大模型实战篇：，的左右互搏之术

文章通过实际源码详细介绍了BasicReflection模式的实现方法，包括构建Generator和Reflector的过程。阅读本文可以帮助读者更好地理解BasicReflection的概念和实现过程。

在前面几篇文章中，风叔依次介绍了REWOO、Plan-and-Execute和LLMCompiler三种更侧重规划能力的AIAgent设计模式。

从最初的ReAct模式出发，加入规划能力即演变成REWOO；再加上Replan能力即演变成Plan-and-Execute；最后再加上DAG和并行处理能力，即演变成LLMCompiler。

从这篇文章开始，我们将转向另外几种侧重反思的AIAgent模式，我们首先从BasicReflection开始。

一、BasicReflection的概念

BasicReflection可以类比于左右互博。左手是Generator，负责根据用户指令生成结果；右手是Reflector，来审查Generator的生成结果并给出建议。在左右互搏的情况下，Generator生成的结果越来越好，Reflector的检查越来越严格，输出的结果也越来越有效。

下图是BasicReflection的原理，非常简单。

二、BasicReflection的实现过程

下面，风叔通过实际的源码，详细介绍BasicReflection模式的实现方法，具体的源代码地址在文章结尾处获取。

第一步构建Generator

在下面的例子中，我们先构建一个能够生成5段话的文章生成器。

首先，我们给Generator约定了Prompt，大模型实战篇：，的左右互搏之术告诉Generator具体的角色和目标，并且要求如果用户给出建议，需要给出修改后的版本

然后选择LLM模型，并构建Generator

最后要求输出Generator生成的内容

fromlangchain_core.messagesimportAIMessage,BaseMessage,HumanMessage

fromlangchain_core.promptsimportChatPromptTemplate,MessagesPlaceholder

fromlangchain_fireworksimportChatFireworks

下面是Generator的InitialResponse，可以看出Generator虽然按照我们的要求生成了初始回复，但是回复的整体质量还有很大的提升空间。

Title:TheRelevanceofTheLittlePrinceinModernChildhood

TheLittlePrince,anovellabyAntoinedeSaint-Exupéry,hasbeenachildhoodfavoriteforgenerations.Despitebeingpublishedoversevendecadesago,itstimelessthemescontinuetoresonatewithmodernchildren,makingithighlyrelevantincontemporarychildhood.

Firstly,thestoryexploresthecomplexnatureofhumanrelationships,whichisparticularlyrelevantformodernchildrengrowingupinanincreasinglyconnectedyetimpersonalworld.Throughthelittleprince’sencounterswithvariousgrown-upsondifferentplanets,thebookhighlightstheimportanceofgenuineconnectionsandunderstanding.Inanagewheredigitalcommunicationoftenreplacesface-to-faceinteraction,thismessageismorepertinentthanever.TheLittlePrinceencourageschildrentolookbeyondsuperficialrelationshipsandseekdeeperconnections,fosteringempathyandemotionalintelligence.

Secondly,thebookdealswiththeconceptofresponsibilityandself-discovery,elementsthatareintegraltoachild’sgrowth.Thelittleprince’sjourneyisessentiallyaquestforself-discovery,leadinghimtorealizehisresponsibilitytowardshisbelovedrose.Thisnarrativeencouragesmodernchildrentoembracetheirindividualitywhileunderstandingthesignificanceoftheiractions.Inasocietythatoftenoverlookstheemotionalwell-beingofchildren,TheLittlePrinceoffersarefreshingperspectiveonpersonalgrowthandresponsibility.

Thirdly,thebookaddressesthechallengingthemeoflossandbereavement.Thelittleprince’sdeparturefromhisasteroidandhissubsequentencounterswiththefoxandthesnakeareprofoundreflectionsontheinevitabilityoflossandtheimportanceofcherishingrelationships.Inatimewhenchildrenareexposedtovariousformsofloss,fromthedeathoflovedonestoenvironmentaldegradation,TheLittlePrinceprovidesagentleyetpowerfulwaytounderstandandcopewiththeseexperiences.

However,somecriticsarguethatthebook’spaceandabstractconceptsmightbechallengingformodernchildrenwithshortattentionspans.Toaddressthis,arevisedversioncouldincorporatemorevisualelementsandinteractiveactivitiestoengageyoungreadersbetter.Additionally,supplementarymaterialsexplainingthebook’sthemesinsimplertermscouldbeprovidedforparentsandeducatorstouseindiscussionswithchildren.

Inconclusion,TheLittlePrinceremainsrelevantinmodernchildhoodduetoitsexplorationofhumanrelationships,self-discovery,andloss.Thesethemes,wrappedinacaptivatingnarrative,offervaluablelessonsformodernchildren.Whilesomeadaptationsmaybenecessarytocatertothepreferencesoftoday’schildren,theessenceofthestoryremainsapowerfultoolforteachingemotionalintelligence,personalgrowth,andresilience.

