Ukusebenzisa ama-agent e-AI ngama-Databricks: umhlahlandlela ophelele osebenzayo

Informatec Digital » Izinsiza » Ukusebenzisa ama-agent e-AI ngama-Databricks: umhlahlandlela ophelele osebenzayo

Ukwakha ama-ejenti e-AI ku-Databricks akuseyona indawo ekhethekile yamaphrofayili obuchwepheshe obukhulu: phakathi kwezimo ezibonakalayo njenge-AI Playground , ukuzenzekela nge-Agent Bricks, kanye nohlaka lwe-MLflow, manje usungakwazi ukusuka emcabangweni uye kowohlobo lweprototype kanye nosuka kowohlobo lweprototype uye kowokukhiqiza ngaphandle kokulahleka enkambisweni. Kulo mhlahlandlela, sihlanganisa konke okudingayo ukudala, ukuhlola, kanye nokuthumela ama-ejenti abiza amathuluzi, abuze idatha, futhi ahlanganiswe nokuphathwa kwe -Unity Catalog.

Ngaphezu kokuhlola indlela yokungasebenzisi ikhodi kanye nokuthuthukiswa kwe-Python, uzofunda ukuthi ungawafaka kanjani ama-ejenti nge -MLflow (ChatAgent kanye ne-ResponsesAgent) , ubhalise izinsiza kukhathalogi, unike amandla ukuqapha ngesikhathi sangempela, kanye nokulinganisa usebenzisa i-Model Serving. Siphinde sifake uhlelo lokusebenza lwesampula olwakhiwe nge -React kanye ne-FastAPI ukuze luhlanganise i-ejenti yakho kalula esibonakalayo sewebhu sebhizinisi.

Izinketho zokudala ama-agent ku-Databricks

Uma ubeka phambili isivinini kanye nobulula, i-Agent Bricks inikeza ukuhamba komsebenzi okuzenzakalelayo kokwakha izinhlelo ze-ejenti ezisezingeni eliphezulu, ezihambisana nesizinda. Ucacisa isimo sakho sokusetshenziswa bese uxhuma idatha yakho ye -Unity Catalog ; ipulatifomu ikhiqiza izinhlobo ze-ejenti, ithuthukise ngokuhlola, futhi ikuvumela ukuthi uthuthukise, konke ngokuphatha nokuphepha okwakhelwe ngaphakathi.

Uma udinga ukulawulwa okuhlelekile futhi ufuna ukusebenzisa kabusha imitapo yolwazi ejwayelekile, indlela esekelwe kukhodi ihlanganisa i-Mosaic AI Agent Framework kanye ne-MLflow . Nge-Python, ungachaza i-agent logic, unike amandla izingcingo zamathuluzi, wenze ukulandelela okuningiliziwe, bese ekugcineni upakisha futhi uhambise ekukhiqizeni nge-Databricks ecosystem.

Uma usuvele usebenzisa izinhlaka zenkampani yangaphandle, kukhona ukwesekwa kwendabuko kwe -LangGraph/ LangChain , LlamaIndex , noma ukuqaliswa kwe-Python ngokwezifiso. Ngenxa yama-signatures emodeli ye-MLflow kanye ne-interfaces ye-ejenti, ungazixhuma emandleni okuhlola nokuqapha ngaphandle kokubhala kabusha kusukela ekuqaleni.

Ufuna ukwenza i-prototype ngesivinini sombani? Nge -AI Playground, ukhetha imodeli kanye nomshini obonakalayo, ungeze amathuluzi ku-LLM ngesibonisi sekhodi ephansi, bese uxoxa ukuze uqinisekise. Ngemuva kwalokho, uthumela i-ejenti kukhodi elungele ukukhiqiza bese uqhubeka nomjikelezo wokuphila ojwayelekile wokuthuthukiswa.

Amasignesha angamamodeli nokuhambisana nezici zeplathifomu

I-Azure Databricks isebenzisa amasiginesha emodeli ye-MLflow ukuchaza uhlelo lokufaka nokukhipha lwama-ejenti. Izici zomkhiqizo—njenge-AI Playground noma ukuhlolwa—zidinga ukuthi i-ejenti ivumelane nesinye samasiginesha asekelwayo ukuqinisekisa ukusebenzisana.

