1. Obuzibu bwa Catastrophic Forgetting ne RAG
AI bw'eteekebwa mu mirimu egy'amazima, ekimu ku bintu ebisinga okweraliikiriza ye mbeera y'okwerabira okw'ekitalo (Catastrophic Forgetting) ebangawo ng'ogolola okumanya kwe yasooka okuyiga ng'omuyigiriza ku data empya.
Kino kye kibeera nga AI eyesigamiziddwa ku nkolagana za neural ebangawo ng'eyiga amawulire amapya, n'efirwa amangu okumanya oba enkola bye yali eyize edda.
Okugeza, bw'ogolola LLM ey'ensibuko enzigule ng'okozesa data ya kkampuni eyeetongodde, okumanya okw'awamu oba obusobozi bw'olulimi buyinza okukendeera.
Okwewala ekizibu kino, tekinologiya wa RAG(Retrieval-Augmented Generation) akoreshwa nnyo, naye ne RAG alina obuzibu.
Obuzibu obukulu obwa RAG
- Obutakola bulungi ku data etegeke (Structured Data QA): Enkola ya RAG etereezeddwa nnyo ku biwandiiko bya lulimi ebitali bitegekeddwa, n’olwekyo tesobola kutegeera oba kukozesa bulungi makulu n’enkolagana za data etegekeddwa (emmeeza, database, spreadsheet, n’ebirala).
- Obuzibu ku PDF ezizibu/biwandiiko ebitali bitegekeddwa (Complex PDFs): PDF ezizibu (ezirimu emmeeza, empandiika ez’empandiika eziwera, ebifaananyi, n’ebirala) ziyinza okufiirwamu amawulire oba okukyusibwa kw’embeera mu nkola ya chunking (okugabanya). Kino kiyinza okuleetera data enkulu obutayingizibwa mu nkalala oba okukaluubiriza okuginoonya.
- Obutaba na model ya Fallback (Fallback Model(s)): Enkola ya RAG bw’emererwa okuzuula eky’okuddamu ekituufu, logic ey’okukyukira ku model endala (ey’okuwagira) eyinza okuba nga tebawo, oba nga tekola bulungi. Kino kireetera omukozesa obutafuna ky’alonda kimala ng’okuddamu kulemereddwa.
- Obunafu mu by’okwerinda (LLM Security): Olw’obutaba na bukuumi bumala ku bunafu bw’okwerinda obuli mu LLM yennyini (prompt injection, okubikkula data, n’ebirala), waliwo akabi k’amawulire amakulu okubikkulwa.
- Obutaba na busobozi bumala bw’okugaziya (Data Ingestion Scalability): Ebizibu by’obusobozi bw’okugaziya bitera okubaawo mu nkola y’okuteeka mu nkalala n’okutereka data nnyingi. Data bwe yeeyongera, obwangu bwa chunking, okutereka, n’okunoonya bukendeera, ate omugugu ku nkola ne gweyongera.
- Okubulwamu amawulire amakulu (Missing Content): Emirundi mingi ebikwata ku nsonga enkulu mu biwandiiko bibulamu mu nkola ya chunking oba ne bitayingizibwa mu nkalala. Kino kireetera omukozesa obutasobola kunoonya mawulire g’ayagala.
- Okubulwamu ebisinga okukwatagana ku ntikko (Missed Top Ranked): Mu bivudde mu kunoonya, chunk esinga okukwatagana mu by’amazima (ey’ekifo ekisinga obwaggulu) eyinza okubulamu. Ensonga ziyinza okuba ekkomo ly’enkola y’okunoonya, obutabeera bulungi bwa embedding, ensobi mu kuteeka mu mutendera, n’ebirala.
- Obutakwatagana na mbeera (Not in Context): Emirundi mingi chunk enooneddwa tekoosa mbeera y’ekibuuzo eky’amazima. Kino kiva ku kkomo ly’okunoonya okwesigamiziddwa ku kufaanagana kwokka, kubanga enkolagana y’amakulu eba ntono.
