【AI】【Agent】来实践

本次就是聚焦到Dify（开源版）进行实践（上次是宏观体验），包含了内容抓取，知识库，绘图和工作流的外部调用，一起来Get!

看效果

如上是对一批网站按照特定分类后做的可视化展示！图像是通过Dify生成的！

话原理

本次的实践流程上，是拿到一批网站，请求网站的既定分类（调用dify的网站分类工作流），再处理汇总数据，调用可视化生成工具（dify的可视化呈现工作流）

其中涉及到几个工具的使用，做个小结

1）知识库

我上传了既定的分类文件，我的上传就包含了2列，ID和Category

经过试了一番下来，最后还是都用了默认的！

创建知识库

这里有三种检索方式，分别是向量检索，全文检索和混合检索！

向量检索里Top k是说在检索时找到k个最相似的文本块，score阈值是相似度来过滤文本块的；Rerank是做文本块的排序，dify种有好几个比如cohere的；

文本检索就是全文字符关键词检索这种方式去找，，这种是弥补向量查询的劣势：比如查询很短的文本，不常用的文本或专业名次，都是向量数据库不擅长的；所以是推荐混合检索

使用时的检索方式

这里可以理解的是，第一个种，就是向量检索后直接找了一个文本块；第二种是会返回多个文本块，结合rerank给到下一步去分析；逻辑上来讲，第二个更加不容易漏掉信息

个人实际体验：

混合检索，检索出来的不准，最后切回了向量检索

Multipath的也没有达到找出更多准确信息，最后切换回了N-to-1

我一开始是上传的txt文件，后面发现大模型在理解字段时不好去分，最后更改上传为excel

PS：这个只是一个项目的体验，只做参考，或者不同的文本效果会不一样

2）内容抓取

我尝试使用了jinareader、Web Scraper，对比下来jinareader和Web Scraper的抓取有的信息够，有的信息抓了很多不相关的

个人实际体验：最后自己使用了GoogleSearch,效果相对好

3）绘图

这块尝试了line、bar和pie，本地项目使用的是pie 饼图

是因为line和bar的输入这块代码判断存在bug，算bug吗，我去看了源码是调用的python的源码，我觉得算

以line为列，下面是一个要求的输入

输入data: 2,3,4; 20,18,32 #组间以英文分号隔开，组内是英文逗号

输入axis: tag1;tag2;tag3; #以英文分号隔开

这里直接把如上的2,3,4去做浮点化了，就报错了；应该先对组内英文逗号隔开处理！

个人体验建议：

建议dify官方优化下修复下/另外显示上能优化下界面呈现等，现在就朴素的没有使用的冲动哈哈哈

4）外部调用api

这块做的很方便，每次发布完可以直接调；log也很方便，可以协助排查问题，看trace

来实践

1、Workflow1 DSL给网站做分类

app:description:websiteCaticon:"\U0001F916"icon_background:'#FFEAD5'mode:workflowname:websiteCatworkflow:features:file_upload:image:enabled:falsenumber_limits:3transfer_methods:-local_file-remote_urlopening_statement:''retriever_resource:enabled:falsesensitive_word_avoidance:enabled:falsespeech_to_text:enabled:falsesuggested_questions:[]suggested_questions_after_answer:enabled:falsetext_to_speech:enabled:falselanguage:''voice:''graph:edges:-data:sourceType:starttargetType:toolid:1717578663813-1717579685305source:'1717578663813'sourceHandle:sourcetarget:'1717579685305'targetHandle:targettype:custom-data:sourceType:tooltargetType:llmid:1717579685305-1717579853399source:'1717579685305'sourceHandle:sourcetarget:'1717579853399'targetHandle:targettype:custom-data:sourceType:llmtargetType:endid:1717579967241-1717579117964source:'1717579967241'sourceHandle:sourcetarget:'1717579117964'targetHandle:targettype:custom-data:sourceType:llmtargetType:knowledge-retrievalid:1717579853399-1717580066887source:'1717579853399'sourceHandle:sourcetarget:'1717580066887'targetHandle:targettype:custom-data:sourceType:knowledge-retrievaltargetType:llmid:1717580066887-1717579967241source:'1717580066887'sourceHandle:sourcetarget:'1717579967241'targetHandle:targettype:customnodes:-data:desc:''selected:falsetitle:Starttype:startvariables:-label:urlsmax_length:256options:[]required:truetype:text-inputvariable:urlsheight:90id:'1717578663813'position:x:112.4877286321854y:197.44333138123648positionAbsolute:x:112.4877286321854y:197.44333138123648selected:falsesourcePosition:righttargetPosition:lefttype:customwidth:244-data:desc:''outputs:-value_selector:-'1717579967241'-textvariable:textselected:falsetitle:Endtype:endheight:90id:'1717579117964'position:x:799.2832195913106y:-74.46777660074163positionAbsolute:x:799.2832195913106y:-74.46777660074163selected:falsesourcePosition:righttargetPosition:lefttype:customwidth:244-data:desc