Uma ulandela indlela enconywayo lapho wakha i-ejenti, i-MLflow inquma ngokuzenzakalelayo isiginesha ehambisanayo ngaphandle kwezinyathelo ezengeziwe. Ngaphandle kwalokho, qinisekisa ukuthi ukusetshenziswa kwakho kuhambisana nesinye seziginesha ku- schema yefa yokufaka nokukhipha i-ejenti, ugweme ukumangala lapho kuhlanganiswa namathuluzi e-Databricks.

Okokufundisa Okusheshayo: Umenzeli Wokuqala Onohlaka Lomenzeli we-Mosaic AI

Umgomo uwukwakha i-ejenti ebonisana ne- I-LLM ibanjwe e-Databricks futhi ukwazi ukusebenzisa amathuluzi, isibonelo umsebenzi owakhelwe ngaphakathi system.ai.python_exec kusuka kukhathalogi yobumbano ukuze usebenzise ikhodi yePython. Ukwengeza, sizovumela i- ukulandela ngomkhondo nge-MLflow ukuze kuhlolwe imikhondo kanye nokuxilonga.

Izidingo eziyisisekelo zendawo yokusebenza: nika amandla izici ze-ejenti futhi ube nazo ukufinyelela ku-Model ServingEsikhathini sokusebenza, faka amaphakheji afana mlflow, databricks-openai kanye ne-agent packer (databricks-agent o databricks-agents(kuya ngomhlahlandlela oyilandelayo kanye nenguqulo), bese uqala kabusha i-Python ukuze ulayishe ukuncika.

Chaza ukuxhumeka kusevisi yemodeli usebenzisa iklayenti le-OpenAI elidalulayo Indawo yokusebenzaIklayenti, esebenzayo mlflow.openai.autolog() okokusebenza okuzenzakalelayo, futhi irekhoda amathuluzi atholakalayo. Phakathi kwazo, system.ai.python_exec Isebenza njengomtoliki we-Python ongenamthetho futhi ivumela i-LLM ukuthi ixazulule izibalo noma ilawule idatha phakathi nengxoxo.

With a function like run_agent(), uthumela ukwaziswa komsebenzisi ekugcineni kwengxoxo (Ngokwesibonelo, imodeli efana «idathabricks-claude-3-7-sonnet«nokuphatha kungenzeka ubiza amathuluzi ukubuyisela imilayezo yomsizi nezimpendulo zamathuluzi ohlwini oluhambisanayo.

Ukuze uhlole, sebenzisa umbuzo odinga ikhodi (isb., ukubala inombolo ye-Fibonacci ye-100 ) bese ubuyekeza kokubili umphumela we-wizard kanye ne- MLflow traces efakwe ku-notebook noma ku-experiment ehlobene. Lokhu kuzokuvumela ukuthi uthuthukise ukuhlelwa ngaphambi kokupakisha.

Ipakishwe nge-MLflow futhi yasetshenziswa Ekusezweni Kwemodeli

Indlela enconyiwe yokulungiselela i-ejenti ukuthi ikhiqize iwukusebenzisa isixhumi esibonakalayo I-MLflow ChatAgentUdinga indlela predict() eyamukela imilayezo, ishayela i-logic yomenzeli wakho, bese ibuyisela a I-ChatAgentResponse ngohlu lwemiyalezo (okuhlanganisa nemiyalezo emaphakathi ehlotshaniswa namathuluzi, uma ikhona).

Ukwenza lula ukugqilazwa, hlanganisa yonke ikhodi ibe yifayela elilodwa lePython (isibonelo, ngokusebenzisa %%writefile quickstart_agent.py (uma usebenza kwi-laptop). Engxenyeni yokugcina, shayela mlflow.models.set_model() ukudlulisa isibonelo sekilasi lakho (isb., QuickstartAgent) ukusungula i indawo yokungena yesiphetho.

Bhalisa i-ejenti Ikhathalogi Yobumbano con mlflow.pyfunc.log_model, kufaka phakathi izinsiza ozodinga ekukhiqizeni: indawo yokugcina yemodeli ye-LLM (isb., DatabricksServingEndpoint) kanye nemisebenzi yekhathalogi umenzeli azoyicela (isb., DatabricksFunction con system.ai.python_execNgakho, Databricks can Lungiselela ukufakazela ubuqiniso kulezi zinsiza lapho kuthunyelwa.

