1. Catastrophic Forgetting e ɗo RAG waawaa
So tawii AI naatnii e golle ɗo waɗata e goonga, gooto e caɗeele mawɗe ko ɗo waawata waɗde e yontere hoore (Catastrophic Forgetting) nde humpito ɗo jannginaama adii waɗetee e data keso ngam fine-tuning.
Ɗum ko waɗde nde AI dogii e neural network waasata yaafde humpito walla paatone ɗo janngini adii haa sahaa juutɗo, tawi o woni e janngude humpito keso.
Misal, so a waɗii fine-tuning e LLM open source huutoraade data gootoonrewɓe, waawde ɗo humpito holluɗo e yimɓe walla waawde ɗemngal waawaa ɓeydude, waawaa.
Ngam haɗde e caɗeele ɗee, teknolosi RAG(Retrieval-Augmented Generation) huutortee no feewi, kono RAG kadi woodi loowdi mum.
Loowɗe mawɗe RAG
- Njoɓdi e huutoraade keɓe cuɓaaɗe (Structured Data QA): RAG system’en ɗon ƴellita e binndol ɗe waawaa heftinde, ngam ɗe mbaɗaa e binndol ɗe waawaa heftinde. Ngam ɗum, ɗe mbaawaa faamde walla huutoraade maana e jokkondiral keɓe cuɓaaɗe (tableaux, database, spreadsheets, ekn.).
- Ɗoŋkaare e PDF ɗoƴƴi/ɗerewol ɗe waawaa heftinde (Complex PDFs): PDF ɗoƴƴi (tableaux, koloomaji ɗuɗɗi, nate e nder mum, ekn.) mbaawi waasde keɓe walla waylude kontekst e sahaa chunking (ɓeydude keɓe e cuɓe). Ngam ɗum, keɓe ɗe njogii nafaa mbaawi waasde indeksing walla waawde yiyteede e saɗɗa.
- Alanaa model Fallback (Fallback Model(s)): So RAG system waawaa yiytude jaabawol moƴƴol, logic nde waylitata e model goɗɗo (backup) waawi waasde walla huutoraade e bonnude. Ngam ɗum, so jaabawol fali, huɗo waawaa heɓde cuɓe ɗe mbaɗi e ɗoƴƴugol mum.
- Ɓeydugol kisal (LLM Security): Ɗoƴƴugol kisal e LLM hoore mum (prompt injection, yaltinde keɓe, ekn.) waawi waasde, tee ɗum waawi waɗde keɓe sirriiɗe njaltina.
- Alanaa ɓeydude (Data Ingestion Scalability): Caɗeele ɓeydude mbaawi waɗde e indeksing e danndugol keɓe mawɗe. So keɓe ɓeydi, no chunking, danndugol e yiytugol waɗata ɓurata loowde, tee koorsol system ɓeydoto.
- Keɓe mawɗe waasde (Missing Content): Keɓe mawɗe e ɗerewol mbaawi waasde e sahaa chunking walla mbaawaa waɗde indeksing. Ngam ɗum, huɗo waawaa yiytude keɓe ɗe ɗoƴƴi.
- Waasaade rank mawɗo (Missed Top Ranked): Chunk (rank mawɗo) mo ɓuri jokkude e njaɓɓorgo waawi waasde e nate njiɗaaɗe. Sababuji ɗi mbaawi wonde ceertuɗe search, ƴellitgol embedding, walla juumre ranking.
- Jokkondiral kontekst waasde (Not in Context): Chunk njiytaaɗo waawi waasde jokkude e kontekst ɗo naamnal ngal waɗi. Ɗum ko caɗeele yiytugol ɗo tiiɗi e similarity tan, ngam jokkondiral maana waawaa heɓde.
- Juume format (Wrong Format): Format chunk njiytaaɗo waawi waasde fotde e LLM ngam huutoraade, walla waasde fotde e format jaabawol mo huɗo yiɗi. Misal, tableau waawi waylude e binndol, tee keɓe ɗe mbaawi waylude.
- Yaltinde keɓe waasde (Not Extracted): Keɓe ɗe sokkaaɗe mbaawi waasde yaltineede no feewi e chunk, walla LLM waawaa anndude ɗe. Ɗum waɗata ko ɓuri e jumlaaji ɗoƴƴi, tableaux e nate.
- Cuɓoraaɗe keɓe waasde (Incorrect Specificity): Jaabawol waawi wonde ɗoƴƴi haa ɓurude walla, e ɗo feewi, waawi wonde e ɗoƴƴugol haa ɓurude, tee ɗum waawaa fotde e ko huɗo sokki. Ɗoƴƴugol fannu e ɓeydude keɓe no saɗi.
- Jaabawol timmataaɗo (Incomplete): Jaabawol mo sosaaɗo e ɗo ɗo waawi waasde timmude walla hoto keɓe tan, tee ɗum waawaa hebbinde ko huɗo sokki. Sababuji mbaawi wonde haɗde jokkondirde keɓe iwɗe e chunks ɗuɗɗi, walla LLM huutoraade keɓe seɗɗa tan.
No feccere RAG, sabu o dowrowa ko e yiytugol kesɗitinal vector e indeksaaji chunk, waɗaanni caɗeele bayɗe e golle goonga: hoolaare, waawde ɓeydude e jaɓɓugol, e cellal.
2. Databaas semantik CLEVI nde waɗi e heɓde laawol e caɗeele RAG
Ngam heɓde laawol e ɗee caɗeele, CLEVI waɗi infrastrukture jeyaaɗe e humpito AI hesere, nde feewi e jokkondiral semantik, miijitagol ɗaɓɓitorde e jaabawol dow seedeeji: Clevi Semantic Database.
