Kwethulwa okukhishwa okuhlelekile kokungenisa imodeli yangokwezifiso e-Amazon Bedrock

nimda November 7, 2025

0 4 6 minutes read

Kwethulwa okukhishwa okuhlelekile kokungenisa imodeli yangokwezifiso e-Amazon Bedrock

Nge-Amazon Bedrock Custom Model Templevol, ungahambisa futhi ulinganise ukukala okuhle noma amamodeli wesisekelo wokuphawula endaweni ephethwe ngokuphelele, engenasici. Ungaletha amamodeli akho embhedeni we-Amazon, uwalinganise ngokuphepha ngaphandle kokuphatha ingqalasizinda, futhi ubahlanganise namanye amandla okulala embhedeni we-Amazon.

Namuhla, sijabule kakhulu ukumemezela ukungezwa kokuphuma okuhlelekile kokungenisa imodeli yangokwezifiso. Ukukhishwa okuhlelekile kucindezela inqubo yesizukulwane semodeli ngesikhathi sangempela ukuze wonke amathokheni akhiqiza ukuvumelana ne-schema oyichaza. Esikhundleni sokuthembela kumaqhinga obunjiniyela abasebenza ngokushesha noma ama-brittle acubungulayo, manje usungakhiqiza imiphumela ehlelekile ngokuqondile ngesikhathi sokuvula.

Ngezinhlelo ezithile zokukhiqiza, ukubikezela kokuphuma kwemodeli kubaluleke kakhulu kunokuguquguquka kwabo kokudala. Isevisi yamakhasimende Chatbot ingahle inzuzo ngezimpendulo ezahlukahlukene, ezokwenziwa kwemvelo, kepha uhlelo lokufaka i-oda ludinga idatha ngqo, ehlelekile evumelana nama-schem achazwe ngaphambilini. Ukukhishwa okukhishwe okuhlelekile leligebe ngokugcina ubuhlakani bamamodeli wesisekelo ngenkathi uqinisekisa imiphumela yabo ihlangabezana nezidingo eziqinile zokufomatha.

Lokhu kufanekisela ukushintshwa kusuka kusizukulwane sefomu lefomu lamahhala kuya kokuphuma okungaguquki, okufundeka ngomshini, futhi kwenzelwe ukuhlanganiswa komthungo nezinhlelo zebhizinisi. Ngenkathi imibhalo yefomu lamahhala le-Excels yokusebenzisa kwabantu, izinhlelo zokusebenza zokukhiqiza zidinga ukunemba okwengeziwe. Amabhizinisi awakwazi ukukhokhela ubuqili bokuhluka kolimi lwemvelo lapho izinhlelo zabo zincike ekuphumeni okuhleliwe ku-interface ethembekile nge-API, imininingwane, kanye nokuhamba komsebenzi okuzenzakalelayo.

Kulokhu okuthunyelwe, uzofunda ukuthi ungawusebenzisa kanjani umphumela ohlelekile wokungenisa amamodeli wangokwezifiso e-Amazon Bedrock. Sizomboza ukuthi yikuphi ukuphuma okuhlelekile, ukuthi ungakuvumela kanjani ngezingcingo zakho ze-API, nokuthi ungayisebenzisa kanjani kwizimo zangempela zomhlaba ezidinga ukuphuma okuhlelekile, okucatshangelwayo.

Ukuqonda okukhishwe okuhlelekile

Ukukhishwa okuhlelekile, okwaziwa nangokuthi ukuqunjelwa okuphoqelekile, kuyindlela eqondisa ukuphuma kwe-LLM ukuze ahambisane ne-schema echazwe ngaphambilini, njenge-json evumelekile. Esikhundleni sokuvumela imodeli ukuthi ikhethe ngokukhululekile amathokheni ngokuya ngokusatshalaliswa okungenzeka, wethula izingqinamba ngesikhathi sesizukulwane esinciphisa izinqumo kulabo kuphela abazogcina ukuba semthethweni. Uma ithokheni ethile lingaphula i-schema ngokukhiqiza i-json engalungile, noma isebenzisa igama lenkambu elingalindelekile Ukukhishwa okuhleliwe kuyakulahla futhi kudinga imodeli ukuthi ikhethe enye inketho evunyelwe. Lokhu kuqinisekiswa kwesikhathi sangempela kusiza ukugcina umphumela wokugcina ungaguquki, ufunde umshini, futhi ngokushesha usebenze ngezicelo ezansi nomgwaqo ngaphandle kwesidingo sokucutshungulwa okwengeziwe.

