Yakha Inkumbulo Ephikelelayo Ye-ejenti ye-AI nge-AIrecall

Yini i-AIrecall?

I-AIrecall iyisendlalelo sememori yesikhathi eside esingezwa kalula kuma-agent e-AI. Igcina ulwazi oluwusizo phakathi kwamaseshini, esikhundleni sokuphoqa i-agent ukuthi ithembele kuphela efasiteleni layo lamanje lomongo. Le phrojekthi ihlanganisa i-Python SDK, iseva yememori ye-Go, isitoreji se-SQLite, kanye ne-vector index esebenza ngaphakathi kwenqubo.

I-AIrecall yakhelwe ukugcina ama-prompt egxile. Esikhundleni sokufaka wonke umlando wokuxhumana esicelweni ngasinye, i-agent yakho ingathola izinkumbulo ezihlobene kakhulu nejika lamanje.

Izinhlobo Nezici Eziyinhloko Zememori

• Imemori ye-episodic: Igcina izehlakalo ngokulandelana, okuhlanganisa izingxoxo, izenzo, nemiphumela.
• Imemori ye-semantic: Igcina amaqiniso nezintandokazi ezihlala ziwusizo, ezikhishwe ezenzakalweni zangaphambilini.
• Ukuthola ngohlanganiso: Kuhlanganisa ukufanisa amagama angukhiye nokufana kwama-vector kuwo womabili amemori e-episodic nawe-semantic.
• Ukufingqa okuzenzakalelayo: Icindezela ngezikhathi ezithile izehlakalo ezindala futhi ikhiphe amaqiniso ahlala ebalulekile esikhundleni sokusula umlando nje.
• Isitoreji sasendaweni esiphikelelayo: Isebenzisa i-SQLite enenkomba yama-vector esebenza ngaphakathi kwenqubo.
• Ukusebenza okuguquguqukayo: Iseva yememori ingasebenza endaweni noma njenge-sidecar.

Isebenza Kanjani i-AIrecall

Kujika ngakunye, i-Python SDK ithumela umbuzo wamanje ku-AIrecall Core ngocingo lwasendaweni noma nge-gRPC. I-core isesha imemori ye-episodic ne-semantic isebenzisa amagama angukhiye nokufana kwama-vector, ihlele imiphumela, bese ibuyisa izinkumbulo eziphezulu ezifanele ukuze zifakwe ku-prompt ye-agent.

Ngemuva, umfingqi ucubungula izehlakalo ezindala. Ukhipha ulwazi oluhlala luwusizo alufake kwimemori ye-semantic futhi ahlanganise umlando osele, okusiza ama-agent asebenza isikhathi eside ukuthi agcine ulwazi oluwusizo ngaphandle kokukhulisa njalo umongo wawo osebenzayo.

Ukufakwa Nokulungiselela

Faka i-Python SDK

Faka iphakheji engeniswa yi-agent yakho ye-Python:

pip install airecall-sdk

Qalisa Isitoreji Sasendaweni

Dala isitoreji sememori sasendaweni esisekelwe ku-SQLite usebenzisa isixhumi somugqa womyalo:

airecall init

Lokhu kulungiselela kukuvumela ukuthi uqale ukugcina nokubuyisa izinkumbulo ngokushesha ngaphandle kokudinga isevisi ehlukile.

Faka noma Qalisa Iseva Yememori

Uma ufuna iseva yememori eyinhloko itholakale endaweni noma njenge-sidecar, faka i-server extra:

pip install airecall-sdk[server]

Ungabe ususebenzisa iseva njengenqubo ezimele:

airecall serve

I-SDK ngokwayo isebenzisa umtapo ojwayelekile kuphela. Iseva isebenzisa ukuncika okukodwa kwe-Go okumsulwa, ngokwe-README yephrojekthi.

Ukusetshenziswa Okuyisisekelo

Dala Isenzakalo Sememori

Ngenisa Memory bese unikeza i-agent isihlonzi:

from airecall import Memory

memory = Memory(agent_id="support-bot")

I-agent_id ikhomba i-agent okugcinwa futhi kubuyiswe kuyo izinkumbulo zayo. Sebenzisa isihlonzi esifanayo uma ufuna amaseshini akamuva afinyelele eqoqweni elifanayo lezinkumbulo.

Gcina Imemori Ye-Episodic

Sebenzisa remember ukurekhoda okuthile okwenzeke ngesikhathi sokuxhumana:

memory.remember(
    "User asked about refund policy for order ORD-9921, told 30-day window applies."
)

Lena imemori ye-episodic ngoba irekhoda izehlakalo ezithile kanye nomphumela wazo.