Ekugcineni, sebenzisa i-Agent API ukuze ukusabalalisa imodeli endaweni yokugcina yesevisi: from databricks import agents y agents.deploy(model_name, model_version)Lapho isiphetho siqala, uzokwazi Xoxa nge-AI Playgroundwabelane ngayo nabathintekayo bese uqala ukuqoqa impendulo.

Uma ufuna ukuqhubeka nokuthuthukisa, i-Databricks inikeza iziqondiso zokulinganisa ikhwalithi ye-ejenti , ukwakha ama-ejenti anokubuyisa (i-RAG) kudatha engahlelekile kanye nokulawula izingxoxo eziphendula izikhathi eziningi, kanye nezinye izindlela nge -LangGraph, i-Python emsulwa noma i-OpenAI.

I-ResponsesAgent: I-interface yesimanje yama-ejenti okukhiqiza

Noma kunjalo I-ChatAgent Isebenza kahle, i-MLflow iyayincoma namuhla I-ResponsesAgent Okwama-agent abalulekile. Lesi sixhumanisi sihambisana nohlelo Responses kusuka ku-OpenAI futhi ivumela ukuhlanganiswa kwama-ejenti avela noma yiluphi uhlaka ngamakhono e-Databricks (ukubhalisa, ukulandelela, ukuhlola, ukuthunyelwa, nokubonakala) ngaphandle kokungqubuzana.

Izinzuzo ezibalulekile: usekelo lwendabuko lwe izinhlelo ama-ejenti amaningi, phuma kumodi Ukusakaza, umlando ophelele wemilayezo nezingcingo zamathuluzi, ukuqinisekiswa kwezingcingo zamathuluzi kanye amathuluzi ahlala isikhathi eside. Ngaphezu kwalokho, inikeza izixhumanisi ezifakiwe Ku-Python, inference yesiginesha ezenzakalelayo, ukulandelela okuzenzakalelayo kwe predict y predict_stream kanye namathebula okukhomba athuthukisiwe ngokusebenzisa I-AI Gateway.

Izidingo zokufaka zale ndlela: databricks-agents 1.2.0 +, mlflow 3.1.3 + y I-Python 3.10+ (usebenzisa i-Serverless noma i-Databricks Runtime 13.3 LTS+). I-Databricks futhi incoma amaphakheji alandelayo: Ibhuloho le-AI ngokuhlanganiswa okujwayelekile.

Ukuhlanganiswa okuwusizo: databricks-openai (OpenAI), databricks-langchain (LangChain/LangGraph), databricks-dspy (DSPy) kanye databricks-ai-bridge (Ama-ejenti kuPython ehlanzekile). Ngalezi zingcezu ungakha noma yini kusuka kuma-ejenti alula kuya amapayipi ayinkimbinkimbi ngokululama kanye nekholi yethuluzi.

Uma usuvele unomenzeli owakhelwe ngaphakathi, usonge ngesigaba esizuza kuso mlflow.pyfunc.ResponsesAgent. Sebenzisa predict ukuguqula okokufaka nokuphumayo kube ResponsesAgent schemaUkuze usakaze, thembela ku predict_stream, likhipha i-deltas kanye nomcimbi wokugcina response.output_item.done Sicela wengeze impendulo ephelele.

Mayelana namaphutha okusakaza-bukhoma, i-Mosaic AI isakaza ukwehluleka kanye nethokheni yokugcina ngaphakathi databricks_output.errorIklayenti lakho kufanele liyithwebule futhi liyibonise ngendlela efanele, ithuthukise i- isipiliyoni se-depuration.

Amasistimu okubuyisela, okokufaka ngokwezifiso, nokulandeleka

Ama-ejenti amaningi ayasebenzisa ama-recyclers kuma-RAG ngaphezulu kwama-vector indices. Uma ububanzi bokubuyisa kwakho kuveza isikimu esihlukile, ungasisebenzisa imephu mlflow.models.set_retriever_schemaememezela primary_key, text_column, doc_uri namanye amakholomu. Ngakho, i-AI Playground izovela Izixhumanisi emithonjeni isitholakele futhi ukuhlolwa kuzosebenza amagoli abalulekile ngokuzenzakalelayo.

