Tussna tafrawant

TI tasnulfant deg tmɣawin n Murakuc: aseqdec, ixaṭaṛen d tarrayt

Sɣur Yasser MsahalYeffeɣ-d ass n 12 Kṭuber 20269 n tesdatin n tɣuri

TI tasnulfant (IA générative) tenfeɛ tamɣawt B2B deg Murakuc mi tt-tesseqdec ɣef yiwet n tbadut iẓeṛṛan: asekker n yisemliyen, aheggi n tririyin i tallalt n yimsaɣen, anadi n telɣut deg yisemliyen igensanen, neɣ tira n tẓegri tamezwarut n uneɣmis. Ur tettak ayen yettwarǧan mi tt-tessnesreḍ deg yal tama war ilugan. Kraḍ n tɣawsiwin ay yesbeddalen kullu: ad tesmeẓleḍ isefka uqbel ad ten-tazneḍ i umudil, ad yeqqim yiwen n umdan d amasay ɣef wayen d-yeffɣen, d ad tebduḍ s upilot ittwaktalen.

Mayd tezmer ad tt-tegg TI tasnulfant i tmɣawt B2B deg Murakuc?

Amudil ameqqran n tutlayt (LLM, large language model) yeqqar, yessewzal, yessmeẓli d yaru aḍris. Deg B2B, ittalli deg twuriwin yettuɣalen anida yettaddar wakud deg tɣuri, deg unadi d usenɣel. Kkuẓ n twaculin n useqdec ay d-yellan, fessusit i usbedded, fessusit i usenqed.

• Asekker n yisemliyen: tifaktuṛin, tuttriwin n usɣa, ikuntraten neɣ ccekwat n yimsaɣen. Amudil yessufuɣ-d iferdisen yeṣṣefḍen, yessmeẓli asemli, d yesmel ayen ixuṣṣen. Deg Murakuc, isemliyen-a ttaseɣ-d s tefṛansist, s taɛrabt neɣ s tesmeskelt n snat, tikkal ttwasekenen (scan): ɛreḍ amudil ɣef yimedyaten-nnek s yiman-nnek, s yal tutlayt i d-yettaseɣ, uqbel ad d-tessufɣeḍ kra n tegzemt.
• Tiẓegriyin n tririyin n tallalt: amudil yeqqar atiki (ticket), yettaf-d ayen yeṣṣefḍen deg yisura-nnek, d yessumer-d tiririt ara yeɣṛ umsallal, ad tt-yeseɣti, sakin ad tt-yazen. Amɣur yella deg unadi d tira, maci deg tegzemt. Ma ttarun-d yimsaɣen-nnek s darija (taɛrabt n Murakuc) daɣen, smel iznan-a ɣer yirmiten-nnek.
• Anadi deg tesmelt n yisemliyen igensanen: isura n tɣara, tisertiyin n tɣerma tamdant (RH), isemliyen itiknikiyen. Amexdam ad yesteqsi s tutlayt n yal ass, ad yawi tiririt d tseddarin n yiɣbula-ns.
• Asenfel n yineɣmisen: ineɣmisen n temlilit, isewzalen n yal ayur, tinɣmisin n ufran n tama. Amudil yessmun-d tẓegri tamezwarut ara yessemd umasay, ad tt-yesentem; imḍanen kkan-d seg yinagrawen-nnek, ur kkin bdan seg tmektit n umudil.

Ayen yemsasan gar kkuẓ n useqdec-a: amudil ittheggi, amdan ay iggzem. D ayen-a ay sen-yeǧǧan ad sbedden war ma ttṛaǧun amudil amagnu.

RAG: asteqsi n yisemliyen n tmɣawt war asmuɣli n umudil

RAG (Retrieval-Augmented Generation, asnulfu yettwasneɣten s unadi) ur yessteqsay ara amudil ɣef wayen « yessen »: yal asteqsi, ittak-as tiseddarin yeṣṣefḍen seg yisemliyen-nnek. Ur ilaq bdan asmuɣli n umudil ɣef yisefka-nnek, d asura yettwaleqqmen yettwaḥseb ɣas ad yales asmel-ns deg umatar (réindexation).

