Gwaro remugadziri wemapatani ema "multi-agent" neADK

Home » Python » Gwaro remugadziri wemapatani ema "multi-agent" neADK

Mapurogiramu eAgent ari kukurumidza kukura kupfuura maitiro ekare e "single mega-prompt", uye vagadziri vanoda modhi yakasimba yepfungwa yekugadzira ma "agents" akawanda pasina kupinda mukusagadzikana. Kit yeGoogle yeAgent Development Kit (ADK) yakagadzirirwa izvi chaizvo: inokutendera kuti ugadzire masisitimu ema "multi-agent" akavimbika, maturusi akabatana uye ndangariro, uye kuisa zvese paGoogle Cloud nekutarisa kugadzirwa, kuchengetedzeka uye kudzora mitengo.

Gwaro iri rinokuratidzai nzira huru dzekushandisa ma "multi-agent" anotsigirwa neADK - kubva ku "simple parent/child hierarchies" kusvika ku "Sequential", "Loop" uye "Parallel workflow agents" - uye rinoratidza kuti dzinoenderana sei ne "wide reference architecture" paGoogle Cloud. Tichakurukurawo mamiriro ekugovana kwemisangano, nzira dzekupa mabasa, mapurani akajairika evamiririri vakawanda, uye zvinhu zvinoshanda pakuisa, kuchengetedza uye kushandisa masisitimu aya munzvimbo chaiyo.

Sei paine masisitimu akawanda eADK?

Kana application ikashandiswa nechinhu chimwe chete, zvinokurumidza kuoma kufunga nezvazvo, kuyedza uye kushanduka. Zvikumbiro zvikuru hazvitani kusimba, zvakaoma kugadzirisa uye zvinorwadza kuchengetedza sezvo zvinodiwa zvichikura. ADK inokusundira kuti ugadzire hurongwa Sisitimu yeVamiririri Vakawanda (MAS) uko mumiriri wega wega ane basa rake rakatarisana, uye kuronga kwacho kwakajeka.

Kugadzira app yako sevashandi vakawanda vanobatsirana kunoita kuti ive nehunyanzvi, hunyanzvi uye ishandiswezve. Unogona kuva nemumiriri wekutsvaga ruzivo, mutsoropodzi, munyori wemafaira, router, mumiriri wekuwana ruzivo, uye wozvishandisazve mumapurojekiti kana mabasa ekushanda pachinzvimbo chekuisa pfungwa imwechete muchirongwa chimwe chete chikuru.

ADK inokupa zvivharo zvekuvaka zvakasimba kuti uite izvi: ma agents anotarisa LLM, ma workflow agents (Sequential, Parallel, Loop) uye ma custom agents anosanganisira non-LLM logic. Vose vanogara nhaka kubva kune mumwe munhu BaseAgent, saka vanobatanidza modhi imwechete ye orchestration, logging, state handling uye tool system.

Sezvo masisitimu achikura, modhi iyi yekunyora inokura zviri nani pane kodhi yekurongeka kwemamiriro ezvinhu kana cheni dzakadzika dzema "function calls" dzakatenderedza modhi imwe chete. Unochengetedza mutoro wepfungwa uchikwanisika: munhu wese ane basa rake rakajeka uye ane mukana wekutaurirana nevamwe zvakanaka.

ADK primitives dzevanogadzira zvinyorwa

ADK inoburitsa zvinhu zvidiki zvepakutanga zvaunogona kusanganisa kuti uratidze magadzirirwo anoshamisa ezvigadzirwa zvakawanda. Kunzwisisa pfungwa idzi huru kunoita kuti zvive nyore kufunga nezvemaitiro epamusoro gare gare.

Chekutanga chekutanga ihurongwa hwemumiriri: mumiriri wese anogona kuzivisa runyorwa rwe sub_agents, kugadzira muti une muti mumwe chete root_agent pamusoro. Kana uchipa vana sub_agents, ADK inozviwaya otomatiki parent_agent zvinoreva uye zvinosimbisa kuti muenzaniso wakapihwa une mubereki mumwe chete (zvikasadaro a ValueError inosimudzwa).