:"\u6293\u53D6\u7F51\u7AD9\u4FE1\u606F"provider_id:googleprovider_name:googleprovider_type:builtinselected:truetitle:GoogleSearchtool_configurations:result_type:linktool_label:GoogleSearchtool_name:google_searchtool_parameters:query:type:mixedvalue:'{{#1717578663813.urls#}}'type:toolheight:120id:'1717579685305'position:x:198.59871386335033y:-74.46777660074163positionAbsolute:x:198.59871386335033y:-74.46777660074163selected:truesourcePosition:righttargetPosition:lefttype:customwidth:244-data:context:enabled:truevariable_selector:-'1717579685305'-textdesc:"\u603B\u7ED3\u7F51\u7AD9\u5173\u952E\u8BCD"model:completion_params:temperature:0mode:chatname:moonshot-v1-8kprovider:moonshotprompt_template:-id:4aa3aab4-11f5-41dc-a382-dd9831658f78role:systemtext:"leaseextractthecorrespondingtitleordescriptionkeywordsbased\\onthecontentof{{#1717579685305.text#}}\n\n1\uFF09Note.Ifthiswebsite\\sellsmanydifferentcategoriesofthings,itcanbeclassifiedasshopping\n\2)Extractnomorethan3keywords,andtheoutputformatis:\nxx,xx,\\xx\n"-id:dcd22bf7-6d91-4e6e-a843-ed266659d8d7role:usertext:/selected:falsetitleLMtype:llmvariables:[]vision:enabled:falseheight:128id:'1717579853399'position:x:463.64815768773747y:197.44333138123648positionAbsolute:x:463.64815768773747y:197.44333138123648selected:falsesourcePosition:righttargetPosition:lefttype:customwidth:244-data:context:enabled:truevariable_selector:-'1717580066887'-resultdesc:"\u6DA6\u8272\u7ED9\u51FA\u6700\u7EC8\u7684\u7F51\u7AD9\u6807\u7B7E"model:completion_params:temperature:0mode:chatname:moonshot-v1-8kprovider:moonshotprompt_template:-id:0bb1c5a6-930e-4f74-be63-06474a77fe48role:systemtext:'leasecombinethewebsite''skeywords{{#1717579853399.text#}}withtheclassificationlabels/proofreadingretrievedfromtheknowledgebasetoprovidethemostsuitableclassificationlabelsretrievedfromtheknowledgebase{{#context#}},Attention:1)Thistagshouldcomefromatagretrievedfromtheknowledgebase.Pleaseselectonlythetagthatbestexpressesthiswebsiteandoutputonlyonecategorytag2)Ifyoudon''tknowwhichcategoryitis,justoutputUncategorized3)Outputonlyneedstooutputthewebsitecategorywithoutdetailedexplanation,withoutIDoutputformat:xx'selected:falsetitleLM2type:llmvariables:[]vision:enabled:falseheight:128id:'1717579967241'position:x:799.2832195913106y:197.44333138123648positionAbsolute:x:799.2832195913106y:197.44333138123648selected:falsesourcePosition:righttargetPosition:lefttype:customwidth:244-data:dataset_ids:-2307d992-5284-4c0b-b270-8436f497bfeadesc:"\u68C0\u7D22\u5DF2\u6709\u7684\u7F51\u7AD9\u5206\u7C7B\u6807\u7B7E"multiple_retrieval_config:reranking_model:model:rerank-english-v2.0provider:coherescore_threshold:0.8top_k:5query_variable_selector:-'1717579853399'-textretrieval_mode:singleselected:falsesingle_retrieval_config:model:completion_params:{}mode:chatname:moonshot-v1-8kprovider:moonshottitle:KnowledgeRetrievaltype:knowledge-retrievalheight:122id:'1717580066887'position:x:503.45190546262575y:-74.46777660074163positionAbsolute:x:503.45190546262575y:-74.46777660074163selected:falsesourcePosition:righttargetPosition:lefttype:customwidth:244viewport:x:40.67810562789691y:254.34055609621913zoom:0.6607535491528895