Ezimeni lapho umenzeli wakho edinga ulwazi olwengeziwe (isibonelo, uhlobo_lweklayenti o iseshini_id), sebenzisa custom_inputs futhi ibuyisela imethadatha ngaphakathi custom_outputs lokho akufanele kungcolise umlando wengxoxo. Kokubili i-AI Playground nohlelo lokusebenza lokubuyekeza likuvumela ukuthi ungene okokufaka_kwezifiso kusuka ku-UI.

Ukuthunyelwa nge-deploy(), izimfuneko nezenzo ezizenzakalelayo

Ukuze ushicilele umenzeli wakho ku-Mosaic AI Model Serving, sebenzisa umsebenzi deploy() yePython API. Ku-MLflow 3, faka mlflow >= 3.1.3 y databricks-agents >= 1.1.0 (uma uphakela ngaphandle kwezincwadi zamanothi). Ku-MLflow 2.x, ubuncane bushintsha (isibonelo, mlflow >= 2.13.1 y databricks-agents >= 0.12.0Qiniseka Bhalisa i-ejenti kukhathalogi ye-Unity ngaphambi kokuthunyelwa.

Shayela ku deploy() yakha iphoyinti lokugcina nge ukulinganisa okuzenzakalelayo, lungiselela i- ukuqinisekiswa okuvikelekile kwezinsiza zeDathabricks (izinkomba zevekhtha, imisebenzi yekhathalogi) -ihlola izimvume zomnikazi-, futhi ivule buyekeza isicelo ngokubambisana nababambe iqhaza.

Iphinde ivumele ukuqapha kwesikhathi sangempela ku-MLflow (ngokulandelwa kokuhlolwa okusebenzayo kanye nasemathebuleni okuqondiswa). Ku-MLflow 3, ukuqapha kokukhiqiza ngokuhlolwa kwekhwalithi okuzenzakalelayo kungalungiswa; izicelo ze-REST kanye namazwana angeniswe kumathebuleni okuqondiswa (kusetshenziswa ama-API empendulo yesimanje).

Ukuze wenze ngendlela oyifisayo, dlulisela izimpikiswano ku deploy() njengoba scale_to_zero_enabled=Trueukunciphisa izindleko ngemva kwesikhathi sokuphumula ngezindleko zokubambezeleka okuphezulu esicelweni sokuqala. Ukwengeza, unezinsiza zokuthi list, thola futhi ususe ukuthunyelwa nge-Agent API.

Imikhuba emihle yezwe, ukumisa, namaphethini okusebenzisa

I-Model Serving isatshalaliswa: gwema izinqolobane zendawo futhi ithatha ngokuthi izimpendulo ezihlukene zingaphatha izikhathi ezihlukene engxoxweni. Yakha kabusha umbuso kusukela ku- isichazamazwi se ResponsesAgentRequest lilinye predict futhi uklame isimo sibe intambo-ephephile.

Iqalisa umbuso ngaphakathi predict futhi hhayi kumakhi we-ejenti, njengoba isifaniso esisodwa singasingatha izingxoxo zabasebenzisi abahlukene. Lokhu kuvimbela ukuvuza kolwazi kanye nokungqubuzana kwezimali ezindaweni ezinomphumela ophezulu.

Hlela ukuziphatha komenzeli nge ModelConfig (isichazamazwi noma i-YAML). Chaza, isibonelo, model_serving_endpointImingcele ye-LLM nezifanekiso ezisheshayo; layisha lokho kulungiselelwa ekuthuthukisweni futhi ukukhuthaze ekukhiqizeni ngaphandle kokuthinta ikhodi.

Gwema ukwethula ama-event loops noma ozimele abangavumelani kuseva ye-ejenti: I-Databricks iphatha i-concurrency ngaphakathi . Sebenzisa ikhodi evumelanisiwe noma ama-callback; ngaphandle kwalokho, ungase uhlangane namaphutha anjengokuthi "Le mcimbi womcimbi usuvele usebenza" kanye nokuziphatha okungajwayelekile.