Ngaphandle kokukhipha okuhlelekile, abathuthukisi bavame ukuzama ukuphoqelela ukwakheka ngemiyalo esheshayo efana nokuthi “Phendula kuphela eJSON.“Ngenkathi le ndlela kwesinye isikhathi isebenza, ihlala ingathembeki ngenxa yemvelo ye-LLMS ye-LLMS. Lezi zimodeli zikhiqiza umbhalo ngesampula kusuka ekusabalaliseni okungenzeka, ukwethula ukuhlukahluka kwemvelo okwenza izimpendulo zizizwe ngomuntu ezenzakalelayo.

Cabanga ngohlelo lokuxhaswa kwamakhasimende olwahlukanisa amathikithi: Uma izimpendulo zihluka phakathi “Lokhu kubukeka sengathi kuyinkinga yokukhokha,“”Ngingakwazi ukuhlukanisa lokhu njengoku: ukukhokhisa,“Futhi”Isigaba = ukukhokhisa,“I-Downtrum Code ayikwazi ukuhumusha ngokuthembekile imiphumela. Yiziphi izinhlelo zokukhiqiza ezidinga esikhundleni salokho zingabanjwa, okuhlelekile. Isibonelo:

{
  "category": "billing",
  "priority": "high",
  "sentiment": "negative"
}

Ngempendulo enjengale, uhlelo lwakho lokusebenza lungakwazi ngokuzenzakalelayo amathikithi omzila, ukugeleza komsebenzi okubangela, noma ukubuyekeza imininingwane ngaphandle kokungenelela komuntu. Ngokuhlinzeka ngezimpendulo eziqondakalayo, ezihleliwe ze-schema, okukhishwa okuhlelekile kuguqula ama-LLM avela kumathuluzi wokuxoxa abe yizakhi zohlelo ezithembekile ezingahlanganiswa nemininingwane, ama-API, kanye ne-legic logic. Lokhu kunamandla kuvula amathuba amasha okuzenzakalela ngenkathi kugcinwa ukucabanga okuqondayo okwehlisa inani lalezi zinhlobo.

Ngaphandle kokwenza ngcono ukwethenjwa nokwenza lula ukucubungula kwangemva kokusebenza, okuhlelekile kunikeza izinzuzo ezengeziwe eziqinisa ukusebenza, ukuphepha kanye nokuphepha ezindaweni zokukhiqiza.

Ukusetshenziswa kwethokheni ephansi nezimpendulo ezisheshayo: Ngokukhulisa isizukulwane ku-schema echaziwe, okukhishwe okuhlelekile kususa i-verbose engadingekile, umbhalo wamafomu wamahhala, okuholela ekutheni kuncishiswe isibalo sethokheni. Ngoba isizukulwane sethokheni sihlelekile, imiphumela emifushane ngokuqondile zihumusha izimpendulo ezisheshayo kanye ne-latency ephansi, ngcono ukusebenza okuphelele kanye nokusebenza kwezindleko.
Ukuphepha okuthuthukile ngokulwa nomjovo osheshayo: Ukukhishwa okuhlelekile kunciphisa isikhala senkulumo yemodeli futhi kusiza ukuvimba ukuthi kungakhiqizi okuqukethwe okuphikisayo noma okungaphephile. Abalingisi ababi abakwazi ukujova imiyalo, ikhodi noma umbhalo obungalindelekile ngaphandle kwesakhiwo esichaziwe. Inkambu ngayinye kufanele ihambisane nohlobo lwayo kanye nefomethi yayo, uqiniseke ukuthi imiphumela ihlala ngaphakathi kwemingcele ephephile.
Izilawuli Zokuphepha Nenqubomgomo: Ukukhishwa okuhlelekile kukuvumela ukuthi uthande ukuklama ama-Schemas asiza ukuvikela okuqukethwe okuyingozi, okunobuthi, noma kwenqubomgomo. Ngokukhawulela amasimu kumanani avunyelwe, ukuphoqelela amaphethini, futhi unqande umbhalo wefomu lamahhala, ama-schemas qiniseka ukuthi imiphumela ihambisane nezidingo zokulawula.