1. Isemliyen yettwasirgen ttwabḍan ɣer tseddarin, ttwasmelen deg umatar.
2. Yal asteqsi, anagraw yettaf-d tiseddarin ay as-iqerben ugar.
3. Amudil yettawi tiseddarin-a d unaḍ ad ur d-yerr ala seg-sent.
4. Tiririt tettban-d d yiɣbula-ns, i usenqed.

Kraḍ n tɣawsiwin ay iggzmen tigezmi. Isemliyen s yiman-nsen: isura iqburen neɣ yemgaraden ad d-awin tiririyin tiqburin neɣ yemgaraden. Izerfan n unekcum: anadi ilaq ad yeḍfer tisirag n usefrek n yisemliyen-nnek, ma ulac aneɣmaru (stagiaire) yezmer ad yawi tiririt yekkan-d seg usemli n tnemhalt n tdrimt. D usenqed: uqbel asenker, heggi tabdart n yisteqsiyen n tidet, s tririt yeṣṣeḥḥan d usemli aɣbalu yeṣṣeḥḥan, sakin ales-tt yal tikkelt ara tbeddleḍ amudil, anaḍ neɣ tarrayt n ubḍu.

Tilisa n LLM deg tmɣawin: hallucinations, tuffɣa n yisefka, tiṣeṛfiwin

Hallucinations (tiririyin yettwasnulfan)

LLM yettaru aḍris ay yettbanen ugar d tidet, maci win yellan d tidet. Yezmer ad d-yini tafurt ur nelli deg ukuntra neɣ amḍan n yidrimen yettwasnulfan am wakken d tiririt yeṣṣeḥḥan. Tifrat: ad tḥetmeḍ iɣbula, ad tesirgeḍ i umudil ad d-yini « ur ssineɣ », ad tesneqdeḍ s wudem awurman iferdisen yettwassufɣen (talɣiwin, imḍanen imatuten, timerniwin), d ad yeɣṛ yiwen n umdan kra n wayen ara yeffɣen seg tmɣawt.

Tuffɣa n yisefka

Ma tsenteḍ akuntra n umsaɣ deg yiwen n ufecku n yal yiwen (grand public), d akken kan ay t-tefkiḍ i wis kraḍ ur tḥekkmeḍ bdan deg yiqeddacen-ns, ula deg wakud n uḥraz, ula dima ayen ara yegg yis. Deg ugensu, amɛiwen (assistant) ur yettwaḥeṛzen ara mliḥ yezmer ad d-yesken talɣut i win ur nesɛi azerf ad tt-yẓer. Amdeg aneggaru, ur yettwassen ara aṭas: asemli neɣ imayl yeqqar umudil yezmer ad yeṭṭef inaḍen yettwaffren ay ibeddlen tiririt-ns. Ur ttak bdan i umɛiwen ugar n unekcum neɣ n tezmert n tigawt ɣef wayen yesɛa umdan ay t-yesseqdacen.

Asmuɣer n tṣeṛfiwin

Imudilen ay d-yettawḍen s API, s umata, ttxellaṣen s useqdec, ɣef teɣzi n uḍris yettwaznen d win d-yeffɣen. Aɣlaf (prototype) ur nesqaṛ ara yezmer ad yeɣlu mi yessnesrar ɣef tmɣawt meṛṛa, ugar ma yella yal tuttra tettawi isemliyen ɣezzifen. Ktel tiṣeṛfiwin n yal asekker deg upilot, d sbedd tilisa uqbel ad telliḍ anekcum i medden.

Loi 09-08 d CNDP: mayd ilugan n tuzna n yisefka i LLM?

Alugen amezwaru n unabaḍ: ad tesmeẓleḍ isefka uqbel ad ten-tazneḍ i umudil. Kraḍ n yiswiren n tḥessast d yiwen n ucraḍ ay yeddan ɣef meṛṛa ay d-yudan i tazwara:

• Imeɣradanen: akatalug, isebtar n usmel, isemliyen n uznuzu yettwasnesren.
• Igensanen: isura, imudilen n yisemliyen, tinɣmisin n tuddsa.
• Tiwḥcin (confidentielles): ikuntraten, tisumar n wazal, isefka n tdrimt, tangalt taɣbalut (code source).
• Udmawen, ɣas mamek iga uswir: yal ayen ay d-yessenten amdan (amsaɣ, amexdam, amettar n uxeddim), ula deg yimayl, deg utiki neɣ deg CV.

