens the number of tokens for long conversations.<\/td>Can result in higher token usage for smaller conversations<\/td><\/tr>
Enables much longer conversations<\/td>Memorization of the conversation history is wholly reliant on the summarization ability of the intermediate summarization LLM<\/td><\/tr>
Relatively straightforward implementation, intuitively simple to understand<\/td>Also requires token usage for the summarization LLM; this increases costs (but does not limit conversation length)<\/td><\/tr><\/tbody><\/table>\n
\u5bf9\u4e8e\u9884\u671f\u6709\u957f\u5bf9\u8bdd\u7684\u60c5\u51b5\uff0c\u5bf9\u8bdd\u603b\u7ed3\u662f\u4e00\u4e2a\u4e0d\u9519\u7684\u65b9\u6cd5\u3002\u7136\u800c\uff0c\u5b83\u4ecd\u7136\u57fa\u672c\u4e0a\u53d7\u5230\u6807\u8bb0\u9650\u5236\u7684\u9650\u5236\u3002\u7ecf\u8fc7\u4e00\u6bb5\u65f6\u95f4\u540e\uff0c\u6211\u4eec\u4ecd\u7136\u4f1a\u8d85\u51fa\u4e0a\u4e0b\u6587\u7a97\u53e3\u7684\u9650\u5236\u3002<\/p>\n
\u5bf9\u8bdd\u7f13\u51b2\u7a97\u53e3\u5185\u5b58<\/h2>\n
“ConversationBufferWindowMemory” \u7684\u4f5c\u7528\u4e0e\u4e4b\u524d\u7684 “\u7f13\u51b2\u5185\u5b58” \u76f8\u540c\uff0c\u4f46\u5b83\u5728\u5185\u5b58\u4e2d\u6dfb\u52a0\u4e86\u4e00\u4e2a\u7a97\u53e3\u3002\u8fd9\u610f\u5473\u7740\u6211\u4eec\u53ea\u4fdd\u7559\u4e00\u5b9a\u6570\u91cf\u7684\u8fc7\u53bb\u4ea4\u4e92\uff0c\u7136\u540e\u5c06\u5176\u201c\u9057\u5fd8\u201d\u3002\u6211\u4eec\u4f7f\u7528\u5b83\u7684\u65b9\u5f0f\u5982\u4e0b\uff1a<\/p>\n
<\/div>
ini<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>from langchain.chains.conversation.memory import ConversationBufferWindowMemory<\/span>\n<\/span>\nconversation<\/span> = ConversationChain(<\/span>\n\tllm<\/span>=llm,<\/span>\n\tmemory<\/span>=ConversationBufferWindowMemory(k=1<\/span>)<\/span>\n)<\/span>\n<\/code><\/pre>\n
\u5728\u8fd9\u4e2a\u793a\u4f8b\u4e2d\uff0c\u6211\u4eec\u8bbe\u7f6e\u4e86 k=1 \u2014 \u8fd9\u610f\u5473\u7740\u7a97\u53e3\u4f1a\u8bb0\u4f4f\u4eba\u7c7b\u548c\u4eba\u5de5\u667a\u80fd\u4e4b\u95f4\u7684\u6700\u65b0\u4e00\u6b21\u4ea4\u4e92\u3002\u4e5f\u5c31\u662f\u6700\u65b0\u7684\u4eba\u7c7b\u56de\u590d\u548c\u6700\u65b0\u7684\u4eba\u5de5\u667a\u80fd\u56de\u590d\u3002\u6211\u4eec\u53ef\u4ee5\u770b\u5230\u8fd9\u4e2a\u6548\u679c\u5982\u4e0b\uff1a<\/p>\n
In[61]:<\/p>\n
<\/div>
scss<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>count_tokens<\/span>(<\/span>\n    conversation_bufw, <\/span>\n    \"Good morning AI!\"<\/span>\n)<\/span>\n<\/code><\/pre>\n
Out[61]:<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>Spent a<\/span> total of 85<\/span> tokens<\/span>\n<\/code><\/pre>\n
Out[61]:<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>\" Good morning! It's a<\/span> beautiful day today, isn't it? How can I<\/span> help you?\"<\/span>\n<\/code><\/pre>\n
In[62]:<\/p>\n
<\/div>