第二步构建Reflector

针对Generator，我们构建一个专门的Reflector。

同样的，我们要先给Reflector约定Prompt，告诉Reflector“你是一名正在给论文评分的老师。请对用户的论文提出批评和建议，包括长度、深度、风格等要求。”

然后选择LLM模型，并构建Reflector

下面是针对Generator的InitialResponse，Reflector反馈的评语和建议。Reflector给出了B 的评级，并且按照我们的要求，从文章的长度、深度、风格等维度给出了具体的建议和总结。

EssayGrade:B

Theessayyousubmittedprovidesaclearandwell-structuredargumentabouttherelevanceofTheLittlePrinceinmodernchildhood.Youhavedemonstratedastrongunderstandingofthetextanditsthemes,andhaveeffectivelyappliedthemtothecontextofcontemporarychildhood.However,therearesomeareaswhereimprovementcouldbemadetoenhancethedepth,style,andoverallflowofyouressay.

1.Length:Whileyouressayiswell-writtenandinformative,itisrelativelybrief.Expandingoneachpointwithmoredetailedanalysisandexampleswouldstrengthenyourargumentanddemonstrateamorecomprehensiveunderstandingofthetext.Aimforaminimumof500wordstoallowforamorein-depthexplorationofyourideas.

2.Depth:Althoughyouhavetouchedupontherelevanceofthenovel’sthemes,furtheranalysisisneededtotrulyestablishitssignificanceinmodernchildhood.Forexample,whendiscussingthecomplexnatureofhumanrelationships,delveintohowthedigitalageaffectschildren’scommunicationskills,andhowTheLittlePrinceaddressesthisissue.Providingconcreteexamplesfromthetextandconnectingthemtoreal-worldscenarioswillmakeyourargumentmorecompelling.

3.Style:Toengageyourreadersmoreeffectively,considervaryingyoursentencestructureandlength.Usingamixofsimple,compound,andcomplexsentenceswillimprovetheflowofyouressayandmakeitmoreengagingtoread.Additionally,watchyourtenseconsistency.Ensurethatyoumaintainthesametensethroughoutyouressaytoavoidconfusion.

4.Recommendations:Whileyoursuggestionsforadaptationareagoodstart,theycouldbeexpandedupontoprovidemorecomprehensiverecommendations.Forexample,youmaywanttodiscussdifferentmethodsofincorporatingvisualelementsandinteractiveactivities,suchasillustrations,quizzes,ordiscussionquestions.Thiswilldemonstratethatyouhavethoughtfullyconsideredtheneedsofmodernchildrenandhavedevelopedstrategiestoaddressthesechallenges.

5.Conclusion:YourconclusioncouldbenefitfromastrongersummarizationofyourkeypointsandanassertivefinalstatementabouttherelevanceofTheLittlePrinceinmodernchildhood.Tyingallyourargumentstogetherinaconciseandpowerfulmannerwillleavealastingimpressiononyourreadersandsolidifyyourposition.

Overall,youressayiswell-researchedandprovidesasolidfoundationforacompellingargumentabouttherelevanceofTheLittlePrinceinmodernchildhood.Withsomeexpansion,deeperanalysis,andstylisticimprovements,youressaycanachieveanevenhigherlevelofexcellence.

第三步循环执行

接下来，我们就可以循环执行这个“生成–检查”的过程，重复规定的次数，或者约定当Generator的生成结果达到多少分时，停止循环。

大家可以看到，在循环过程中，我们也加入了人类的反馈建议，有助于Generator和Reflector学习迭代。

forchunkingenerate.stream(

{“messages”:[request,AIMessage(content=essay),HumanMessage(content=reflection)]}

):

print(chunk.content,end=””)

下面，我们构建流程图，将Generator、Reflector等节点添加进来，循环执行并输出结果。

classState(TypedDict):

messages:Annotated[list,add_messages]

asyncdefgeneration_node(state:Sequence[BaseMessage]):

returnawaitgenerate.ainvoke({“messages”:state})

builder=StateGraph(State)

builder.add_node(“generate”,generation_node)

builder.add_node(“reflect”,reflection_node)

builder.add_edge(START,“generate”)

defshould_continue(state:List[BaseMessage]):

iflen(state)>6:

#Endafter3iterations

returnEND

return“reflect”

builder.add_conditional_edges(“generate”,should_continue)

builder.add_edge(“reflect”,“generate”)

graph=builder.compile()

asyncforeventingraph.astream(

[

HumanMessage(

content=”GenerateanessayonthetopicalityofTheLittlePrinceanditsmessageinmodernlife”

)

],

print(event)

print(“—“)

至此，BasicReflection的完整流程就介绍完了，非常简单，相信哪怕是没有AI基础的同学也能看懂。

BasicReflection的架构，非常适合于进行相对比较发散的内容生成类工作，比如文章写作、图片生成、代码生成等等。

总体而言，BasicReflection是一种非常高效的反思类AIAgent设计模式。BasicReflection的思路非常朴素，使用成本较低。但是在实际应用中，BasicReflection也面临着一些缺陷：

对于一些比较复杂的问题，显然需要Generator具备更强大的推理能力

Generator生成的结果可能会过于发散，和我们要求的结果相去甚远

在一些复杂场景下，Generator和Reflector之间的循环次数不太好定义，如果次数太少，生成效果不够理想；如果次数太多，对token的消耗会很大。

我们有两种方法来优化BasicReflection，一种是边推理边执行的SelfDiscover模式，一种是增加了强化学习的Reflexion模式。

题图来自Unsplash，基于CC0协议。