Izitini zomenzeli: "akukho khodi" okuzenzakalelayo ukusheshisa inani

I-Agent Bricks, eyethulwa ngo-2025, inikeza umsebenzi wokudala i-ejenti ongenamthungo ongadingi ukufaka ikhodi . Ufaka umsebenzi ngolimi lwemvelo, uxhumanise idatha yakho, bese ipulatifomu ikhetha amamodeli, ikhiqize idatha yokwenziwa , ihlole, futhi ithuthukise inqubo kuze kube yilapho iletha i-ejenti elungele ukukhiqiza enokuphathwa kwe -Unity Catalog.

Phakathi kwezinzuzo zayo kukhona ukuzenzekela kusukela ekuqaleni kuya ekugcineni , ukukwazi ukukhulisa i-Databricks — kanye nama-endpoints angafinyelela ku-zero ngemva kwezinsuku ezi-3 zokungasebenzi —, ukusetshenziswa kanye nokuhlaziywa kwezindleko ku-Lakehouse kanye nezinhlobo zama-ejenti alungele ukusetshenziswa (ukukhishwa kolwazi, i-LLM yangokwezifiso , umsizi wolwazi kanye nomphathi wama-ejenti amaningi wokuhlelwa okuyinkimbinkimbi).

Imikhawulo ifaka phakathi iqiniso lokuthi ikhodi ekhiqiziwe ayilandeki okwamanje , okunciphisa ukwenza ngokwezifiso okujulile, futhi ikhathalogi yesifanekiso iqondiswe ezinhlanganweni esezihlanganisiwe kakade neLakehouse kanye ne-Unity Catalog. Abanye abasebenzisi bathola nokuthi izilinganiso zezindleko azinawo ama-granularity .

Izibonelo, ama-ejenti amaningi kanye nama-ejenti asemthethweni

Kumadokhumenti asemthethweni, uzothola ama-notebook ane- OpenAI (izingxoxo ezilula kanye nezingcingo zamathuluzi, kokubili amamodeli aphethwe yi-Databricks kanye ne-OpenAI), i-LangGraph (isicelo samathuluzi), kanye ne -DSPy (izingcingo zamathuluzi ngesikhathi esisodwa). Lezi ziyisisekelo esihle kakhulu sokwandisa ukusebenza ngamathuluzi enziwe ngokwezifiso kanye nokuhlolwa kwekhwalithi.

Ukuze uhlele ochwepheshe abaningi, ungasetha izinhlelo ezisebenzisanayo usebenzisa amathuluzi afana ne-Genie. Futhi uma unentshisekelo kwimemori yengxoxo, kukhona ukwesekwa kwama-ejenti azimele avumela izindawo zokuhlola kanye nokuqhubeka kwengxoxo.

Kumadokhumenti asemthethweni, uzothola ama-notebook ane- OpenAI (izingxoxo ezilula kanye nezingcingo zamathuluzi, kokubili amamodeli aphethwe yi-Databricks kanye ne-OpenAI), i-LangGraph (isicelo samathuluzi), kanye ne -DSPy (izingcingo zamathuluzi ngesikhathi esisodwa). Lezi ziyisisekelo esihle kakhulu sokwandisa ukusebenza ngamathuluzi enziwe ngokwezifiso kanye nokuhlolwa kwekhwalithi.

Izinhlelo zokusebenza ezikuDathabricks: I-UI ene-React ne-backend ku-FastAPI

Izinhlelo zokusebenza ze -Databricks zinikeza ipulatifomu eqinile yokuletha izixhumi ezicebile ngaphandle kokusetha ingqalasizinda yangaphandle. Zihlanganiswa ngokwendabuko ne-Databricks SQL, i-Unity Catalog , i-Model Serving, kanye ne-Jobs, futhi zizuza njengezilawuli zokuphepha nokuhambisana ne-platform.

Iphethini eliwusizo kakhulu ukudala i-frontend nge -React (esekela ukusakaza, ukuhanjiswa kwemiyalezo, kanye ne-Markdown UX) kanye ne-backend ene -FastAPI esebenza njengebhuloho eliya endaweni yokugcina ye-ejenti. I-backend idlulisela isicelo somsebenzisi kusevisi yemodeli futhi ibuyisela impendulo ehlelekile kuklayenti.