Urongwa uhwu hunopfuura kushongedza chete: hunotsanangura kuti ndeapi maagent anotenderwa kupa mabasa api, uye ndiyo nzvimbo inoshandiswa nevamiriri vebasa uye kutamiswa kunotungamirirwa neLLM. Kubva kune chero mumiririri, unogona kuyambuka uchienda kumusoro kuburikidza agent.parent_agent kana kuwana vana vane agent.find_agent(name), iyo inobatsira zvikuru pakugadzirisa matambudziko.

Pamusoro pevamiririri veLLM, ADK inosuma vamiririri vebasa vakazvitsaurira - SequentialAgent, ParallelAgent uye LoopAgent - basa ravo harisi re "kufunga" asi kuronga vamiririri vechidiki. Dzese dzine interface imwe chete asi dzinoshandisa nzira dzakasiyana dzekuita: kumhanya zvakarongwa, kuburitsa mafafitera zvakafanana, kana kuita kudzokorora mumutsara nemirairo yakajeka yekugumisa.

Chinhu chechitatu chakakosha chepakutanga inyaya yekutaurirana, yakatarisana neyakagovaniswa Session uye iyo state duramazwi. Mamiriro echikamu anoita sebhodhi chena rakajairika: chero mumiririri kana chishandiso chinogona kunyora mhedzisiro yepakati kana mireza, uye vamwe vamiririri vari mukukumbira kumwe chete vanogona kuaverenga, kazhinji kuburikidza nekiyi-templating mukati memirairo yavo (semuenzaniso). {PLOT_OUTLINE?}).

Matauriro anoita vamiririri kune mumwe nemumwe muADK

ADK inotsigira nzira nhatu dzekutaurirana dzinoenderana pakati pevamiririri: mamiriro echikamu chakagovaniswa, kutamiswa kunotungamirirwa neLLM, uye kukumbirwa pachena kuburikidza ne AgentTool. Kusarudza chakakodzera pakutaurirana kwega kwega kunoita kuti system yako ive pachena uye inonyatso fungidzirwa.

Mamiriro echikamu chakagovaniswa (session.state) ndiyo nzira iri nyore uye inonyanya kushandiswa. Mukati mekukumbira kumwe chete, vamiririri vese vanoona zvakafanana Session chinhu kuburikidza ne InvocationContextChishandiso kana callback zvinogona kuita context.state = value, uye mumiririri wepashure anogona kuitora ne context.state.get("data_key")Kune vamiririri veLLM, gadziriro output_key vanongoramba vachipa mhinduro yavo yekupedzisira pasi pekiyi iyoyo.

Kutumirwa kwemashoko kunotungamirirwa neLLM, dzimwe nguva kunonzi "kutamiswa kwevamiririri", kunoita kuti LLM isarudze nguva yekupa kune mumwe mumiririri zvichibva pamirairo netsananguro dzevamiririri. Mukati, modhi inoburitsa runhare rwakakosha rwebasa senge transfer_to_agent(agent_name="screenwriter")ADK's AutoFlow inodzivisa kufona uku uye inodzosera mashandiro kune mumiririri akasarudzwa mukati mechikamu chinobvumidzwa (mubereki, vana, vanun'una zvichienderana nechimiro).

Kunyengetera pachena ne AgentTool inokupa nzira inodzorwa zvakanyanya, yakaita seyekushanda, yekufonera mumiririri mumwe kubva kune mumwe. Unoputira muenzaniso wemumiriri anotarirwa mukati AgentTool, iwedzere kune yemunhu ari kufona tools runyorwa, uye LLM inogona kusarudza chishandiso ichocho sechero rimwe basa. Kana yakumbirwa, AgentTool.run_async inoita kuti subagent ishande, inobatanidza mamiriro ezvinhu uye maartifacts, uye inodzosera mhinduro yesubagent semhedzisiro yechishandiso.

Aya machaneli matatu anofukidza zvinodiwa nevashandisi vakawanda: data risingachinjiki richipfuura nepamamiriro ezvinhu, nzira inochinjika kuburikidza nekutumira data, uye mafoni akabatana achishandisa maturusi. Mumagadzirirwo akaomarara, unowanzo sanganisa mukati memuti mumwe chete: router inotamisirwa kuvana, nyanzvi dzinoshandisa mamiriro ekutaurirana, uye mumiririri mumwe kana vaviri vanowanikwa sezvishandiso zvekutumira zvinhu pachivande.

