1. Gabatarwa
Wannan takarda tana magance ƙalubalen samar da damar amfani da Hanyoyin Haɗin Aikace-aikacen Software (APIs) ta hanyar amfani da Manyan Samfuran Harshe (LLMs). Haɗin kai na gargajiya na API yana buƙatar ilimin fasaha na tsari, sigogi, da takamaiman kira, yana haifar da shinge ga masu amfani marasa fasaha. Tsarin da ake shawarwari yana amfani da LLMs don ayyuka biyu na farko: 1) Rarraba shigarwar harshe na halitta na masu amfani zuwa kiran API masu dacewa, da 2) Sarrafa samar da bayanan ƙirƙira, na musamman don aiki don kimanta aikin LLM don ayyukan rarraba API. Wannan hanyar biyu tana nufin rage shingen amfani da software yayin samar da kayan aiki mai amfani ga masu haɓakawa don tantance dacewar LLM don sarrafa API na musamman.
2. Ayyukan Da Suka Gabata
Binciken ya ginu akan ayyukan da suka wanzu a cikin NLP da injiniyan software, yana mai da hankali kan haɗa harshen ɗan adam da umarnin da na'ura za ta iya aiwatarwa.
2.1 LLMs don Haɗa Harshe na Halitta zuwa Taswirar API
Nazarin da suka gabata sun bincika amfani da samfuran jeri-zuwa-jeri da ingantattun nau'ikan BERT don taswirar harshe na halitta zuwa lamba ko jerin API. Zuwan LLMs masu ƙarfi, na gaba ɗaya kamar GPT-4 ya canza tsarin, yana ba da damar taswira mai sassauƙa da fahimtar mahallin ba tare da horo mai zurfi na musamman ba.
2.2 Samar da Bayanan Ƙirƙira a cikin NLP
Samar da bayanan ƙirƙira, mahimmanci don horo da kimantawa inda ainihin bayanai ba su da yawa, ya samo asali daga samfuran tushen ƙa'ida zuwa samar da LLM. Samfura kamar GPT-4 na iya samar da misalan rubutu iri-iri, masu dacewa da mahallin, waɗanda ake amfani da su a cikin wannan aikin don ƙirƙirar bayanai don ayyukan API na musamman.
3. Tsarin Da Ake Shawarwari
Babban ƙirƙira shine tsarin haɗin kai wanda ke sarrafa duka aikin rarrabuwa da ƙirƙirar ma'aunin kimantawa nasa.
3.1 Tsarin Tsarin Tsarin
Tsarin ya ƙunshi sassa biyu masu haɗin kai: Sashen Rarrabuwa da Sashen Samar da Bayanan Ƙirƙira. Babban mai tsarawa yana sarrafa aikin, yana ɗaukar ƙayyadaddun API a matsayin shigarwa kuma yana fitar da ko dai kiran API da aka rarraba ko bayanan kimantawa da aka samar.
3.2 Rarraba Harshe na Halitta zuwa API
Idan aka ba da tambayar harshe na halitta $q$ da saitin yiwuwar kiran API $A = \{a_1, a_2, ..., a_n\}$, LLM yana aiki azaman mai rarrabuwa $C$. Manufar ita ce nemo API $a_i$ wanda ke haɓaka yuwuwar sharadi: $a^* = \arg\max_{a_i \in A} P(a_i | q, \theta)$, inda $\theta$ ke wakiltar sigogin LLM. Tsarin yana amfani da ƙaramin nuna misalai tare da misalai don jagorantar samfurin.
3.3 Hanyar Samar da Bayanan Ƙirƙira
Don aikin API da aka yi niyya, sashen samarwa yana amfani da LLM (misali, GPT-4-turbo) don ƙirƙirar saiti iri-iri na tambayoyin harshe na halitta $Q = \{q_1, q_2, ..., q_m\}$ waɗanda suka dace da wannan API. Ana jagorantar tsarin ta hanyar nuna cewa an ƙayyade manufar API, sigogi, da bambance-bambancen da ake so a cikin jumla, rikitarwa, da niyyar mai amfani.
4. Tsarin Gwaji & Sakamako
4.1 Tsarin Samar da Bayanai
An samar da samfuran bayanai don ayyukan API da yawa (misali, dawo da yanayi, tambayar bayanai, sarrafa biyan kuɗi) ta amfani da GPT-4-turbo. Kowane bayanai ya ƙunshi ɗaruruwan tambayoyin harshe na halitta waɗanda aka haɗa da madaidaicin alamar kiran API, suna rufe kewayon fassarori da maganganun mai amfani.