Khumbula Umongo Ofanele

Eseshini yakamuva, shayela recall ngombuzo wamanje:

context = memory.recall(
    "what did we tell this user about refunds?"
)

I-AIrecall isesha izinkumbulo ezigciniwe bese ibuyisa isenzakalo esifanele, esihlelwe ngokuhambisana kwaso. Uhlelo lwakho lungabe selungeza lowo mongo obuyisiwe ku-prompt ethunyelwa kumodeli.

Gcina Iqiniso Elihlala Lisebenza

Ulwazi okufanele luhlale luwusizo nangemva kwesenzakalo esisodwa lungaphakanyiswa ngokucacile nge-remember_fact:

memory.remember_fact(
    key="preferred_contact",
    value="email, not phone"
)

Lokhu kudala imemori ye-semantic yentandokazi ezinzile, esikhundleni sokurekhoda esinye isenzakalo esilandelana ngesikhathi kuphela.

Ukusebenzisa I-adapter ye-LangChain

I-AIrecall ihlanganisa i-adapter ye-LangChain ebizwa ngokuthi MemoryRetriever. Goqa isenzakalo esikhona se-Memory ukuze wakhe i-retriever:

from airecall import Memory
from airecall.adapters.langchain import MemoryRetriever

memory = Memory(agent_id="support-bot")
retriever = MemoryRetriever(memory)

I-README ikhombisa ukwakha i-adapter kodwa ayicacisi okunye ukulungiselela kwe-chain noma kwe-retriever. Landela i-API yamanje ye-adapter lapho uhlanganisa into ewumphumela kuhlelo olukhulu lwe-LangChain.

Indlela Esebenzayo Yokuhlanganisa

1. Dala isenzakalo se-Memory se-agent.
2. Ngaphambi kokukhiqiza impendulo, shayela recall ngombuzo wamanje womsebenzisi.
3. Faka izinkumbulo ezifanele ezibuyisiwe ku-prompt yemodeli njengomongo osekelayo.
4. Ngemva kokuxhumana, shayela remember ukuze ulondoloze izehlakalo, isenzo, noma umphumela obalulekile.
5. Sebenzisa remember_fact lapho uhlelo lwakho luthola intandokazi noma iqiniso elihlala isikhathi eside okufanele ligcinwe ngokucacile.

Le ndlela igwema ukufaka wonke umlando kuwo wonke ama-prompt, kuyilapho isanikeza i-agent ukufinyelela ezinqumweni zangaphambilini nasezintandokazini zabasebenzisi lapho kudingeka.

Amathiphu Athuthukile

Hlukanisa Izehlakalo Olwazini Oluhlala Lusebenza

Sebenzisa imemori ye-episodic kulokho okwenzeke ngesikhathi sokuxhumana okuthile. Sebenzisa imemori ye-semantic kumaqiniso nasezintandokazini ezizinzile. Ngokwesibonelo, ingxoxo ngembuyiselo ingeyesenzakalo, kanti intandokazi yomsebenzisi yokuxhumana nge-imeyili iyiqiniso elihlala lusebenza.

Gcina Izihlonzi Zama-agent Zifana

Khetha i-agent_id ezinzile yamaseshini adinga ukwabelana ngememori. Ukushintsha lesi sikhombi kudala ububanzi bememori obuhlukile ngokombono wohlelo lokusebenza.

Khetha Imodi Efanele Yokusetshenziswa

Ngokuhamba komsebenzi okulula kwasendaweni, qalisa isitolo esisekelwe ku-SQLite nge-airecall init. Ngokwakhiwa okudinga inqubo yememori ezinikele, faka i-server extra bese usebenzisa i-airecall serve endaweni noma njenge-sidecar.

Vumela I-Index Ifudumale

Uma umbuzo wokuqala uphakamisa i-RecallTimeout, i-vector index kungenzeka isafudumala. Zama futhi ngemva kokuba i-health endpoint ibike ukuthi status: ready.

Sebenzisa Indlela Ethembekile Yokugcina

Uma izinkumbulo zibonakala zishoda ngemva kokuqalisa kabusha, hlola indlela ye-storage.backend. Indlela ehlobene ihunyushwa ngokusekelwe ku-working directory yenqubo ye-server, ngakho ukuqala i-server kwenye i-directory kungayikhomba kwesinye isitolo. Sebenzisa i-working directory noma indlela yokugcina engaguquki.

Isiphetho

AIrecall inikeza ama-AI agent imemori eqhubekayo ye-episodic ne-semantic ngaphandle kokudinga ukuthi konke ukuxhumana kwangaphambilini kuhlale ku-prompt. Nge-Python API encane, ukugcinwa okusekelwe ku-SQLite, ukubuyisa okuhlanganisiwe, kanye nokufingqa okuzenzakalelayo, inikeza indlela esebenzayo yokugcina umongo ofanele phakathi kwamasession. Qala nge-airecall init, rekhoda ama-episode nge-remember, uwabuyise nge-recall, bese ugcina ulwazi oluhlala isikhathi eside nge-remember_fact.