scss<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>count_tokens<\/span>(<\/span>\n    conversation_bufw, <\/span>\n    \"My interest here is to explore the potential of integrating Large Language Models with external knowledge\"<\/span>\n)<\/span>\n<\/code><\/pre>\n
Out[62]:<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>Spent a<\/span> total of 178<\/span> tokens<\/span>\n<\/code><\/pre>\n
Out[62]:<\/p>\n
<\/div>
vbnet<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>' Interesting! Large Language Models are a type of artificial intelligence that can process natural language and generate text. They can be used to generate text from a given context, or to answer questions about a given context. Integrating them with external knowledge can help them to better understand the context and generate more accurate results. Do you have any specific questions about this integration?'<\/span><\/span>\n<\/code><\/pre>\n
In[63]:<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>count_tokens(<\/span>\n    conversation_bufw, <\/span>\n    \"I<\/span> just want to<\/span> analyze the different possibilities. What can you think of?\"<\/span>\n)<\/span>\n<\/code><\/pre>\n
Out[63]:<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>Spent a<\/span> total of 233<\/span> tokens<\/span>\n<\/code><\/pre>\n
Out[63]:<\/p>\n
<\/div>
vbnet<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>' There are many possibilities for integrating Large Language Models with external knowledge. For example, you could use external knowledge to provide additional context to the model, or to provide additional training data. You could also use external knowledge to help the model better understand the context of a given text, or to help it generate more accurate results.'<\/span><\/span>\n<\/code><\/pre>\n
In[64]:<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>count_tokens(<\/span>\n    conversation_bufw, <\/span>\n    \"Which data source types could be used to<\/span> give context to<\/span> the model?\"<\/span>\n)<\/span>\n<\/code><\/pre>\n
Out[64]:<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>Spent a<\/span> total of 245<\/span> tokens<\/span>\n<\/code><\/pre>\n
Out[64]:<\/p>\n
<\/div>
vbnet<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>' Data sources that could be used to give context to the model include text corpora, structured databases, and ontologies. Text corpora provide a large amount of text data that can be used to train the model and provide additional context. Structured databases provide structured data that can be used to provide additional context to the model. Ontologies provide a structured representation of knowledge that can be used to provide additional context to the model.'<\/span><\/span>\n<\/code><\/pre>\n
In[65]:<\/p>\n
<\/div>
scss<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>count_tokens<\/span>(<\/span>\n    conversation_bufw, <\/span>\n    \"What is my aim again?\"<\/span>\n)<\/span>\n<\/code><\/pre>\n