- Ensobi mu ngeri (Wrong Format): Enkola ya chunk enooneddwa eyinza obutaba nnungi LLM okugikwatamu, oba eyawukana ku ngeri y’okuddamu omukozesa gy’ayagala. Okugeza, emmeeza bwe ekyusibwa okuba ekiwandiiko, amawulire gayinza okukyusibwamu amakulu.
- Okulemererwa okuggya amawulire (Not Extracted): Amawulire ageetaagisa gayinza obutaggibwa bulungi mu chunk, oba LLM ne eremererwa okugategeera. Kino kitera okubaawo mu nsengeka z’ebigambo ezizibu, emmeeza, ebifaananyi, n’ebirala.
- Obunene/obuziba bw’amawulire obutasaana (Incorrect Specificity): Eky’okuddamu kiyinza okuba ekigazi nnyo ku kibuuzo, oba okukka mu buli kimu nnyo, ne kiremererwa okukwatagana n’obwetaavu bw’omukozesa. Okusalawo obunene n’obuziba bw’amawulire obutuufu kiba kizibu.
- Eky’okuddamu ekitaggwa (Incomplete): Eky’okuddamu ekisembayo ekikolebwa kiyinza obutaggwa, oba okuwaayo ekitundu ky’amawulire kyokka, ne kiremererwa okutuukiriza obwetaavu bw’omukozesa. Ensonga ziyinza okuba obutakungaanya mawulire okuva mu chunks eziwera, oba LLM okukozesa ekitundu kyokka eky’amawulire.
Mu ngeri eno, kubanga RAG yeesigamye nnyo ku kunoonya okusinziira ku kufaanagana kwa vector n’okussaamu index okusinziira ku chunk, erina obuzibu obweyoleka mu mirimu egy’amazima ku ludda lw’obwesigwa, okugaziya n’obutuufu.
2. Clevi Semantic Database eyamalawo obuzibu bwa RAG
Okumalawo obuzibu buno, CLEVI yatondawo omusingi gw’amakubo g’ amakulu, okutegeera okugatta ebintu ebingi n’okuddamu okusinziira ku bujulizi, nga ye musingi gw’amawulire ga AI ogw’omulembe oguddako ogwategekebwa ku bino: Clevi Semantic Database.
Kino kisoboka kubanga tulina Reasoning models, multimodal AI ne agentic AI models ezaffe zonna, era kye kimu ku tekinologiya ow’amaanyi ow’enjawulo owa CLEVI akasobozesa AI okuyamba ddala mu bulamu obwa nnamaddala.
Mu kifaananyi, singa database omuterekeddwamu amawulire tugiyita etterekero ly’ebitabo, osobola okulowooza nti Clevi Semantic Database erina omuterekeddwamu w’ebitabo omukugu nnyo. (Bw’osaba omuterekeddwamu amawulire ge weetaaga, ofuna eky’okuddamu ekituufu era eky’amangu.)
Abakozesa basobola okwongeramu amawulire amapya mu tterekero ly’ebitabo n’okuyita mu mawulire ge beetaaga ekiseera kyonna.
Era basobola okuteekawo engeri gye balungamyaamu enkyusa za data n’olukusa lw’okuyingira nga bwe baagala.
Buli kikolwa ky’omuterekeddwamu ebikwata ku bitabo kirekebwa mu log (wandiiko), n’olwekyo n’eky’Okujjamu ensobi bwe kibaawo, kisobola okutereezebwa.