Esigabeni esilandelayo, sizohlola ukuthi imiphumela ehlelwe isebenza kanjani nge-Custom Model Entrock e-Amazon Bedrock futhi ihamba ngesibonelo sokuyinika amandla kumakholi akho we-API.

Kusetshenziswa okukhishwe okuhlelekile nge-Custom Model Lection in Amazon Bedrock

Ake siqale ngokucabanga usuvele ungenise imodeli yobuso be-hugging e-Amazon Bedrock usebenzisa isici sokungenisa imodeli yangokwezifiso.

Izimfuneko

Ngaphambi kokuqhubeka, qiniseka ukuthi unayo:

I-Akhawunti ye-AWS esebenzayo ngokufinyelela ku-Amazon Bedrock
Imodeli yangokwezifiso eyenziwe e-Amazon Bedrock isebenzisa isici sokungenisa imodeli yangokwezifiso
Ubunikazi be-AWS obufanele kanye nokuphathwa kokufinyelela (i-IAM) ukunxusa amamodeli ngokusebenzisa i-Amazon Bedrock Runtime

Ngalezi zidingo ezisendaweni, ake sihlole ukuthi singawusebenzisa kanjani umphumela ohlelekile ngemodeli yakho engenisiwe.

Ukuqala ukusebenzisa umphumela ohlelekile nge-Custom Model Leanrock e-Amazon Bedrock, qala ngokumisa imvelo yakho. Ku-Python, lokhu kubandakanya ukudala iklayenti le-Bedrock Runtime futhi liqale ithokheni kusuka kumodeli yakho yobuso obungenisiwe.

Iklayenti le-Bedrock Ruretime Client linikeza ukufinyelela kwimodeli yakho engenisiwe usebenzisa umbhede InvokeModel API. I-Tokenizer isebenzisa ithempulethi yengxoxo efanele evumelana nemodeli engenisiwe, echaza ukuthi umsebenzisi, uhlelo, nemiyalezo yomsizi ehlanganiswa kanjani ne-stream eyodwa, ukuthi izindima zinjani (ngokwesibonelo, <|user|>, <|assistant|>) Kufakiwe, futhi lapho impendulo yemodeli kufanele iqale khona.

Ngokubiza tokenizer.apply_chat_template(messages, tokenize=False) Ungakhiqiza ngokushesha okufana nefomethi yokufaka ngqo imodeli yakho ilindelwe, okubalulekile ekutholeni okungaguquki futhi okuthembekile, ikakhulukazi lapho kufakwa khona inkokhelo ehlelekile.

import boto3
from transformers import AutoTokenizer
from botocore.config import Config

# HF model identifier imported into Bedrock
hf_model_id = "<>" # Example: "deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
model_arn = "arn:aws:bedrock:<>:<>:imported-model/your-model-id"
region      = "<>"

# Initialize tokenizer aligned with your imported model 
tokenizer = AutoTokenizer.from_pretrained(hf_model_id)

# Initialize Bedrock client
bedrock_runtime = boto3.client(
    service_name="bedrock-runtime",
    region_name=region)

Ukuqalisa ukusebenza okuhlelekile

Lapho ucela imodeli yangokwezifiso e-Amazon Bedrock, unenketho yokunika amandla okukhishwa okuhlelekile ngokungeza a response_format vimba kwisicelo sokulayisha. Leli bhulokhi lamukela i-json schema echaza impendulo yemodeli. Ngesikhathi sokuthobeka, imodeli igcizelela le schema ngesikhathi sangempela, iqinisekisa ukuthi ithokheni ngalinye elikhiqizwa livumelana nesakhiwo esichaziwe. Ngezansi kuwukuhamba ngezinyawo okukhombisa ukuthi ungawusebenzisa kanjani umphumela ohlelekile usebenzisa umsebenzi wokukhipha ikheli elilula.

