EmbeddingGemma 2 Uyasebenza

I-EmbeddingGemma 2 iyimodeli ye-Embedding yomdabu enemodi eziningi, evulekile yakwa-Google, esekelwe ekwakhiweni kwe-Gemma 4. Imephesha umbhalo, ikhodi, izithombe, amavidiyo nomsindo endaweni eyodwa yamavekhtha, ukuze kwenziwe ukusesha kwencazelo phakathi kwezinhlobo ezahlukene zemidiya. I-EmbeddingGemma 2 yakhelwe ngendlela eyimodulayo; yonke imodi inamapharamitha angu-740M kuphela, inewindi lomongo elingu-8K token, futhi ngemva kokucindezelwa ithatha cishe u-567MB wememori ocingweni. I-EmbeddingGemma 2 iyimodeli enamandla kunazo zonke yamavekhtha enemodi eziningi engaphansi kwe-1B, kanti ukusesha ikhodi kuthuthuke cishe ngamaphuzu angu-10 uma kuqhathaniswa nesizukulwane esedlule. Ilungele i-RAG esebenza ocingweni, ukusesha kwencazelo, ukusesha ama-albhamu ezithombe nezinye izimo zokusetshenziswa.

Ukusesha kwencazelo phakathi kwamamodi: Imephesha umbhalo, ikhodi, izithombe, amavidiyo nomsindo endaweni efanayo yamavekhtha, futhi isekela ukusebenzisa noma iyiphi imodi ukusesha amanye amamodi.
Ukusebenza kwasendaweni ocingweni: Yonke imodi inamapharamitha angu-740M kuphela. Ngemva kokucindezelwa, ithatha cishe u-567MB wememori ku-Pixel 11 Pro, futhi ingenza i-inference ocingweni noma kwi-laptop ngaphandle kokuxhumeka ku-inthanethi.
Ipayipi le-RAG lasendaweni: Isebenzisa i-tokenizer ne-encoder yomsindo efanayo ne-Gemma 4. Lezi zinhlobo ezimbili zingahlanganiswa ukuze kwakhiwe isistimu ye-RAG enemodi eziningi endaweni, kugcinwe ubumfihlo bedatha.
Ukusesha kwama-albhamu nemidiya ngencazelo: Ku-Instant Media Search ye-Google AI Edge Gallery, ungafaka umbhalo ukuze useshe izithombe eziku-albhamu yocingo, noma usebenzise ikhamera ukuthatha isithombe sento bese usesha izithombe ezifanayo.
Ukuthola izikhathi kumavidiyo: I-Video Moments Finder isekela imibuzo yombhalo noma yomsindo, futhi imaka ngokuzenzakalelayo izingxenye zevidiyo ezihambisana nencazelo.
Injini yezinqumo yesikhathi sangempela: Nge-MediaPipe Decision Task API, yenza ukuhlukanisa, ukuqondisa nokubikezela ngesikhathi sangempela. Ekubonisweni, isinqumo esisodwa sithatha ama-millisecond angu-43 kuphela, okushesha kakhulu kunama-millisecond angu-640 e-LLM.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokuxhumana lamaphrojekthi e-AI avulekile
Indawo ehlanganisiwe yokushumeka enemodi eziningi: Imodeli ilandela indlela yobuchwepheshe yochungechunge lwe-Gemini Embedding, ifake okokufaka kwamamodi ahlukene endaweni efanayo yamavekhtha enobukhulu obuphezulu. Okuqukethwe okunencazelo esondelene kuba seduze endaweni yamavekhtha, kanti okungenabudlelwano bencazelo kuba kude. Ukusesha phakathi kwamamodi kungancishiselwa ekusesheni komakhelwane abaseduze bamavekhtha, ngaphandle kokugcina imodeli ehlukile yemodi ngayinye.
Ukwakhiwa kwamamoduli ama-encoder nama-decoder: Imodeli yakhiwe phezu kwe-Gemma 4 futhi ihlanganisa izingxenye ezintathu: i-text backbone enamapharamitha angu-270M, i-visual encoder enamapharamitha angu-170M, kanye ne-audio encoder enamapharamitha angu-300M. Amamoduli okubona nawomsindo angalayishwa lapho kudingeka, ngakho ekusetshenzisweni kunezinhlobo ezine: umbhalo kuphela, umbhalo+izithombe (440M), umbhalo+umsindo, kanye nemodi eziningi. Imisebenzi yombhalo ayidingi izindleko zememori zamamoduli okubona nawomsindo.
Ukwabelana ngezinto ne-Gemma 4: I-EmbeddingGemma 2 nemodeli ekhiqizayo i-Gemma 4 zisebenzisa i-text tokenizer ne-audio encoder efanayo. Zingaxhunywa ukuze zakhe ipayipi le-RAG elihlanganisiwe elisebenza ocingweni. I-EmbeddingGemma 2 ifaka amafayela endawo anemodi eziningi enkombeni, kanti i-Gemma 4 yenza i-inference kumongo otholakele. Ukwabelana ngezinto kunciphisa ukusetshenziswa kwememori okuphelele lapho zihlanganiswa.
Ukufunda izethulo ze-Matryoshka nokunquma okuguquguqukayo: Imodeli isebenzisa ukuqeqeshwa kwe-MRL, ikhiphe amavekhtha anobukhulu obungu-768, futhi isekele ukunqunywa kube ngu-512, 256 noma 128 ngesikhathi se-inference. Uma kunqunywa kube ngu-256, ukusebenza cishe akulahleki, kanti ukugcinwa kwenkomba yamavekhtha kungancipha kube ingxenye eyodwa kwezintathu noma ngaphansi, kuze kufike kokukodwa kokuyisithupha kosayizi wokuqala.
5. Umongo omude nokusikwa kokokufaka okude: Imodeli isekela iwindi lomongo elingu-8K token, eliphindwe kane kunesizukulwane esedlule. Ingacubungula ngesikhathi esisodwa cishe imizuzu engu-5.5 yomsindo, izithombe ezingu-29, amafreyimu evidiyo angu-58 noma inhlanganisela eshintshanayo yalokhu. Okokufaka kombhalo kusekela neziqalo zemisebenzi, ukuze umbhalo ofanayo ukhiqize izethulo zamavekhtha ezahlukene zemisebenzi ehlukene.