<Clevi Semantic Database Structure>
Ebikulu ebikwata ku nkola n’enzimba ya tekinologiya
Okutereka data okusinziira ku makulu
- Si chunk vector DB yokka, wabula etereka enkolagana z’amakulu wakati wa data (emikwataganyo, emitendera, ensonga n’ebivaamu, n’ebirala) mu graph DB
- Kyangu okugaziya n’okwongerako mu ngeri ya knowledge graph/semantic network
Okunoonya n’okutegeera okugatta enkola eziwerako
- CTA+Context Query: Eteeka awamu ensonga ezikwata ku Context ne CTA mu kibuuuzo
- Hybrid Retrieval: Egatta vector similarity search n’okunoonya okusinziira ku mateeka/graph
- Intelligent Folder Hits: Enoonya yokka folders/categories ezikwatagana mu makulu
Okukola data ez’ebika eby’enjawulo mu kiseera kye kimu
- TextReader Pool, VisionReader Pool n’ebirala bisobozesa okwekenneenya mu kiseera kye kimu data ez’enjawulo nga text, ebifaananyi, emmeeza n’amaloboozi
- Wadde ng’obukola bwa RAG obwa bulijjo busobola okukola okusinga ku text yokka, semantic database esinga nnyo mu kukola data ez’ebika eby’enjawulo
Okuddamu okusinziira ku bujulizi n’okukakasa obwesigwa
- Ennyinyonnyola z’Ebiwandiiko n’Ebijuliziddwa, Okuzuula Ebivudde mu Ttebulo n’ebirala — okukola mu ngeri ey’otoma ennyinyonnyola z’ebiwandiiko n’ensibuko z’amawulire
- Merge & Verify: okugatta amakulu g’amawulire okuva mu nsibuko eziwerako n’okukakasa obwesigwa bwago
- Evidence Pack: okuwa ekipapula ky’okuddamu okwesigamiziddwa ku bujulizi
Okulowooza okw’enjawulo n’omuteesiteesi
- Planner/DAG: okukola enteekateeka ey’emitendera n’okulowooza nga kwesigamiziddwa ku DAG (Directed Acyclic Graph) ku bibuuzo ebizibu
Enteekateeka y’aba-agenti egattiddwa
- cip-5-agent Supervisor: omuwagizi omukulu addukanya era n’alungamya enkola yonna ey’okunoonya, okulowooza n’okugatta
3. Ennyinyonnyola n’okugeraageranya kw’enteekateeka
Mu bufunze, Semantic DB efuna ebikwata ku data mu ngeri ez’enjawulo (eddoboozi, ekiwandiiko, ekifaananyi, vidiyo n’ebirala), n’etegeka okumanya ng’egatta enkolagana z’amakulu wakati wa data (emirimu, emitendera, ensibuko n’ebirala), so si kugabanyamu Chunk yokka.
Ku musingi guno, kubanga ekola okunoonya n’okulowooza nga ekozesa enkolagana z’amakulu so si kunoonya okwesigamiziddwa ku bigambo ebikulu oba okufaanagana nga RAG, osobola okufuna amawulire agasinga obutuufu n’okuddamu okusinga obulungi okuva mu AI.
Era kubanga eterekebwa mu ngeri ya giraafu y’okumanya oba semantic network, kyangu okugigaziya n’okugyongerako ebintu ebipya obutasalako.
CLEVI yaffe yazuula obuzibu bwa sisitemu za RAG mu kiseera ky’enkyukakyuka ennene ey’obukugu bwa AI, era okuyita mu semantic database ne model ya AI eyakolebwa mu ngeri ey’enjawulo, kati tusobola okuddamu mangu eri bakasitoma nga tubawa data entuufu era eyesigika ennyo.
Sisitemu ya AI ya CLEVI yaffe esobola okukola data y’omuwendo ey’omugaso eri bakasitoma mu mbeera yonna, awatali kusinziira ku kika kya kitundu.
CLEVI egenda mu maaso n’okukola kyonna ekisoboka okukozesa obulungi omuwendo oguli mu data ya bakasitoma n’okubawa obumanyirivu bwa AI obuyiiya.
Mweyagalire okuwona omulembe omupya ogwa AI awamu ne CLEVI.
Okubuuza n’okusaba okulaga: [email protected]
Copyright© 2025 Clevi Inc. All rights reserved.