Zvivako: Vamiririri veLLM, vamiririri vebasa uye vamiririri vezvinhu zvakasiyana

Mazhinji ma MAS topologies muADK musanganiswa wemhando nhatu dze agents: LLM-based, workflow uye custom agents. Chikamu chimwe nechimwe chinogadzirisa chikamu chakasiyana chedambudziko rehurongwa hwekugadzira.

Vamiririri veLLM vanoputira modhi yemutauro mukuru uye maturusi esarudzo, callbacks uye kuburitsa zvinobuda mukati BaseAgent. Funga nezvazvo sezvinhu zvako "zvekufunga": zvinotsanangura zvinopihwa nemushandisi, maturusi ekufona, kugadzirisa mamiriro uye zvinopindura mushandisi kana kupa kune mumwe mumiririri.

Vamiririri vebasa vanoshanda sevatungamiriri kwete sevashandi: havazvifungi ivo pachavo, asi vanodzora kurongeka, kufanana uye kudzokorora kwekuita kwevashandisi. SequentialAgent inomhanya nevana vayo mumwe mushure memumwe ichigovana zvakafanana InvocationContext, ParallelAgent vanopararira mumapazi akasiyana-siyana ane nyika imwe chete asi ane matavi akasiyana enhoroondo, uye LoopAgent inoita kutevedzana kakawanda kusvika mamiriro ekumira asangana.

Vamiririri vetsika vanowedzera BaseAgent nepfungwa dzisina kurongwa dzeLLM asi nzira dzekugadzira dzakavakirwa mukati hadzina kukwana. Semuenzaniso, unogona kushandisa custom scheduler iyo inoshandisa vamiririri zvichienderana nemamiriro ezvinhu, kana kubatanidza injini yemitemo yebhizinesi iyo inosarudza kuti ndeipi sub-flow inofanira kushanda zvichienderana nemitemo.

Musanganiswa uyu we generic orchestration primitives uye pluggable logic ndizvo zvinoita kuti ADK ikodzere mabasa akakomba ebhizinesi, kwete ma demo chete. Unogona kutanga nevamiririri vemazuva ano vemabasa ekushanda, uye kana zvinodiwa zvave zvisingawanzoitiki ndipo paunosvika pazviri. CustomAgent.

Mamiriro echikamu uye maitiro ekurangarira

Mamiriro echikamu muADK anotsigira ndangariro yekukurukurirana kwenguva pfupi uye data rakarongeka rinopfuudzwa pakati pevamiririri. Hurukuro yega yega inoshandisa Session chinhu chakabata nhoroondo yemeseji uye chinochinjika state duramazwi rinowanikwa kune vese vamiririri vari muchirongwa ichocho.

Kunyorera ku state kunowanzoitwa mukati mematurusi kana ma callbacks, uchishandisa ToolContext or CallbackContext chinhu. Semuenzaniso, chishandiso chakaita se save_attractions_to_state(tool_context, attractions: List) inogona kubatanidza zvinhu zvitsva zvinokwezva neizvo zvakatochengetwa pasi state, achidzorera meseji iri nyore kumumiriri uku ADK ichiramba ichiita delta yemamiriro ezvinhu muchikamu.

Kuverenga kubva pamamiriro ezvinhu kunogadzirwa nenzira ye ergonomic kuburikidza nema templates akakosha akaiswa mumirairo. Kana murairo uine {my_key?}, ADK ichabaya state kana iripo; chiratidzo chemubvunzo chinoita kuti ive sarudzo kuitira kuti mumiririri asakundikana kana kiyi isipo. Izvi zvakakosha mumabasa ekushanda akadai sekuti "tsvagurudzo → nyora → ongorora" apo danho rega rega rinoverenga zvakachengetwa nedanho rekare.