Isinyathelo 1: Chaza ukwakheka kwedatha

Ungachaza umphumela wakho olindelekile usebenzisa imodeli yePydantic, esebenza njengenkontileka ethayishiwe yedatha ofuna ukuyikhipha.

from pydantic import BaseModel, Field

class Address(BaseModel):
    street_number: str = Field(description="Street number")
    street_name: str = Field(description="Street name including type (Ave, St, Rd, etc.)")
    city: str = Field(description="City name")
    state: str = Field(description="Two-letter state abbreviation")
    zip_code: str = Field(description="5-digit ZIP code")

Isinyathelo 2: khiqiza i-json schema

I-Pydantic ingaguqula ngokuzenzakalelayo imodeli yakho yedatha ibe yi-json schema:

schema = Address.model_json_schema()
address_schema = {
    "name": "Address",
    "schema": schema
}

Le-schema ichaza uhlobo lwensimu ngayinye, nesidingo, kanye nesidingo, idale i-bleeticprint ukuthi imodeli izolandela ezizukulwaneni.

Isinyathelo 3: Lungiselela imiyalezo yakho yokufaka

Fometha okokufaka kwakho usebenzisa ifomethi yengxoxo elindelwe yimodeli yakho:

messages = [{
    "role": "user",
    "content": "Extract the address: 456 Tech Boulevard, San Francisco, CA 94105"
}]

Isinyathelo 4: Faka ithempulethi yengxoxo

Sebenzisa i-tokazizer yemodeli yakho ukukhiqiza ngokushesha okufomethiwe:

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

Isinyathelo 5: Yakha isicelo

Dala umzimba wakho wesicelo, kufaka phakathi response_format Lokho kubhekisa i-schema yakho:

request_body = {
    'prompt': prompt,
    'temperature': 0.1,
    'max_gen_len': 1000,
    'top_p': 0.9,
    'response_format': {
        "type": "json_schema",
        "json_schema": address_schema
    }
}

Isinyathelo 6: cela imodeli

Thumela isicelo usebenzisa InvokeModel API:

response = bedrock_runtime.invoke_model(
    modelId=model_arn,
    body=json.dumps(request_body),
    accept="application/json",
    contentType="application/json"
)

Isinyathelo 7: Pupse impendulo

Khipha umbhalo okhiqizwayo kusuka ekuphenduleni:

result = json.loads(response['body'].read().decode('utf-8'))
raw_output = result['choices'][0]['text']
print(raw_output)

Ngoba i-schema ichaza izinkambu ezidingekayo, impendulo yemodeli izoba ne-BOOLK

{
"street_number": "456",
"street_name": "Tech Boulevard",
"city": "San Francisco",
"state": "CA",
"zip_code": "94105"
}

Ukukhishwa kuhlanzekile, i-json evumelekile engadliwa ngqo ngohlelo lwakho lokusebenza ngaphandle kwe-parsing eyengeziwe, ukuhlunga, noma ukuhlanza okudingekayo.

Ukugcina

Ukukhishwa okuhlelekile nge-Custom Model Engemuva e-Amazon Bedrock kunikeza indlela ephumelelayo yokukhiqiza izakhiwo, ukuphuma okuhambisana ne-schema kusuka kumamodeli akho. Ngokushintsha kokuqinisekiswa kumodeli ngokwayo, okukhishwe okuhlelekile kunciphisa isidingo sokuhamba komsebenzi okuyinkimbinkimbi kanye nekhodi yokuphatha iphutha.

Ukukhishwa okuhlelekile kwakha ukuphuma okuvezwa futhi okuqondile ukuhlanganisa ezinhlelweni zakho futhi kusekela amacala ahlukahlukene okusebenzisa, ngokwesibonelo, izinhlelo zokusebenza zezezimali ezidinga amadokhumenti emitholampilo aqondile, noma izinhlelo zensizakalo yamakhasimende ezidinga ukuhlukaniswa kwamathikithi okungaguquki.