Out[65]:<\/p>\n
<\/div>
css<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>Spent a<\/span> total of 186<\/span> tokens<\/span>\n<\/code><\/pre>\n
Out[65]:<\/p>\n
<\/div>
vbnet<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>' Your aim is to use data sources to give context to the model.'<\/span><\/span>\n<\/code><\/pre>\n
\u5728\u5bf9\u8bdd\u7ed3\u675f\u65f6\uff0c\u5f53\u6211\u4eec\u95ee\u201c\u6211\u7684\u76ee\u6807\u662f\u4ec0\u4e48\uff1f\u201d\u65f6\uff0c\u7b54\u6848\u5305\u542b\u5728\u4e09\u6b21\u4ea4\u4e92\u524d\u7684\u4eba\u7c7b\u56de\u5e94\u4e2d\u3002\u7531\u4e8e\u6211\u4eec\u53ea\u4fdd\u7559\u4e86\u6700\u8fd1\u7684\u4e00\u6b21\u4ea4\u4e92\uff08k=1\uff09\uff0c\u6a21\u578b\u5df2\u7ecf\u5fd8\u8bb0\u4e86\uff0c\u65e0\u6cd5\u7ed9\u51fa\u6b63\u786e\u7684\u7b54\u6848\u3002<\/p>\n
\u6211\u4eec\u53ef\u4ee5\u8fd9\u6837\u770b\u5230\u6a21\u578b\u7684\u6709\u6548\u201c\u8bb0\u5fc6\u201d\uff1a<\/p>\n
In[66]:<\/p>\n
<\/div>
ini<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>bufw_history<\/span> = conversation_bufw.memory.load_memory_variables(<\/span>\n    inputs<\/span>=[]<\/span>\n)['history']<\/span><\/span>\n<\/code><\/pre>\n
In[67]:<\/p>\n
<\/div>
scss<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>print<\/span>(bufw_history)<\/span>\n<\/code><\/pre>\n
Out[67]:<\/p>\n
<\/div>
vbnet<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>Human:<\/span> What is<\/span> my aim again?<\/span>\nAI:<\/span>  Your aim is<\/span> to<\/span> use data sources to<\/span> give context to<\/span> the model.<\/span>\n<\/code><\/pre>\n
\u5c3d\u7ba1\u8fd9\u79cd\u65b9\u6cd5\u4e0d\u9002\u7528\u4e8e\u8bb0\u4f4f\u8fdc\u8ddd\u79bb\u7684\u4ea4\u4e92\uff0c\u4f46\u5b83\u5f88\u64c5\u957f\u9650\u5236\u4f7f\u7528\u7684\u6807\u8bb0\u6570\u91cf \u2014 \u8fd9\u662f\u4e00\u4e2a\u53ef\u4ee5\u6839\u636e\u6211\u4eec\u7684\u9700\u6c42\u589e\u52a0\u6216\u51cf\u5c11\u7684\u6570\u91cf\u3002\u5bf9\u4e8e\u6211\u4eec\u4e4b\u524d\u6bd4\u8f83\u4e2d\u4f7f\u7528\u7684\u66f4\u957f\u5bf9\u8bdd\uff0c\u6211\u4eec\u53ef\u4ee5\u8bbe\u7f6e k=6\uff0c\u5728\u603b\u517127\u6b21\u4ea4\u4e92\u540e\u8fbe\u5230\u6bcf\u6b21\u4ea4\u4e92\u7ea61.5K\u6807\u8bb0\uff1a<\/p>\n
\n
\"image.png\"<\/figure>\n<\/p>\n
\u5982\u679c\u6211\u4eec\u53ea\u9700\u8981\u6700\u8fd1\u4ea4\u4e92\u7684\u8bb0\u5fc6\uff0c\u8fd9\u662f\u4e00\u4e2a\u5f88\u597d\u7684\u9009\u62e9\u3002\u7136\u800c\uff0c\u5982\u679c\u9700\u8981\u540c\u65f6\u8003\u8651\u8fdc\u8ddd\u79bb\u548c\u6700\u8fd1\u4ea4\u4e92\uff0c\u8fd8\u6709\u5176\u4ed6\u9009\u62e9\u3002<\/p>\n
\u5bf9\u8bdd\u603b\u7ed3\u7f13\u51b2\u5185\u5b58<\/h2>\n
“ConversationSummaryBufferMemory” \u662f\u5bf9\u8bdd\u603b\u7ed3\u5185\u5b58\u548c\u5bf9\u8bdd\u7f13\u51b2\u7a97\u53e3\u5185\u5b58\u7684\u6df7\u5408\u4f53\u3002\u5b83\u603b\u7ed3\u4e86\u5bf9\u8bdd\u4e2d\u6700\u65e9\u7684\u4ea4\u4e92\uff0c\u540c\u65f6\u4fdd\u7559\u4e86\u5bf9\u8bdd\u4e2d\u6700\u8fd1\u7684\u6807\u8bb0\u6570\u4e0d\u8d85\u8fc7\u6700\u5927\u6807\u8bb0\u9650\u5236\u3002\u521d\u59cb\u5316\u5982\u4e0b\u6240\u793a\uff1a<\/p>\n