Ku-FastAPI, ukuhleleka komhlangano wezicelo ezingaphansi kubalulekile: app.mount("/api", api_app) iqondisa imizila ye-API (isibonelo, /api/chat) futhi app.mount("/", ui_app) Inikeza amafayela amile asuka ku-React build kumkhombandlela wempande. Lolu hlelo luyaqhubeka ukucaca kwemizila futhi ihlukanise i-UI ezindaweni zokugcina.

Lungiselela okuguquguqukayo kwendawo (isb., SERVING_ENDPOINT_NAME noma i-FQN yemodeli yomenzeli ku-Unity Catalogue) kusukela kuncazelo yohlelo lokusebenza kokuthi Databricks. Hlanganisa i-frontend (npm run build), ivumelanisa amafayela, isho okuncikile (requirements.txt) futhi iqala nge I-Gunicorn + Uvicorn abasebenzi be-ASGI.

Icala langempela: i-chatbot yemisebenzi yokukhiqiza

Ezimweni zokuphathwa kwemisebenzi yokukhiqiza (i-MOM), i-chatbot ye-Databricks ingasiza abaphathi bokukhiqiza ngokuhlaziywa kwezinkinga , ukulandelela impahla, kanye nokukhiqizwa kwemibiko. Imisebenzi ivezwa njengamathuluzi ku-Unity Catalog (nge-SQL noma i-Python), futhi i-ejenti iyayibiza ngokusekelwe kumongo.

Izibonelo: umsebenzi identify_bottleneck_station Isebenzisa imibuzo ku-Databricks SQL ukuthola isigaba esibangela ukubambezeleka okukhulu; omunye, check_inventory_levelsIbuyisela isitoko sesikhathi sangempela. I-ejenti ingakwazi futhi thumela ama-imeyili, khiqiza imibiko futhi wenze ukuhlaziya imizwa, konke kuhlelwe ngamathuluzi abizayo.

Ukusebenzisa i-Model Serving endpoint kusebenza njenge- backend ejwayelekile . Uhlelo lokusebenza lwe-React lusebenzisa le endpoint, lubonisa izimpendulo zesikhathi sangempela ngolwazi lomsebenzisi oluthuthukisiwe kanye nokusekela ukusakazwa kwamathokheni uma kunikwe amandla ku-ResponsesAgent yakho.

Ukufuduka ku-ChatCompletions nokuhambisana

Uma umenzeli wakho wefa esebenzisa i-OpenAI's ChatCompletions API , ungathuthela ku- ResponsesAgent ngesitsha esihumusha imiyalezo engenayo/ephumayo, futhi, ngokuzithandela, sifaka amamephu okusakaza ama-delta ku- ResponsesAgentStreamEvent . Lokhu kuthuthukisa i-telemetry kanye nokubonakala ngaphandle kokubhala kabusha i-logic eyinhloko.

Ukuze ubhale futhi uphathe ukuthunyelwa, i-Agent API ikuvumela ukuthi ubhale ukuthunyelwa okukhona, uthole ukuthunyelwa okuthile (ngegama nenguqulo), bese ususa ama-endpoints ongasawadingi. Khumbula ukuphatha iziqinisekiso zezinsiza zangaphandle usebenzisa iziguquguquko zemvelo nezimfihlo.

I-Databricks inikeza indlela ephelele: kusukela ekucabangeni nge -AI Playground noma ukuzenzekela nge-Agent Bricks, ngokuthuthukiswa kwekhodi (i-Mosaic AI, i-MLflow, i-LangGraph, i-DSPy), kuya ekupakisheni, ukubhaliswa ku- Unity Catalog , ukuthunyelwa nge -deploy() , kanye nokuqapha ukukhiqizwa. Ngemikhuba emihle efana nokusayina kwemodeli efanele , isimo esiphephile ezindaweni ezisatshalalisiwe, kanye nokucushwa okunemingcele, ungaletha ama-ejenti athembekile, alawulwayo alungele ukuhlanganiswa kuzinhlelo zokusebenza zewebhu epulatifomu.

I-athikili ehlobene:

UClaude 4: I-Anthropic icabanga kabusha ubuhlakani bokwenziwa ngamamodeli athuthukile ohlelo nama-ejenti azimele