Deg Murakuc, isefka udmawen ddaw n loi 09-08 (asaḍuf n Murakuc ɣef ugdal n yisefka udmawen), d CNDP (Tamesmunt Tanamurt n Usenqed n Ugdal n Yisefka Udmawen) ay yettḍefren asemres-ns. Asekker s TI iqqim d asekker: iswi iban, isefka ur zeggiren ayen yessefken, asselɣu n yimdanen yerzan, d, ɣef tegnatin, asuṭṭu neɣ tuttra n tsiregt ɣer CNDP. Kra n tewsatin, ttusemman tiḥessasin (tadawsa, tiẓri, liman n ddin, aẓar…), ddaw n yilugan iqquḥen ugar. Asiweḍ n yisefka udmawen ɣer berra n Murakuc, yettilin aṭas d yimeẓla n TI deg uẓeṭṭa, yettwasbedded daɣen s yilugan. Senqed tagnit-nnek ɣer CNDP neɣ ɣer umsalel-nnek n usaḍuf uqbel asenker n unagraw.

S tidet, yal aswir yettawi alugen n useqdec: isefka imeɣradanen ttwasirgen deg yal amkan; isefka igensanen s yiwen n umeẓlu kan ay tesentem tesga n IT n tmɣawt; isefka tiwḥcin d udmawen ala deffir n usenqen n yismawen (anonymisation), neɣ s tifrat ay ttwasenqden uḥraz-ns d ukuntra-ns.

Anabaḍ n TI tasnulfant: amdan deg tlawt d usekles

• Yiwen n umasay i yal aseqdec, seg tama n umussu, ittesentam tama d yigmuḍen.
• Amdan deg tlawt (human in the loop): yal agmuḍ ara yerzen tamɣawt (tiririt i umsaɣ, tagzemt n RH, asemli n ukuntra) ittwasentam uqbel tuzna.
• Asekles (journalisation): tuttra, iɣbula yettwasqedcen, alqem n umudil d unaḍ, tiririt d usentem, ttwaḥeṛzen akken ad tezmreḍ ad tesfehmeḍ agmuḍ deffir-s. Iɣmisen-a s yiman-nsen llan deg-sen isefka: ḥrez-ten, d sbedd akud n uḥraz-nsen.
• Izerfan n unekcum: afecku yettawi tisirag yellan, ur d-yettesnulfu ara tid imaynutnin.
• Tamsasit n useqdec: ayen zemren yimexdamen ad t-sekcmen neɣ ur t-zemren ara deg yifecka n TI, ula d wid n yal yiwen (grand public).

Ma yekcem-d yiwen n umsenfar (prestataire), tarrayin-ns sɛant azal am tid-nnek: ẓer ISO 27001 d tɣellist n yisefka ɣer umsenfar.

Amek ara tesmeleḍ TI tasnulfant deg tbadut n umussu, aḍar s uḍar

1. Fren yiwet n tbadut kan: tameqqrant, tettuɣal, d aḍris, anida tuccḍa tettban d tettwaseɣta. Asmeẓli n ccekwat yemmal-as; tagzemt n uṛḍil (crédit), maci i tazwara.
2. Ktel ayen yellan: akud n usekker, akud n tririt, amḍan n yiɣawasen yettuɣalen. War tasebdadt n tazwara, ur tezmireḍ ad tesbeyyneḍ amɣur.
3. Bdu apilot yettwaḥeṛzen: yiwet n tgrawt, isemliyen n tidet, isefka yettwasmeẓlin, asentem n umdan yal tikkelt, asekles yettemɣuren.
4. Serwes: imsekta-nni s yiman-nsen, d tiṣeṛfiwin n yal asekker d tmuɣli n tririyin ay seɣtan yimseqdacen.
5. Gzem: ad tessimɣureḍ, ad tseggmeḍ neɣ ad tḥebseḍ. Aḥbas n upilot ur d-nefki ayen yettwarǧan d agmuḍ, maci d aɛeṭṭel.
6. Sbedd s tidet: asmel ɣer unagraw n telɣut, aḍfar, tadrimt, d uḍfar n yimseqdacen.

