AI ƒe Nudɔdɔ 12: Aleke míate ŋu atrɔ Prompt?

Prompt trɔtrɔ (Prompt Engineering / Optimization) nye nyateƒe si na gbegbe gã la ɖo to, vevietɔ le RAG sistemowo me, ena wòdze gbeɖeɖe ŋu, eye wòle be wòawɔ ɖe nyateƒe siwo wòxɔ la nu, mewɔ susu aɖeke o, eye wòaɖo do go ƒe mɔnu nyui aɖe.

ɖeka, Prompt Trɔtrɔ ƒe Nuhiaɖeɖe Seviwo

Kɔkɔ > Eɖokpe: Ŋutinya kpui siwo me nu le wɔwɔ le eme la wɔa dɔ nyuie wu esiwo me nu le wɔwɔ le ɖeɖem o.
Na Mɔniameɖoɖo Sɔ: Na gbegbe la meɖo eƒe nɔnɔme, “Aleke wòate ŋu awɔ” kple “Aleke wòmate ŋu awɔ o.”
Na Kpɔɖeŋuwo: Few-shot wɔa dɔ nyuie wu Zero-shot.
Wɔe be woate ŋu akpɔe: Na gbegbe la nado gome kple ka woɖoe, alo xɔse, bene woate ŋu ato eme akpɔe.
Trɔtrɔme: Dze le gɔmeɖoɖoa, ɖe asi le eteƒe ɖeka ko dzi, trɔ trɔ nu.

Eve, Trɔtrɔ Nyateƒewo (Tso bɔbɔe yi sesẽ)

1. Mɔnuɖoɖo (System Prompt)

Wò nye ame kpekpeɖeŋu gã aɖe le dɔwɔwɔ me. Àte ŋu aɖo eteƒe si wòxɔ tso 【Nudɔdɔwo】 me ko. Ne mènya eɖoɖo o la, gblɔ be “Nudɔdɔɖoɖoa me nya aɖeke mele o,” meka nu ɖe ŋutsuwò o.

Dɔwɔwɔ: Eɖoa li kple gbeƒãɖeɖe.
Trɔtrɔme: Gbeƒãɖeɖe (wɔwɔ / lɔlɔ̃), mɔniameɖoɖo sesẽ (sesẽ / gblɔe).

2. Ŋutinya Kɔkɔ

❌ Vɔ̃: “Ðo nya na ame.”
✅ Nyui: “Dzini ko ɖe 【Nudɔdɔwo】 siawo me. Ne nudɔdɔɖoɖoa mele eteƒe o, gblɔ ‘Mate ŋu aɖo nya sia o.’”

3. Do Go Mɔnu Ƒoƒo

Míedi be nàɖo go le JSON mɔnu sia nu:
{
  “answer”: “Wò ŋuɖoɖoa”,
  “confidence”: “kɔkɔ / titina / bɔbɔ”,
  “sources”: [1, 3]
}

Dɔwɔwɔ: Ena woate ŋu ahe eme, aɖo ŋkume, ahe le eteƒe.

4. Kpɔɖeŋu Few-shot (Ewɔa dɔ nyuie)

Kpɔɖeŋu 1:
Nya: Ŋkeke nɔnɔme le afi?
Nudɔdɔwo: Ŋkeke nɔnɔme: Ƒe ɖeka 5, Ƒe ewo 10.
Ŋuɖoɖo: Ƒe ɖeka 5, Ƒe ewo 10.

Kpɔɖeŋu 2:
Nya: Aleke woatsɔ ga aɖo dɔwɔwɔ?
Nudɔdɔwo: Dɔwɔɣi 1.5 tɛ, gbeɖé 2 tɛ.
Ŋuɖoɖo: Dɔwɔɣi 1.5 tɛ, gbeɖé 2 tɛ.

Azɔ míaɖo nya:
Nya: {Ame ƒe nya}
Nudɔdɔwo: {Nudɔdɔ si wokpɔ}
Ŋuɖoɖo:

Nyateƒe: Kpɔɖeŋuwo nato mɔ vovovowo dzi, eye wònyo be wòna kpɔɖeŋu aɖe si me womate ŋu aɖo nya o la.