4.2 Kwatancen Aikin Samfurin
An kimanta LLMs da yawa akan bayanan da aka samar ta amfani da daidaitaccen daidaiton rarrabuwa.
GPT-4
0.996
Daidaito
GPT-4o-mini
0.982
Daidaito
Gemini-1.5
0.961
Daidaito
LLaMA-3-8B
0.759
Daidaito
4.3 Binciken Sakamako
Sakamakon ya nuna babban gibin aiki tsakanin babban samfurin mallakar mallaka (GPT-4) da babban mai fafatawa na buɗe tushe (LLaMA-3-8B). Wannan yana nuna mahimmancin ƙarfin samfurin don amintaccen turawa a duniyar gaske. Babban daidaiton manyan samfura yana tabbatar da yuwuwar amfani da LLMs don daidaitaccen rarrabuwar kiran API.
5. Binciken Fasaha & Fahimta ta Asali
Fahimta ta Asali: Wannan takarda ba kawai game da amfani da LLM a matsayin mai rarraba API ba ne; tsarin meta ne don kimantawa wane LLM za a yi amfani da shi don wannan takamaiman aikin. Ainihin samfurin shine injin samar da bayanan ƙirƙira, wanda ke juya matsalar "dacewar LLM" zuwa ma'auni mai aunawa, mai ma'auni. Wannan motsi ne mai hikima, yana gane cewa a zamanin LLM, ikon ƙirƙirar bayanan kimantawa masu inganci na kanku yana da daraja kamar samfurin kansa.
Kwararar Ma'ana: Hujjar tana da da'ira mai kyau kuma tana ƙarfafa kanta: 1) Muna buƙatar LLMs don fahimtar harshe na halitta don APIs. 2) Don zaɓar LLM da ya dace, muna buƙatar bayanai na musamman na aiki. 3) Ainihin bayanai suna da wahala a samu. 4) Saboda haka, muna amfani da LLM mai ƙarfi (GPT-4-turbo) don samar da wannan bayanan. 5) Sa'an nan muna amfani da wannan bayanan don gwada wasu LLMs. Tsarin farawa ne wanda ke amfani da samfurin mafi ƙarfi da ake da shi don tantance fagen.
Ƙarfi & Kurakurai: Babban ƙarfin shine aiki. Wannan tsarin yana ba da mafita da za a iya amfani da ita nan da nan ga kamfanoni masu kallon jerin APIs da dashboard na LLMs masu samuwa (OpenAI, Anthropic, Google, buɗe tushe). Kurakuri, wanda marubutan suka yarda da shi, shine haɗarin "LLM-inception": amfani da LLM don samar da bayanai don gwada LLMs na iya gaji da haɓaka son zuciya. Idan GPT-4 yana da makaho a fahimtar wani nau'in tambaya, zai samar da bayanan gwaji marasa kyau, kuma duk samfuran za a yi musu hukunci bisa ma'auni mara kyau. Wannan yana kwatanta ƙalubalen da aka gani a wasu yankuna na samarwa, kamar zagayowar horarwar GANs inda mai samarwa da mai nuna bambanci zasu iya haɓaka cututtuka tare.
Fahimta Mai Aiki: Ga CTOs da manajoji samfura, abin da za a ɗauka a bayyane yake: Kada kawai kuyi gwajin GPT-4 don mu'amalar harshe na halitta na API. Gwada wannan tsarin. Yi amfani da shi don gudanar da gasa tsakanin GPT-4o, Claude 3, da Gemini akan ainihin ƙayyadaddun API ɗinku. Gibin daidaiton maki 24 tsakanin GPT-4 da LLaMA-3-8B shine gargadi mai tsanani cewa zaɓin samfurin ba ƙaramin abu bane kuma farashi (kyauta vs. biya) wakili ne mai haɗari ga aiki. Tsarin yana ba da shaidar ƙididdiga da ake buƙata don yin wannan yanke shawara na dandalin miliyan-miliyan.
6. Misalin Aiwatar da Tsarin
Yanayi: Kamfanin fasahar kuɗi yana son ƙara mu'amalar harshe na halitta zuwa "API na Binciken Ma'amala" na ciki wanda ke da ayyuka kamar get_transactions_by_date(date_range, user_id), flag_anomalous_transaction(transaction_id, reason), da generate_spending_report(user_id, category).