<\/div>
ini<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>conversation_sum_bufw<\/span> = ConversationChain(<\/span>\n    llm<\/span>=llm, memory=ConversationSummaryBufferMemory(<\/span>\n        llm<\/span>=llm,<\/span>\n        max_token_limit<\/span>=650<\/span><\/span>\n)<\/span>\n<\/code><\/pre>\n
\u5c06\u8fd9\u4e2a\u65b9\u6cd5\u5e94\u7528\u5230\u4e4b\u524d\u7684\u5bf9\u8bdd\u4e2d\uff0c\u6211\u4eec\u53ef\u4ee5\u5c06max_token_limit\u8bbe\u7f6e\u4e3a\u4e00\u4e2a\u8f83\u5c0f\u7684\u6570\u5b57\uff0c\u4f46\u8bed\u8a00\u6a21\u578b\u4ecd\u7136\u53ef\u4ee5\u8bb0\u4f4f\u6211\u4eec\u4e4b\u524d\u7684\u201c\u76ee\u6807\u201d\u3002<\/p>\n
\u8fd9\u662f\u56e0\u4e3a\u8fd9\u4e9b\u4fe1\u606f\u88ab\u201c\u603b\u7ed3\u201d\u7ec4\u4ef6\u6355\u83b7\uff0c\u5c3d\u7ba1\u88ab\u201c\u7f13\u51b2\u7a97\u53e3\u201d\u7ec4\u4ef6\u5ffd\u7565\u4e86\u3002<\/p>\n
\u5f53\u7136\uff0c\u8fd9\u4e2a\u7ec4\u4ef6\u7684\u4f18\u7f3a\u70b9\u662f\u57fa\u4e8e\u5176\u6240\u4f9d\u8d56\u7684\u4e4b\u524d\u7684\u7ec4\u4ef6\u7684\u6df7\u5408\u3002<\/p>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Pros<\/th>Cons<\/th><\/tr><\/thead>
Summarizer means we can remember distant interactions<\/td>Summarizer increases token count for shorter conversations<\/td><\/tr>
Buffer prevents us from missing information from the most recent interactions<\/td>Storing the raw interactions \u2014 even if just the most recent interactions \u2014 increases token count<\/td><\/tr><\/tbody><\/table>\n
\u5c3d\u7ba1\u9700\u8981\u66f4\u591a\u7684\u8c03\u6574\u6765\u786e\u5b9a\u8981\u603b\u7ed3\u548c\u5728\u7f13\u51b2\u7a97\u53e3\u5185\u4fdd\u7559\u7684\u5185\u5bb9\uff0c\u4f46ConversationSummaryBufferMemory\u786e\u5b9e\u4e3a\u6211\u4eec\u63d0\u4f9b\u4e86\u5f88\u5927\u7684\u7075\u6d3b\u6027\uff0c\u800c\u4e14\u5b83\u662f\u6211\u4eec\u76ee\u524d\u552f\u4e00\u80fd\u591f\u8bb0\u4f4f\u8fdc\u8ddd\u79bb\u4ea4\u4e92\u5e76\u4ee5\u539f\u59cb\u548c\u6700\u4fe1\u606f\u4e30\u5bcc\u7684\u5f62\u5f0f\u5b58\u50a8\u6700\u8fd1\u4ea4\u4e92\u7684\u8bb0\u5fc6\u7c7b\u578b\u4e4b\u4e00\u3002<\/p>\n
\n
\"image.png\"<\/figure>\n<\/p>\n
\u6211\u4eec\u8fd8\u53ef\u4ee5\u770b\u5230\uff0c\u5c3d\u7ba1\u5305\u62ec\u4e86\u5bf9\u8fc7\u53bb\u4ea4\u4e92\u7684\u603b\u7ed3\u548c\u6700\u8fd1\u4ea4\u4e92\u7684\u539f\u59cb\u5f62\u5f0f\uff0cConversationSummaryBufferMemory\u7684\u6807\u8bb0\u6570\u91cf\u589e\u52a0\u4e0e\u5176\u4ed6\u65b9\u6cd5\u76f8\u7ade\u4e89\u3002<\/p>\n
\u5176\u4ed6\u8bb0\u5fc6\u7c7b\u578b<\/h2>\n
\u6211\u4eec\u5728\u8fd9\u91cc\u4ecb\u7ecd\u7684\u8bb0\u5fc6\u7c7b\u578b\u975e\u5e38\u9002\u5408\u5165\u95e8\uff0c\u80fd\u591f\u5728\u5c3d\u91cf\u8bb0\u4f4f\u5c3d\u53ef\u80fd\u591a\u7684\u4fe1\u606f\u548c\u6700\u5c0f\u5316\u6807\u8bb0\u4e4b\u95f4\u53d6\u5f97\u826f\u597d\u7684\u5e73\u8861\u3002<\/p>\n