Kuti muyeuke kukurukurirana, pfungwa huru ndeyekushandisazve zvakafanana Session kune mamwe mameseji evashandisi pane kugadzira imwe itsva nguva dzese. Nemusangano wakagovaniswa, mumiririri anoona nguva dzakatarwa uye anogona kugadzirisa mibvunzo yekutevera, kugadzirisa uye kuronga matanho akawanda. Kana ukaita netsaona musangano mutsva parwendo rumwe chete, mumiririri anoita sekunge ane dambudziko rekusaziva: haakwanise kubatanidza nguva dzekutevera nemamiriro ezvinhu apfuura.

Hurumende inewo basa guru mumabasa ekubatsira vanhu kuita mabasa akadai se LoopAgent, iyo inoshandisa makiyi anoenderera mberi akadai semakaunzi, runyorwa rwemhinduro kana mireza kuti isarudze kana paine zvimwe zvinodiwa kuti zvidzokororwe. Mumiririri wevatsoropodzi anogona kuwedzera mashoko ake mu CRITICAL_FEEDBACK pane imwe neimwe nzira, apo murongi kana mugadziri wezvinhu anoverenga kiyi iyoyo yekuvandudza hurongwa muchikamu chinotevera.

SequentialAgent: mafambisirwo ebasa akarongwa akajeka

SequentialAgent ndiyo nzira yaunofanira kushandisa kana uine matanho akatevedzana anofanira kuitika nenzira yakatarwa. Funga nezvemapombi akaita sekuti “ongorora chikumbiro → tsvakiridzo → draft → chengetedza kuti ufaire” kana kuti “ziva kwaunoenda → ronga nzira → bhuka kutakurwa”.

MuADK, a SequentialAgent ane runyorwa rwe sub_agents uye anovamhanyisa mumwe nemumwe, achipfuura zvakafanana InvocationContext kuburikidza necheni yese. Nokuti the Session uye state zvinogoverwa, unogona kuita kuti mumiririri wekutanga achengete mhedzisiro yacho pasi pe output_key="destination" uye mumiriri anotevera akaiverenga kuburikidza {destination} mumirayiridzo yayo isina kana glue code.

Muenzaniso wekare ijenareta yefirimu: mushandisi wemidzi anotaura nemushandisi, wobva washanda nemaoko kune SequentialAgent inodana muongorori, wozobva wanyora zvinyorwa, wozobva wanyora mafaira. Mushandisi anoona mhedzisiro yekupedzisira chete, asi girafu yechiitiko muADK Web inoratidza muti wemukati: greeter → film_concept_team → .

Zvichienzaniswa nekurongeka kwemaoko nemaoko ne explicit if/elif mabhuroko uye mafoni ebasa, SequentialAgent inochengetedza kudzora kwekufamba kwemvura uye inoderedza boilerplate. Unozivisa kutevedzana kamwe chete woibata semumiriri mumwe chete anogona kufonwa murunner yako kana UI, uku uchishandisa mamiriro emusangano kupfuudza data pakati pematanho.

Maitiro ekushanda anoenderana zvakanakawo nemamwe ma workflow agents: unogona kuisa loop kana parallel fan out seimwe yematanho ari mucheni refu. Izvi ndizvo zvinovaka mafambiro epamusoro akadai sekuti "kudzokorora mhando yenyaya, wobva watanga box-office uye ongororo yevatambi, wobva wanyora mushumo wakabatanidzwa".

LoopAgent: kugadzirisa mafambiro enyaya uye makamuri evanyori

LoopAgent yakagadzirirwa mabasa anobatsirwa nekupfuura kwakawanda pamusoro pebasa kusvika muganho wemhando wasvika. Pane kushandisa nzira imwe chete yekuti "gadzira kamwe chete wotarisira zvakanakisisa", unogona kunyora maitiro ekukurudzira, kutsoropodza uye kugadzirisa.

Maitiro akajairika ekuita ongororo anosanganisira vamiririri vakaita semuongorori, jenareta (semuenzaniso munyori wechinyorwa) uye mutsoropodzi vanoshanda pamwe chete kwenguva dzakawanda. Pakudzokorora kwega kwega, muongorori anogona kugadzirisa ruzivo rwekare, munyori wenyaya anogadzirisa hurongwa kana chirongwa, uye mutsoropodzi anochiongorora achienzanisa nemirairo yakajeka, achisarudza kana zvimwe zvichidikanwa.