Aiwatar da Tsarin:
- Samar da Bayanai: Kamfanin yana amfani da Sashen Samar da Bayanan Ƙirƙira (wanda GPT-4-turbo ke iko da shi) tare da nuna bayanin kowane aikin API. Don
get_transactions_by_date, yana iya samar da tambayoyi kamar: "Nuna mini sayayyata daga makon da ya gabata," "Me na kashe tsakanin Maris 1st da 10th?", "Zan iya ganin tarihin ciniki na na watan da ya gabata?" - Kimanta Samfurin: Suna amfani da bayanan da aka samar (misali, tambayoyi 500 a cikin ayyukan API 3) don gwada LLMs masu yuwuwa: GPT-4o, Claude 3 Sonnet, da Llama 3 na ciki da aka inganta. Suna auna daidaito da jinkiri.
- Zaɓi & Turawa: Sakamakon ya nuna Claude 3 Sonnet ya cimma daidaiton 98.5% a rabin farashin kira na GPT-4o, yana mai da shi zaɓi mafi kyau. Ingantaccen Llama 3 ya zira maki 89% amma yana ba da sirrin bayanai. Fitowar ƙididdiga tana jagorantar yanke shawara mai bayyane, bisa shaida.
7. Aiwatar da Gaba & Hanyoyi
Tasirin wannan aikin ya wuce rarrabuwar API mai sauƙi:
- Haɓaka Dandalin Ƙarancin Lamba/Babu Lamba: Haɗa wannan tsarin cikin dandamali kamar Zapier ko Microsoft Power Platform zai iya ba masu amfani damar gina sarrafawa masu rikitarwa ta amfani da harshe na halitta kawai, wanda tsarin ke fassara zuwa jerin kiran API a cikin ayyuka daban-daban.
- Democratization na Software na Kamfani: Rukunin software na kamfani masu rikitarwa (misali, SAP, Salesforce) tare da ɗaruruwan APIs za su iya zama masu isa ga masu nazarin kasuwanci ta hanyar mu'amalar tattaunawa, yana rage girman horo sosai da faɗaɗa amfani.
- Tsarin Halittu na API Mai Ƙarfi: A cikin tsarin IoT ko microservices inda APIs sukan canza ko an ƙara sababbi, ana iya gudanar da sashen samar da bayanan ƙirƙira lokaci-lokaci don sabunta bayanan kimantawa da sake tantance LLM mafi kyawun aiki, ƙirƙirar Layer na mu'amala mai daidaita kanta.
- Hanyar Bincike - Rage Mafarki: Mataki na gaba mai mahimmanci shine haɗa tabbatarwa na yau da kullun ko duba ƙuntatawa, wanda aka yi wahayi daga dabarun haɗa shirye-shirye, don tabbatar da cewa kiran API da aka rarraba ba kawai mai yuwuwa ba ne har ma ingantacce a ma'ana kuma lafiya don aiwatarwa.
- Hanyar Bincike - Shigarwa Masu Nau'i Daban-daban: Tsare-tsare na gaba na iya karɓar tambayoyi masu nau'i daban-daban (misali, mai amfani yana nuna alamar dashboard yayin yin tambaya) kuma su taswira su zuwa kiran API gama gari, haɗa hangen nesa na kwamfuta tare da NLP.
8. Nassoshi
- Brown, T. B., et al. (2020). Samfuran Harshe Ɗalibai ne Kaɗan. Ci gaba a cikin Tsarin Bayanai na Neural, 33.
- OpenAI. (2023). Rahoton Fasaha na GPT-4. arXiv:2303.08774.
- Zhu, J. Y., Park, T., Isola, P., & Efros, A. A. (2017). Fassarar Hotuna-zuwa-Hoto mara Haɗin gwiwa ta amfani da Cibiyoyin Adawa na Da'ira-Daidaitacce. Proceedings of the IEEE International Conference on Computer Vision.
- Raffel, C., et al. (2020). Bincika Iyakokin Canja wurin Koyo tare da Mai Fassara Rubutu-zuwa-Rubutu ɗaya. Journal of Machine Learning Research, 21.
- Schick, T., & Schütze, H. (2021). Samar da Bayanai tare da Samfuran Harshe da aka riga aka horar. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing.
- Microsoft Research. (2023). Zamanin Copilots: Haɓaka Software Mai Ƙarfin AI. An samo daga Microsoft Research Blog.
- Google AI. (2024). Gemini: Iyali na Samfuran Nau'i Daban-daban Masu Ƙarfi. Rahoton Fasaha.