AI ƒe Kpɔɖeŋu Biawo 6: Gblɔ AI Agent ƒe Dɔwɔna Mɔnu vevi etɔ̃: ReAct, Plan-and-Solve kple Reflection

AI Agent ƒe Dɔwɔna Mɔnu vevi etɔ̃: ReAct, Plan-and-Solve kple Reflection

AI Agent nye nugɔmesese si ate ŋu adzro eɖokui, aɖo nyatakaka me, eye wòawɔ dɔ. Eƒe dɔwɔna mɔnu vevi etɔ̃ nye: ReAct, Plan-and-Solve kple Reflection. Le gɔmeɖeɖe siawo me, míakpɔ nɔnɔmetata kple kpɔɖeŋu kɔd siwo le eme.

1. ReAct (Reasoning + Acting)

Tameɖoɖo vevi: Bubu ɖe nu ŋlɔɖi (Reasoning) kple dɔwɔwɔ (Acting) siwo wɔna ɖekae dzi. Agent la le afi sia afi gbɔna gbɔna ɖoa tame ɖe nɔnɔme si li fifia kple nu si wòawɔ le afɔɖeɖe si gbɔna (nu ŋlɔɖi), emegbe wɔa dɔ aɖe (abe yɔ ɖoɖowɔnu, di nyatakaka ene), eye wòatsɔ se nu si do goe ayi edzi aɖo tame.

Nɔnɔmetata:

[Gɔmedzedze nɔnɔme] → [Nu ŋlɔɖi: Bu afɔɖeɖe si gbɔna] → [Dɔwɔwɔ: Wɔ dɔ la] → [Kpɔ nu si do go] → [Nu ŋlɔɖi: Trɔ asi le ɖoɖo ŋu] → ... → [Ŋuɖoɖo mamlɛtɔ]

Kpɔɖeŋu kɔd (kɔd si me nyateƒe mele o) :

def react_agent(question):
    context = []
    while not solved:
        # Nu ŋlɔɖi: Wɔ susu ƒe afɔɖeɖe
        thought = llm.generate_thought(question, context)
        # Dɔwɔwɔ: Tia dɔ si wòawɔ le susu ŋu
        action = llm.choose_action(thought)
        # Wɔ dɔ la, eye nàkpɔ nu si do go
        observation = execute_action(action)
        # Tsɔ susu, dɔwɔwɔ, kple nu si do go kpe ɖe nɔnɔme ŋu
        context.append((thought, action, observation))
    return final_answer

Kpɔɖeŋu:
- Amebia: “Aleke yame le egbe le Beijing?”
- Agent ƒe nu ŋlɔɖi: “Mehiã be mabia yame ƒe API, mehiã du ƒe ŋkɔ kple ŋkeke.”
- Dɔwɔwɔ: Yɔ yame ƒe API (ŋusẽtɔwo: Beijing, egbe)
- Nu si do go: Trɔ: “Ɣe le dzo, 25°C”
- Nu ŋlɔɖi: “Enye nyatakaka si mele be matrɔ.”
- Do go: “Egbe le Beijing, 25°C, ɣe le dzo.”

2. Plan-and-Solve

Tameɖoɖo vevi: Gbã la, wɔ ɖoɖo bliboa (Plan), emegbe wɔe afɔɖeɖe ɖe afɔɖeɖe me (Solve). Le ɖoɖo ƒe ɣeyiɣi me, woma dɔ si me veve le la ɖe afɔɖeɖe sue sue me, eye le dɔwɔwɔ ƒe ɣeyiɣi me, wowɔna ɖe ɖoɖo nu, eye wòate ŋu atrɔ asi le ɖoɖo ŋu le emeɖokui si do go nu.

Nɔnɔmetata:

[Dɔ] → [Wɔ ɖoɖo: Ma dɔ la ɖe afɔɖeɖe sue sue me] → [Wɔ afɔɖeɖe 1] → [Wɔ afɔɖeɖe 2] → ... → [Wɔ afɔɖeɖe N] → [Ŋuɖoɖo mamlɛtɔ]

Kpɔɖeŋu kɔd:

def plan_and_solve(task):
    # ɖoɖo ƒe ɣeyiɣi
    plan = llm.generate_plan(task)  # Abegã: ["Di nyatakaka", "Ƒo nu siwo le eme nu", "Ŋlɔ nyatakakaɖeɖe"]
    context = {}
    for step in plan:
        # Wɔ afɔɖeɖe ɖe sia ɖe
        result = execute_step(step, context)
        context[step] = result
    # Ƒo nu siwo katã do go la nu
    final = llm.synthesize(context)
    return final

Kpɔɖeŋu:
- Dɔ: “Ŋlɔ blog si ku ɖe AI Agent ŋu”
- ɖoɖo:
1. Di AI Agent ƒe gɔmeɖeɖe kple nu yeyewo
2. Xlẽ eye nàƒo nu vevi siwo le eme nu
3. Ŋlɔ blog la ƒe alɔdzedro
4. Yɔ nu me kple nya
5. Trɔ asi le eme eye nàɖe mɔ
- Dɔwɔwɔ: Wɔ afɔɖeɖe ɖe sia ɖe ɖe ɖoɖo nu, eye wòana blog la le nuwuwu.

3. Reflection

Tameɖoɖo vevi: Agent la le dɔwɔwɔ me alo le emegbe, bua eɖokui ƒe dɔwɔwɔ ŋu (Reflection), srɔ̃a nu si do go, eye wòtrɔa asi le dɔwɔwɔ siwo gbɔna ŋu. Esia me ɖea eɖokui ƒe moɖeɖe, vodada ɖɔɖɔɖo, alo dɔwɔna ƒe mɔnu nyuiyiyi.

Nɔnɔmetata:

[Dɔwɔwɔ] → [Kpɔ nu si do go] → [Bu eɖokui ŋu: Kpɔe ɖa be edze alo o] → [Ne edze o: Trɔ asi le dɔwɔna ŋu] → [Gawɔ dɔ la] → ... → [Edze]

Kpɔɖeŋu kɔd:

def reflection_agent(task):
    max_attempts = 3
    for attempt in range(max_attempts):
        action = llm.generate_action(task)
        result = execute(action)
        # Bu eɖokui ŋu
        reflection = llm.reflect(task, action, result)
        if reflection['success']:
            return result
        else:
            # Trɔ asi le dɔ alo dɔwɔna ŋu le eɖokui ŋuɖoɖo nu
            task = reflection['improved_task']
    return None

Kpɔɖeŋu:
- Dɔ: “Xlẽ 1234 * 5678”
- Dɔwɔwɔ: Xlẽe enumake, eye nàkpɔ 7006652
- Eɖokui ŋuɖoɖo: Kpɔ xlẽɖeɖe ƒe mɔnu, eye nàkpɔ be vodada le afi aɖe
- Trɔ asi: Gaxlẽe, eye nàkpɔ nu si le eme nyuie 7006652 (nyateƒe la, enye nyuie)
- Ne vodada li haɖe, yi edzi nàbu eɖokui ŋu va se ɖe eme wòdze.

Ƒoƒo ƒe tatia

Le dɔwɔwɔ me la, zi geɖe wowɔa mɔnu etɔ̃ la ɖekae, abe ReAct me wotsɔa eɖokui ŋuɖoɖo kpe ɖe eŋu, alo Plan-and-Solve me wole eɖokui ŋuɖoɖo wɔna le afɔɖeɖe ɖe sia ɖe megbe.