Zvisungo zvinomira pasi pemamiriro maviri: kusvika max_iterations kana chiratidzo chemubatsiri chinoratidza kuti basa rapera. ADK inoburitsa chishandiso chakavakirwa mukati chakaita se exit_loop izvo mutsoropodzi anogona kudana kana chirongwa, purani kana dhizaini zvapfuura runyorwa rwayo rwemukati. LoopAgent inoremekedzawo escalate=True mureza mukati Event zviito, zvichikupa imwe nzira yekubuda nekukurumidza.

Mamiriro echikamu chinogara chiripo akakosha pano: vamiririri vanoverenga makiyi akadai se PLOT_OUTLINE, research or CRITICAL_FEEDBACK uye nyora shanduro dzakavandudzwa kana mamwe mashoko pane imwe neimwe pass. Maitiro aya anotevedzera "nzvimbo yevanyori" apo nyanzvi dzinofunga, dzinotsoropodza uye dzinoongorora kusvika mumwe munhu ataura kuti basa ragadzirira.

Nekubatanidza LoopAgent ne SequentialAgent, unogona kuisa denderedzwa rese rekudzokorora sedanho rimwe chete mukushanda kukuru kubva pakutanga kusvika pakupedzisira. Semuyenzaniso, writers_room (LoopAgent) inogona kutanga yatanga kugadzira hurongwa hwakasimba hwenyaya, mushure mezvo a file_writer mumiririri anochengetedza mhinduro yacho uye anonamira mimwe mishumo.

ParallelAgent: buritsai uye unganidzai mabasa ega

ParallelAgent inoshandisa maitiro ekare ekuti "fan-out / gather" emabasa akazvimiririra asi achigovana mamiriro ezvinhu. Pane kumhanyisa matanho ekutsvagisa eN akatevedzana, unoamhanyisa ese panguva imwe chete womirira kuti ese apedzwe, wobva waunganidza zvabuda.

Mukati, ParallelAgent inopa mumiririri wega wega chinhu chakasiyana InvocationContext.branch – kufanana ParentBranch.ChildName - vachiri kugovera zvakafanana session.state. Izvi zvinoreva kuti vese vanogona kuverenga mamiriro ekutanga se PLOT_OUTLINE, asi inofanira kunyora zvinobuda kumakiyi ehurumende akasiyana (semuenzaniso box_office_report, casting_report) kudzivirira kusawirirana.

Muenzaniso unowanzoonekwa ndewekuti "chikwata chisati chagadzirwa" chefirimu: mumwe mumiririri anofungidzira mukana wekuti vatambi vatore mafirimu zvichibva pamafirimu akafanana, mumwe anopa sarudzo dzekusarudza vatambi, ese ari maviri achienderera mberi. Inotevera file_writer wobva wanyora mushumo uchishandisa matemplate akakosha emhedzisiro yega yega wobva waenderera mberi nawo padhisiki.

Kufamba kwebasa kwakafanana kunoderedza zvakanyanya kunonoka kwemibvunzo yakawanda uye mune escenarios de real-time data analysis: kana uchida mazano emuseum, sarudzo dzemakonzati uye mazano eresitorendi ekupera kwevhiki, kushandisa vamiririri vatatu vane hunyanzvi panguva imwe chete zvinokurumidza kupfuura kuvabvunza zvakatevedzana. Mushure mekubuda kwefeni, mushandisi we synthesis anoverenga zvese zvabuda kubva kumamiriro ezvinhu uye oburitsa mhinduro yakabatana kumushandisi.

Matanho akafanana anowanzo shandiswa mukati me SequentialAgent iyo inotanga yagadzirira mamiriro ezvinhu, yozotanga ParallelAgent, wobva waenderera mberi nekuunganidzwa kwezvinyorwa uye kushuma. Maitiro aya ari nyore kuona uye kushandisazve kana wava kunyatsogadzikana nevamiririri veADK vekushanda.

Maitiro ekutamba ane ADK primitives

Kana wangonzwisisa hierarchy, workflow agents uye state, unogona kushandisa akawanda classic multi-agent patterns zvakananga muADK. Mapatani aya haasi ekare asi magadzirirwo akavakwa nemabhuroko akafanana.