Ko solushon waawi waɗde sabu o woodi e dow Reasoning model mum, AI multimodal e AI model agent; ko kadi gooto e teknolooji semantik CLEVI ɗi ɓurɗi sembe, ɗi waɗata AI waawi wallude e nguurndam goonga.
So en waɗii misal, nde en waɗi databaas nde humpito woni ɗo woni ɗo, ko laawol e hoore ɗo woni e ɗoɓɓere. Databaas semantik CLEVI waawi waɗde no ɗoɓɓere woodi e hoore mum ɗoɓɓere ɓurɗo waawde. (So a naamnii ɗoɓɓere humpito ɗo a sokki, a heɓa jaabawol cello e yaawde.)
Kuutortooɓe mbaawi ɓeydude humpito keso e ɗoɓɓere nde ɓe yiɗi, kadi ɓe mbaawi noddude humpito ɗo ɓe sokki.
Kadi, ɓe mbaawi waɗde cuɓeji ndeertugol version e jamirooje naatgol e dow cuɓagol maɓɓe.
Ko ɗoɓɓere waɗata fow waɗa e log (dokument), ngam so juumre waɗii, waawde feewtinde nde woni.
<Clevi Semantic Database Structure>
Nteeriiji ɓurɗi e no teknolooji nde siryii
Dañndugol humpito semantik
- Hoto wonaa chunk vectorDB tan, jokkondire semantik hakkunde keɓe (kontekst, darnde, sababuuji e goɗɗe) ndee waɗa e graph DB
- Waawi ɓeydude e ɓeydude no knowledge graph walla semantik network
Yiytugol hybrid e miijitagol
- CTA+Context Query: Humpito (Context) e CTA ɗaɓɓitannde njogii e hakkunde
- Hybrid Retrieval: Jokkondiral yiytugol dow kesɗitinal vector e yiytugol dow laawol/rézo graph
- Intelligent Folder Hits: Yiytugol e jaɓɓugol e cuuɗi/kategoriiji ɗi jokkondiri semantik
Ɗaɓɓitorde multimodal e sahaa gooto
- TextReader Pool, VisionReader Pool e goɗɗe, fassitde e sahaa gooto keɓe ɗe ɓuri no feccere: binndol, natal, taabal, sawtu e goɗɗe
- Tawi RAG ɓuri waɗde e binndol tan, databaas semantik ɓuri waawde e ɗaɓɓitorde multimodal
Jaabawol dow seedeeji e ƴeewtagol hoolaare
- Ɗaɗiɗe Takkooje & Cite: waɗde takkooje e ceŋe ɗe nder ɗaɗiɗe e hoore mum
- Merge & Verify: jokkondirde maanaa e ƴeewtagol koolaare humpito iwɗo e ɗaɗiɗe ɗuɗɗe
- Evidence Pack: hokkude pakke jaabawuuji ɗaɓɓitaade dow dalil
Ɗaɓɓugol cuɓoraaɗo e Planner
- Planner/DAG: waɗde laawol e pelle-pelle e ɗaɓɓugol dow DAG (Directed Acyclic Graph) ngam naamne ɗaɓɓitiiɗe
Naatnude agent ɓe e cuɓoraaɗe
- cip-5-agent Supervisor: agent mawɗo jeyɗo e ɗowtugol timmunde nder ɗaɓɓugol, ɗaɓɓugol maanaa e naatnugol
3. Takkooje e keɓe cuɓoraaɗe
E fannu gooto, Semantic DB jaɓɓata keɓe e mbaydiiji ɗuɗɗi (haala, binndi, natal, widewoo, e goɗɗe), tee waawaa taƴde ɗe tan e Chunk, kono jokkondinta ɗe e jokkondiral maanaa hakkunde keɓe ɗe (ɗaɓɓugol, lewru, sabab e ko waɗi, e goɗɗe) ngam renndinde andal.
Dandii ɗum, ngam ɗum huutora jokkondiral maanaa e ɗaɓɓugol e cuɓoraaɗe, wonaa ɗaɓɓugol dow kelme walla nanndugol no RAG waɗata, aɗa waawi heɓde ɗaɓɓugol humpito e jaabawuuji AI ɓurɗi moƴƴude.
Kadi, sabu andal ngal mooftaa no graph andal walla semantic network, ɓeydude e ƴeewtaade ngal e yeeso waawi waɗeede no weeɓi.
Minen CLEVI, nder jamanu AI ɗo woni waylude mawnde, min njiytii kiite RAG, tee min mbaɗi waawde hokkude kiliyaaji humpito ɓurɗo ƴeewtugol e koolaare e law, huutoraade Semantic Database e model AI amen keeriiɗo ngam ɗaɓɓude kiite ɗe.
Sistem AI amen waawi waɗde keɓe mawɗe kiliyaaji ɓe ngoodi e nafaa, nder fannu kala e ndeer ngonka kala, so wonaa fannu ɗo ɗe njogii.
Yeeso kadi, CLEVI maa waɗa ko waawi fof ngam ɓeydude nafaa keɓe kiliyaaji e hokkude ɗeɓɓugol AI kesol e ndimaaku.
Njaaɓnuɗon aduna AI keso e CLEVI.
Naamne e ɗaɓɓugol demo: [email protected]
Copyright© 2025 Clevi Inc. Hakkunde fow kisnaaɗe.