2、workflow2统计分类展示

app:description:Batch&Visusalizationicon:"\U0001F916"icon_background:'#FFEAD5'mode:workflowname:Batch&Visusalizationworkflow:features:file_upload:image:enabled:falsenumber_limits:3transfer_methods:-local_file-remote_urlopening_statement:''retriever_resource:enabled:falsesensitive_word_avoidance:enabled:falsespeech_to_text:enabled:falsesuggested_questions:[]suggested_questions_after_answer:enabled:falsetext_to_speech:enabled:falselanguage:''voice:''graph:edges:-data:sourceType:starttargetType:toolid:1717587078340-1717596240088source:'1717587078340'sourceHandle:sourcetarget:'1717596240088'targetHandle:targettype:custom-data:sourceType:tooltargetType:endid:1717596240088-1717587757492source:'1717596240088'sourceHandle:sourcetarget:'1717587757492'targetHandle:targettype:customnodes:-data:desc:''selected:falsetitle:Starttype:startvariables:-label:tagsmax_length:33024options:[]required:truetype:paragraphvariable:tags-label:countmax_length:256options:[]required:truetype:text-inputvariable:countheight:116id:'1717587078340'position:x:92.06703286554381y:80positionAbsolute:x:92.06703286554381y:80selected:falsesourcePosition:righttargetPosition:lefttype:customwidth:244-data:desc:''provider_id:chartprovider_name:chartprovider_type:builtinselected:falsetitleieCharttool_configurations:{}tool_labelieCharttool_name:pie_charttool_parameters:categories:type:mixedvalue:'{{#1717587078340.tags#}}'data:type:mixedvalue:'{{#1717587078340.count#}}'type:toolheight:54id:'1717596240088'position:x:413.7520931625256y:162.0976621428633positionAbsolute:x:413.7520931625256y:162.0976621428633selected:falsesourcePosition:righttargetPosition:lefttype:customwidth:244-data:desc:''outputs:-value_selector:-'1717596240088'-filesvariable:textselected:falsetitle:Endtype:endheight:90id:'1717587757492'position:x:663.9837808989655y:169.00026104872103positionAbsolute:x:663.9837808989655y:169.00026104872103selected:falsesourcePosition:righttargetPosition:lefttype:customwidth:244viewport:x:-76.8851566982803y:181.9020021881131zoom:1.0705699835378846

3、本地的python代码

importrequestsimportcsvfromcollectionsimportCounterimportwebbrowserimporttime#Step1:Readsite.csvfilesites=[]withopen('sites.csv',mode='r',encoding='utf-8')asfile:reader=csv.DictReader(file)forrowinreader:sites.append(row['site'])#Step2:Requesttogetthecategoryforeachsiteheaders={"Authorization":"Bearerxxxxxxxx","Content-Type":"application/json"}categories=[]forsiteinsites:body={"inputs":{"urls":site},"response_mode":"blocking","user":"abc-123"}response=requests.post("http://localhost/v1/workflows/run",headers=headers,json=body)ifresponse.status_code==200:response_data=response.json()ifresponse_data['data']andresponse_data['data']['outputs']:category=response_data['data']['outputs']['text']categories.append(category)else:print(f"No'outputs'inresponseforsite:{site}")print(response_data)else:print(f"Failedtogetcategoryforsite:{site}")print(response.text)#time.sleep(5)#Waitfor5secondsbeforethenextrequest#Step3:Countthecategoriesandsortthemcategory_counts=Counter(categories)sorted_categories=sorted(category_counts.items(),key=lambdax:x[1],reverse=True)#Formtherequiredstringssorted_sites=";".join(siteforsite,countinsorted_categories)sorted_counts=";".join(str(count)forsite,countinsorted_categories)#Step4:CallthevisualizationAPIheaders_vis={"Authorization":"Bearerxxxxxxxxxxxx","Content-Type":"application/json"}body_vis={"inputs":{"tags":sorted_sites,"count":sorted_counts},"response_mode":"blocking","user":"abc-123"}response_vis=requests.post("http://localhost/v1/workflows/run",headers=headers_vis,json=body_vis)ifresponse_vis.status_code==200:response_data_vis=response_vis.json()ifresponse_data_vis['data']andresponse_data_vis['data']['outputs']:image_url=response_data_vis['data']['outputs']['text'][0]['url']print(f"VisualizationURL:{image_url}")webbrowser.open(image_url)else:print("No'outputs'invisualizationresponse.")print(response_data_vis)else:print("Failedtogeneratevisualization.")print(response_vis.text)

写在最后

本次实践，Googlesearch的免费API额度100次request过程；过程中发现了dify的line好难用，最后经过一番查找python处理找到了小小bug，昨晚还第一次在Github上提了pr...

此处有点快乐，使用过程种，也期望dify能越来越完善吧，工作流能嵌套调用，增加上数据库？国产之光，期待！

最近就是在哲学，就是竟然能看进去，我也是有点佩服我自己了...

快到端午节了，俺们都端午安康！