Maitiro emurongi/muparidzi anoshandisa mumiririri weLLM se "router" yemibvunzo yemushandisi, achitsigirwa nema subagents akawanda ane hunyanzvi. Murongi anoverenga chikumbiro, obva aendesa kutonga kumumiriri ari pasi pake kuburikidza nekutumira vanhu kunotungamirwa neLLM kana kufonera nyanzvi zvakananga vachishandisa AgentToolMienzaniso inowanzoonekwa nevaya vanoda zvekudya, vekufambisa kana vanotungamira mafambiro evhiki.

Maitiro emapaipi akatevedzana ingori SequentialAgent vana vavo vanoita danho rakanyatsorongwa rekuita. Kuyerera kwejenareta nekuongorora imhando yekare: mumiririri wekutanga anonyora draft oichengeta pasi pe output_key, mumiririri wechipiri anoongorora ochengeta mhinduro, uye pamwe mumiririri wechitatu anogadzirisa mhedzisiro zvichibva pane mhinduro iyoyo.

Maitiro ekubatanidza fan-out/gather anoratidzwa se ParallelAgent yakavakirwa mukati mekushanda kwakatevedzana. Vana vanonyora mhinduro mumakiyi akasiyana ehurumende; mumwe mushandisi we synth anozoverenga mhinduro dzacho obva aratidza mhinduro dzakabatana.

Kupatsanurwa kwemabasa ehurongwa hwepamusoro kunobuda pachena kubva pamuti wemubereki/mwana. Vamiririri vepamusoro vanopatsanura zvinangwa kuita zvidiki-diki vozvipa vana (kuburikidza nekugovera kana zvishandiso), uye mhedzisiro yacho inodzoka zvakare. Izvi zvinonyanya kubatsira muvabatsiri vekutsvagisa, vagadziri vezvinhu kana masisitimu evapi vemazano emari uko subdomain yega yega ine mumiriri wayo chaiye.

Kugadziriswazve kwekudzokorora ne LoopAgent inoita kuti loop yejenareta-mutsoropodzi ive pateni inogona kushandiswazve. Chinotenderera chinoshandisa varongi, vatsoropodzi, uye vagadziri vezvinhu kakawanda, vachishandisa makiyi ehurumende kuti varambe vachitevera hurongwa hwazvino uye mhinduro dzinoenderana, zvichimira kana paine chiyero chemhando yepamusoro kana muganho wekudzokorora.

Magadzirirwo ezvinyorwa zvema "multi-agent systems" paGoogle Cloud

Kunze kwekufunga nezvemumiriri, uchiri kuda kumhanyisa sisitimu yako kune imwe nzvimbo, uye Google Cloud inopa dhizaini yakanyatsotsanangurwa yekushandiswa kwema agent akawanda emhando yekugadzira. Padanho repamusoro, mhinduro iyi inosanganisa frontend, agent runtimes, Vertex AI models, security services uye optional tool frameworks seMCP.

Kugadzirisa kwakajairika kunotanga ne frontend - kazhinji chat interface - ichishanda pa Cloud Run. Vashandisi vanotaura neUI iyi, iyo inotumira zvikumbiro kumumiriri wekubatanidza anoratidzwa sebasa. Murongi uyu anobva asarudza pakati pemabasa akasiyana emumiriri zvichienderana nechinangwa chemushandisi, kusanganisira nzira dzesarudzo dzevanhu-mu-the-loop uko vanhu vanogona kusimbisa kana kudzima sarudzo dzemumiririri.

MaAgents pachawo anogona kushanda munzvimbo dzakasiyana siyana: Cloud Run services, Google Kubernetes Engine (GKE) kana Vertex AI Agent Engine. ADK inoshandisa sarudzo idzi, ichibvisa zvimwe zveruzivo rwenguva yekushandisa kuitira kuti mugadziri atarisire pane hunyanzvi hwemumiriri pane kuisa mapombi ezvivakwa.