Aru tama n usenfar uqbel ad tesqesreḍ imsenfaren: tira n udlis n tewtilin (cahier des charges) teddu ula i TI. Mi yerkez useqdec amezwaru, aḍar d-iteddun, s umata, d asmezdi n mennaw n tigawin: d ayen-a ay d asentel n imsewwqen n TI (agents IA) deg tmɣawin. I tmuɣli tamatut, ẓer abeddel adigan n tmɣawin timeẓyanin d tilemmasin deg Murakuc.

LLM deg tmɣawt deg Dar Lbiḍa: mayd tiḍmanin ara tsuterḍ ɣer umsenfar?

Uqbel ad tesɣelseḍ, ssuter isteqsiyen iẓeṛṛan:

• Mani ay ttwaḥeṛzen yimudilen, imataren d yiɣmisen, d mayd isefka ay iḍerrun s yimeẓla n wiyaḍ?
• Is tiwtilin n umeẓlu n umudil sirgent asales n useqdec n yisefka-nnek, amedya i usmuɣli?
• Amek ay ttwabḍan yinekcumen, d mayd ay yettwaseklasen?
• Amek ay ttwaḍefrent tṣeṛfiwin, d mamek ay sent-ttwasbeddant tilisa?
• Wi ay yesɛan tangalt, inaḍen d tsedbadin yettwafkan, d is tzemreḍ ad tbeddleḍ amudil neɣ amsenfar war ma tɛawdeḍ ad tebnuḍ kullu?

Codevista Innovations, deg Dar Lbiḍa, tesɛa iselɣun ISO/IEC 27001:2022, tessmal TI tasnulfant d LLM deg tbadutin n umussu yellan, s yiwen n usentem yettbanen: ur ttḥuzu bdan taɣellist. I usbedded n useqdec amezwaru, nermes-aɣ.

Isteqsiyen yettuɣalen

Is yezmer ad yettwaseqdec ufecku n TI n yal yiwen (grand public) d yisefka n yimsaɣen deg Murakuc?

Ur yelhi ara war asbedded. Isefka n yimsaɣen, aṭas n tikkal, d isefka udmawen ɣef wayen inna loi 09-08, d yifecka n yal yiwen ttsekkaren-ten aṭas deg yiqeddacen yellan berra n Murakuc, ɣef tiwtilin-nsen s yiman-nsen. Smeẓli isefka di tazwara, senqed uḥraz d ukuntra n umeẓlu, ssezwar asenqen n yismawen (anonymisation), d steqsi CNDP neɣ amsalel-nnek n usaḍuf ɣef tgawin ilaqen.

Is ilaq ad tesmuɣleḍ amudil-nnek n TI akken ad t-tesseqdceḍ ɣef yisemliyen n tmɣawt?

Aṭas n tikkal, ala. Tarrayt n RAG tettak i umudil, yal asteqsi, tiseddarin yeṣṣefḍen seg yisemliyen-nnek, war asmuɣli ɣef yisefka-nnek. Tesɣiwil i usbedded, tettḍfar ileqman n yisemliyen-nnek d tesken-d iɣbula n yal tiririt. Asmuɣli uslig yemmal-as ugar i talɣiwin neɣ i umawal ixuṣṣen aṭas.

Amek ara tesneqseḍ hallucinations n LLM deg tbadut n umussu?

Ur ttekksent ara, ttwaḥeṛzent. Efk iɣbula i umudil d ḥtem fell-as ad ur iffeɣ seg-sen, sireg-as ad d-yini ur yessin ara, senqed s wudem awurman imḍanen n yidrimen, izemzen d tmerniwin yettwassufɣen, d ad yesentem yiwen n umdan kra n ugmuḍ ara yerzen tamɣawt. Sekles tiririyin akken ad tafeḍ tuccḍiwin yettuɣalen.

S mayd aseqdec ara tebduḍ s TI tasnulfant deg tmɣawt?

Fren tabadut tameqqrant, tettuɣal, d aḍris, anida tuccḍa tettban d tettwaseɣta s tifses: asmeẓli n ccekwat, asufeɣ n tifaktuṛin, tiẓegriyin n tririyin i tallalt n yimsaɣen neɣ anadi deg yisura igensanen. Ktel akud d tɣara n wass-a, bdu apilot ɣef yiwet n tgrawt kan, sakin serwes igmuḍen uqbel ad tgezmeḍ ad tessimɣureḍ.