5. Ekpea Ne Wozã Nudɔdɔwo

Le ŋuɖoɖoa ƒe nuwuwu, zã [citation:X] na fia sɔhɛ si wòxɔ tso. Abe: “Ŋkeke nɔnɔme nye 5 [citation:1].” Ne wòxɔ tso nudɔdɔ geɖe me, ɖe wo dometɔ ɖesiaɖe ŋu.

6. Na Mɔniameɖoɖo Si Me Womate Ŋu Aɖo Nya O

Sesẽ: “Ne nudɔdɔɖoɖoa melɔ̃ kple nya la o, gblɔ ‘Nudɔdɔwo melɔ̃ o.’”
Gblɔe: Zã xɔse si wɔ akɔnta, ne wòle teƒe si wòdi o, tso mɔniameɖoɖo si me womate ŋu aɖo nya o la.

7. Susugbalẽ (Chain-of-Thought) na Nusiwo Tete Di

Nya: Ame kae nye Zhang ƒe aƒe? 
Dzidzime: 1. Di Zhang ƒe dɔwɔƒe. 2. Di dɔwɔƒea ƒe kplɔla. 3. Na ŋuɖoɖo mamlɛtɔ.
Míedi be nàsusui ɖe dzidzimea nu hena ŋuɖoɖo.

8. Seɖoɖo Si Nye Gbeɖe (Negative Prompting)

Meka nu ɖe ŋutsuwò o. Mekpɔ nyakpakpa abe ‘ate ŋu’, ‘ɖewohu’ o. Mègado xexlẽme siwo mele nudɔdɔɖoɖoa me o.

Etɔ̃, Aleke Míate Ŋu Akpɔ Esi Prompt Nyui?

Nɔnɔme	Gɔmeɖeɖe	Aleke Míate Ŋu Asɛ
Nuteƒefa	Ŋuɖoɖoa le nudɔdɔwo nu sɔsɔ̃a?	Ame alo RAGAS ƒe Nuteƒefa
Mɔniameɖoɖo me ɖoɖo	Ne wòle be woaɖo nya o, wòɖoaa?	Le test set si me nya mele o la, bu akɔnta
Mɔnu Ƒoƒo ƒe Se	Eɖo go le JSON / citation mɔnua nu?	Regex ƒo nu
Ame ƒe Dzidzɔ	Ŋuɖoɖoa le nyuiea?	Le dɔ me / A/B testing

Aɖaŋu: Da test set sue aɖe (20-50 sɔxɛ), tsɔ prompt ɖesiaɖe ŋlɔ eme, ŋlɔ nɔnɔme siwo trɔ.

Ene, Nuɖoɖi Geɖewo Kple Trɔtrɔ Mɔnuwo

Dɔwɔwɔ	Nuɖoɖi	Trɔtrɔ Mɔnu
Gbegbe la mekpɔ nudɔdɔwo o, eye wòɖo nya ɖeka	Ðoɖoɖoa menye sesẽ o	Trɔ ɖe “Dzini ko ɖe nudɔdɔwo”, eye wòzã few-shot si me womate ŋu aɖo nya o
Gbegbe la gblɔna “Ménya o” ɖaa	Mɔniameɖoɖo si me womate ŋu aɖo nya o la le dzi	Eɖe mɔniameɖoɖoa ɖe anyi alo kpɔ nudɔdɔ si wokpɔ la
Do go ƒe mɔnu mele bɔbɔe o, mele JSON nu o	Ðoɖoɖoa mekɔ o	Tsɔ kpɔɖeŋu sesẽ aɖe, alo zã function calling
Ŋuɖoɖo le didi / kpui	Mèɖo didime o	“Zã nya etɔ̃ meɖeɖe o”
Susugbalẽ si tete di me wɔ o	Gbegbe la ƒe nuhiaɖeɖe mekpe o	Bi be wòaɖe dzidzime afia, alo trɔ ɖe gbegbe sesẽ aɖe
Xexlẽme / ŋkeke siwo mele o le ŋuɖoɖoa me	Gbegbe la ɖo ea ŋutɔ ƒe nyanu	Ðo ŋu be “Mègazã xexlẽme si le wò susu me o, kpɔ nudɔdɔwo ko”