Mafoni ese emumiririri anoshandisa Vertex AI kana mamwe mamodheru ekushandisa kuti aonekwe, anowanzo shandiswa neModel Armor kuti achenese zvinokurudzirwa uye mhinduro. Model Armor inobatsira kusefa kuedza kwekubaya majekiseni nekukurumidza, kubuda kwedata rakavanzika kana zvinhu zvinokuvadza pamberi kana mushure mekufona kwemodheru, ichishanda senzira yekuchengetedza zvinhu zvinogadzira.

Zvishandiso nemaseva eMCP (Model Context Protocol) zvinopinda mumufananidzo kana vamiririri vachifanira kutaura nehurongwa hwekunze - madhatabhesi, masisitimu efaira, kana maSaaS API - nenzira yakajairika. MCP inotsanangura chibvumirano chakafanana pakati pemumiriri nesevha yezvishandiso, saka mutengi mumwe chete weMCP mumumiriri wako anogona kuwana zvishandiso zvakawanda zvakavakwa nemapoka akasiyana pasina kubatana kwakasimba; izvi zvinosanganisira kufunga nezvazvo. Zvirongwa zvekuchengetedza data y como exponerlos de forma segura.

Kuchengetedzeka uye kutonga kwezvikumbiro zvevamiririri

Masisitimu ehurumende anounza matambudziko ekuchengetedza anopfuura masevhisi madiki echinyakare, nekuti maLLM anogona kunyengerwa kushandisa maturusi zvisirizvo kana kuburitsa data kana ukasangwarira. Maitiro anokurudzirwa neGoogle anoisa zvidziviriro zvekuchengetedza zvine hungwaru nekudzivirira kwakasimba kunobva muLLM; también es clave entender los límites, sesgos y riesgos de los modelos al diseñar estas defensas.

Kutarisirwa kwevanhu kuchiri kukosha: kufambiswa kwemashoko kunofanirwa kusanganisira matanho ekubvumidza apo munhu anogona kumbomira, kuongorora kana kuramba chiito chakapihwa nemumiririri. Izvi zvinogona kuenzaniswa sechishandiso chakazvipira che "kusimbisa munhu" chinogamuchira zvikumbiro kuUI uye chinotangazve kuita kana munhu apindura.

Kudzora mukana wekupinda kwevamiririri kunobatwa kuburikidza neIAM: mumiriri wega wega kana account yebasa inofanirwa kunge iine mvumo shoma inodiwa kuti iite mabasa ayo. Kana mumiririri akapihwa akakanganiswa kana kushandiswa zvisirizvo, kuputika kwemhepo kunogumira nekuti account yake yebasa haigone kuwana zviwanikwa kana maturusi asina hukama.

Kuvharira maturusi kunotungamirirwa nemutemo, kunoitwa nezvinhu zvakaita se SecurityPlugin ndinoiitawo munhu PolicyEngine, inokutendera kuti ubvume kusimbiswa kwemushandisi usati watanga kushandisa mamwe maturusi. Kana mutemo ukaratidza kuti pane nyaya inonetsa, plugin inoibata, yobva yatumira chikumbiro chako chekuti "kumbira kusimbiswa", uye yomirira kuti chikumbiro chako chidzose mutongo, zvichiita kuti munhu akwanise kuita mabasa ane njodzi huru.

Zvinhu zvakajairika zvekuchengetedza zveGoogle Cloud zvinopedzisa mufananidzo uyu: VPC Service Controls yekuderedza njodzi yekubuda kwedata, CMEK yemakiyi ekunyorera anotungamirwa nevatengi, Cloud Armor yekudzivirira WAF neDDoS, IAP kana Identity Platform yekusimbisa vashandisi uye granular IAM yekuwana zviwanikwa. Pakutaurirana kwemumiriri nemumiriri kuburikidza neA2A, TLS 1.2+ uye OAuth based authentication zvinodiwa kana kuti zvinokurudzirwa mukugadzirwa.

Kuvimbika, kucherechedzwa uye kugadzirisa mitengo

Kushandiswa kweMAS mukugadzira kunofanirwa kuve kwakavimbika, kunoonekwa uye kusingadhuri; ADK inobatanidzwa zvakanaka neGoogle Cloud's operational tooling kuti izvozvo zvikwanisike. Unogona kushandisa zvishandiso, masesheni, uye maturusi kuitira kuti magwaro avo nezviratidzo zvionekwe muCloud Logging neCloud Trace.