Qala ukuzama ngokuphuma okuhlelekile nge-Model Commor Model Tractom namuhla futhi uguqule ukuthi izinhlelo zakho ze-AI ziletha kanjani ukuguqulwa, imiphumela elungele ukukhiqizwa.

Mayelana nababhali

UManoj Selvakumar Ingabe ukwakhiwa kwezixazululo ze-AI akhiqizayo ku-AWS, lapho asiza khona izinhlangano ezakhiwe, i-prototype, futhi abeke izixazululo ezinamandla anikwe amandla efwini. Ngobuchwepheshe ekufundeni okujulile, amasistimu we-scarlwable Cloud-Nature-Naturenchest, ugxile ekuguqukeni okusha ekwakhiweni kwezakhiwo ezilungele ukukhiqizwa kwenani lebhizinisi elilinganiswayo elishayela inani lebhizinisi elilinganisekayo. Unothando ngokwenza imiqondo eyinkimbinkimbi ye-AI esebenzayo futhi inika amandla amakhasimende ukuthi asuse amakhono ngokuzibonela – ukusuka ekuhlolweni kusenesikhathi kokuqala kwebhizinisi. Ngaphambi kokujoyina ama-AWS, i-Manoj isebenze ngokubonisana, ukuhambisa iSayensi yeSayensi nama-AI Solutions amaklayenti okufunda amabhizinisi, amasistimu wokufunda womshini wokugcina asekelwa yimikhuba eqinile ye-Mlops yokuqeqeshwa, ukuthunyelwa kanye nokuqapha ekukhiqizeni.

Yanyan zhang Ungusosayensi wedatha we-AI aphezulu e-AI e-Amazon Web Services e-Amazon Web Services, lapho ubesebenza khona kubuchwepheshe be-AI / ML onqenqemeni njengochwepheshe we-AI obulalayo, abasize amakhasimende asebenzise i-generative AI ukufeza imiphumela edingekayo. U-Yanyan uthweswe iziqu eTexas A & M University nge-PHD enjiniyela kagesi. Ngaphandle komsebenzi, uyakuthanda ukuhamba, ukusebenza, nokuhlola izinto ezintsha.

Lokeshwaran ravi Ingabe unjiniyela ophambili wokufunda ojulile we-aw aw aw aw aw, abhekelele ukusebenza kahle kwe-ML, ukushesha kwemodeli, kanye nokuphepha kwe-AI. Ugxile ekuthuthukiseni ukusebenza kahle, ukunciphisa izindleko, kanye nokwakha izimo zemvelo eziphephile ku-DemoundAtize ai Technologies, okwenza ukunqunywa kwe-ML kufinyeleleke futhi kube nomthelela kuzo zonke izimboni.

I-Revendra Kumar Unjiniyela wokuthuthukisa isoftware aphezulu ezinsizakalweni zewebhu ze-Amazon. Endimeni yakhe yamanje, ugxile ekusingatheni okuyisibonelo nokuzithoba embhedeni we-amazon. Ngaphambi kwalokhu, wasebenza njengonjiniyela kumakhompiyutha we-quantum efwini nasekuthuthukiseni izixazululo zengqalasizinda zezindawo ezisezingeni eliphakeme. Ngaphandle kwezinto zakhe ezingochwepheshe, i-Revelodra ijabulela ukuhlala isebenza ngokudlala i-tennis nokuhamba ngezinyawo.

Muzart tuman Ingabe unjiniyela wesoftware esebenzisa isipiliyoni sakhe emikhakheni efana nokufunda okujulile, ukusebenza kahle komshini, kanye nezicelo eziqhutshwa ngumshini zokusiza ukuxazulula izinkinga zangempela zomhlaba ngendlela ebonakalayo, esebenza kahle, futhi efinyelelekayo. Umgomo wakhe ukudala amathuluzi anomthelela ongagcini ngokuthuthukisa amakhono e-technical kuphela kepha abuye akhuthaze ushintsho olunengqondo kuzo zonke izimboni nemiphakathi.