\u7136\u800c\uff0c\u6211\u4eec\u8fd8\u6709\u5176\u4ed6\u9009\u62e9\uff0c\u7279\u522b\u662fConversationKnowledgeGraphMemory\u548cConversationEntityMemory\u3002\u6211\u4eec\u5c06\u5728\u63a5\u4e0b\u6765\u7684\u7ae0\u8282\u4e2d\u66f4\u8be6\u7ec6\u5730\u4ecb\u7ecd\u8fd9\u4e9b\u4e0d\u540c\u5f62\u5f0f\u7684\u8bb0\u5fc6\u3002<\/p>\n
\u8fd9\u5c31\u662f\u5173\u4e8e\u4f7f\u7528LangChain\u7684LLMs\u7684\u5bf9\u8bdd\u8bb0\u5fc6\u7684\u7b80\u4ecb\u3002\u6b63\u5982\u6211\u4eec\u6240\u89c1\uff0c\u6709\u5f88\u591a\u9009\u9879\u53ef\u4ee5\u5e2e\u52a9\u65e0\u72b6\u6001\u7684LLMs\u4ee5\u4eff\u4f5b\u5728\u6709\u72b6\u6001\u7684\u73af\u5883\u4e2d\u8fdb\u884c\u4ea4\u4e92\uff0c\u80fd\u591f\u8003\u8651\u548c\u53c2\u8003\u4ee5\u524d\u7684\u4ea4\u4e92\u3002<\/p>\n
\u6b63\u5982\u63d0\u5230\u7684\uff0c\u8fd8\u6709\u5176\u4ed6\u5f62\u5f0f\u7684\u8bb0\u5fc6\u6211\u4eec\u53ef\u4ee5\u6db5\u76d6\u3002\u6211\u4eec\u8fd8\u53ef\u4ee5\u5b9e\u73b0\u81ea\u5df1\u7684\u8bb0\u5fc6\u6a21\u5757\uff0c\u5728\u540c\u4e00\u94fe\u4e2d\u4f7f\u7528\u591a\u79cd\u7c7b\u578b\u7684\u8bb0\u5fc6\uff0c\u5c06\u5b83\u4eec\u4e0e\u4ee3\u7406\u7ed3\u5408\u4f7f\u7528\uff0c\u7b49\u7b49\u3002\u6240\u6709\u8fd9\u4e9b\u5185\u5bb9\u90fd\u5c06\u5728\u672a\u6765\u7684\u7ae0\u8282\u4e2d\u4ecb\u7ecd\u3002<\/p>","protected":false},"excerpt":{"rendered":"
\u5bf9\u8bdd\u8bb0\u5fc6\u662f\u6307\u804a\u5929\u673a\u5668\u4eba\u5982\u4f55\u4ee5\u5bf9\u8bdd\u65b9\u5f0f\u54cd\u5e94\u591a\u4e2a\u67e5\u8be2\u3002\u5b83\u4f7f\u5bf9\u8bdd\u8fde\u8d2f\uff0c\u5e76\u4e14\u5982\u679c\u6ca1\u6709\u5b83\uff0c\u6bcf\u4e2a\u67e5\u8be2\u90fd\u5c06\u88ab\u89c6\u4e3a\u5b8c\u5168\u72ec\u7acb\u7684\u8f93\u5165\uff0c\u800c\u4e0d\u8003\u8651\u8fc7\u53bb\u7684\u4ea4\u4e92\u3002 \u8fd9\u79cd\u8bb0\u5fc6\u4f7f\u5f97\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u80fd\u591f\u8bb0\u4f4f\u4e0e\u7528\u6237\u4e4b\u524d\u7684\u4e92\u52a8\u3002\u9ed8\u8ba4<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"rank_math_title":"","rank_math_description":"","rank_math_focus_keyword":"","views":"","footnotes":""},"categories":[3],"tags":[126,127,128,129,136],"collection":[],"class_list":["post-1358","post","type-post","status-publish","format-standard","hentry","category-fenlei2","tag-gpt","tag-ai","tag-128","tag-129","tag-136"],"acf":[],"_links":{"self":[{"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/posts\/1358","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/comments?post=1358"}],"version-history":[{"count":0,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/posts\/1358\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/media?parent=1358"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/categories?post=1358"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/tags?post=1358"},{"taxonomy":"collection","embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/collection?post=1358"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}