Kubva pamaonero ekuvimbika, gadzira girafu yako yemumiriri kuti igone kushivirira kutadza muzvikamu zvakasiyana-siyana. Kana zvichibvira, dzivisa uropi humwe chete husingatsiviwi; rega zvinhu zvakazvimiririra zviite mabasa emunharaunda kuitira kuti kudzima kwechinhu chimwe chete kusadzikise kushanda kwese; además, emplea técnicas como balanceo de carga en búsqueda distribuida para distribuir carga y reducir puntos de fallo. Tevedzera kutadza mustaging kusimbisa hunhu hwekudyidzana pasi pekushushikana.

Kune mafoni emhando, Vertex AI inotsigira dynamic shared quotas uye provisioned throughput. Zvikamu zvinogoverwa zvinodzivirira miganhu yakaoma pabasa rega rega mumamiriro ezvinhu ekubhadhara sezvaunofamba, nepo throughput yakatarwa ichikosha kune mabasa eQPS ane latency yakanyanya uye asingafaniri kudzorwa. Kutarisa mitengo yezvikumbiro uye mashandisirwo etokeni kunokubatsira kusarudza kuti uchachinja riini kubva pakudiwa kuenda pakukwanisa kwakagadziriswa.

Kudzora mutengo kunonyanya kuenderana nekusarudza mamodheru akangwara, dhizaini yekungwarira nekukurumidza uye kudzivirira ma tokeni asina basa. Tanga nemamodeli asingadhuri zvakanyanya paunogona, chengeta zvinokurudzirwa zviri nyore asi zvine ruzivo, kumbira zvakajeka kuti zvibudiswe muchidimbu kana zvichibvira, shandisa nguva yekuchengetedza zvinyorwa kuti uwane zvinokurudzirwa zvikuru, uye funga nezvekufungidzira kwebatch kana mabasa akawanda achibvumira.

Kugadzirisa zviwanikwa zveCloud Run uye zvidzikiso zvenguva refu zvinowedzera mari yekushandisa panguva yekushanda. Tanga neCPU/memory yakajairika, tarisa mashandisirwo chaiwo uye gadzirisa. Pamabasa anotarisirwa, zvidzikiso zvekushandisa zvakavimbiswa zvinoderedza zvakanyanya mari inoshandiswa.

Panyaya yekucherechedza, unofanira kubata vamiririri sevanhu vepamusoro-soro muhurongwa hwako hwekutarisa. Nyora zvavanoisa, zvisarudzo zvavo (semuenzaniso maturusi avanodaidza, mumiririri wavanopa), uye taura shanduko. Shandisa magirafu ezviitiko zveADK muwebhu UI kuti ugadzirise matambudziko ezvikamu zvega zvega, uye Cloud Logging pamwe nemadhibhodhi akagadzirwa emafambiro ezvikepe.

Zvaitwa zvakanaka, maitiro aya anokupa maonero akajeka eMAS yako: unogona kuona kuti ndeapi maekisesaizi anononoka, ndeapi maturusi anoshandiswa zvakanyanya, apo zvinokurudzirwa zvinotora nguva yakareba, uye apo maitiro ekudzora hunhu akafanana neaya. LoopAgent kudzokorora kupfuura zvaitarisirwa. Kutenderera kwemhinduro kwakakosha pakugadzirisa mhando uye mutengo nekufamba kwenguva.

Nekubatanidza ma "primitives" e "agent primitives" eADK, maitiro e "workflow" uye nzira dzehurumende ne "reference architecture" yeGoogle Cloud, security stack uye operational tooling, unogona kugadzira masisitimu e "multi-agent" asina kungwara papepa chete asiwo anogona kushandiswa, anodzorwa uye anoshanda zvakanaka mukugadzira. Kutanga kubva kune vamiririri vevabereki/vana vari nyore uye kufambira mberi kuburikidza neSequential, Loop uye Parallel orchestrations, unowana chishandiso chekushandura mazano evamiririri kuita maapplication akasimba, anochengetedza ayo anopa kukosha kwebhizinesi.