Landa izisindo zemodeli: Thola amafayela emodeli ka-google/embeddinggemma-2 ku-Hugging Face noma e-Kaggle. Uma udinga inguqulo elungiselelwe ukusebenza ocingweni, vakashela ikhasi le-LiteRT Community.
Khetha uhlaka lokusebenza: Ungalayisha imodeli ngokuqondile ngamathuluzi ajwayelekile afana ne-transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama ne-LM Studio.
Bhala ikhodi ye-inference: Usebenzisa i-sentence-transformers njengesibonelo, layisha imodeli bese usebenzisa i-prompt_name="SearchQuery" ne-"Document" ukuze ufake ikhodi embuzweni nasembhalweni, bese ubiza i-similarity() ukubala ukufana kwamavekhtha.
Yakha isistimu yokusesha: Gcina i-embedding ekhiqiziwe kusizindalwazi samavekhtha njenge-Qdrant ukuze wenze ukusesha kwencazelo, ukukhomba ikhodi noma ukusesha okunemodi eziningi.
Ukusetshenziswa ocingweni: Edivayisini ephathwayo, sebenzisa i-Google AI Edge MediaPipe ukuze uhlanganise imisebenzi yokushumeka, ukusesha nezinqumo, noma usebenzise i-LiteRT ukuhlanganisa ngendlela oyifisayo. Esipheqululini, sebenzisa i-transformers.js / WebGPU.
Yakha i-RAG yasendaweni: Yisebenzise ne-Gemma 4, lapho i-EmbeddingGemma 2 iphethe ukusesha amafayela endawo bese i-Gemma 4 yenza i-inference kumongo. Ukwabelana ngezinto kunciphisa ukusetshenziswa kwememori okuphelele.

Ukuhlanganiswa komdabu kwawo wonke amamodi: Imodeli eyodwa ihlanganisa umbhalo, ikhodi, izithombe, amavidiyo nomsindo, ikufake endaweni efanayo yamavekhtha, ithathe indawo yendlela yakudala yokusebenzisa imodeli eyodwa ngemodi ngayinye.
Ilula kakhulu futhi iyasebenza ocingweni: Yonke imodi inamapharamitha angu-740M kuphela. Ngemva kokucindezelwa, ithatha cishe u-567MB wememori ku-Pixel 11 Pro, futhi ingenza i-inference enemodi eziningi ungaxhunyiwe ku-inthanethi.
Ukusebenza okuphezulu ngosayizi ofanayo: Iyimodeli ye-Embedding enemodi eziningi enamandla kunazo zonke engaphansi kwepharamitha engu-1B, futhi kweminye imisebenzi idlula ngisho namamodeli akhethekile aphindwe ngaphezu kokuphindwe kabili ngosayizi.
Ikhono lekhodi lithuthuke kakhulu: Isikolo sokusesha ikhodi sikhuphuke cishe ngamaphuzu angu-10 uma kuqhathaniswa nesizukulwane esedlule, okuyenza ilungele ukukhomba ikhodi yasendaweni nokusesha kwe-coding agent.
Ukulayisha amamoduli lapho kudingeka: Ama-encoder okubona nawomsindo angasuswa ngokuzimela. Imisebenzi yombhalo idinga u-270M kuphela, kugwenywe izindleko zememori zamamodi angadingeki.
Izindleko zokugcina ziphansi: Ngokusekelwe ekufundeni izethulo ze-Matryoshka, amavekhtha angu-768-dimensional anganqunywa abe ngu-128, futhi ukugcinwa kwenkomba yamavekhtha kungancipha kuze kufike kokukodwa kokuyisithupha kosayizi wokuqala.
Ukusekelwa komongo omude: Iwindi lomongo elingu-8K token liphindwe kane kunesizukulwane esedlule, futhi lingacubungula ngesikhathi esisodwa cishe imizuzu engu-5.5 yomsindo noma izithombe ezingu-29.
Ukusebenzisana kwemvelo ne-Gemma 4: Zabelana nge-tokenizer ne-audio encoder, ngakho ukwakha i-RAG yasendaweni ngazo zombili kudinga imemori encane iyonke.

Iwebhusayithi yephrojekthi: https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/
Ilabhulali yamamodeli ye-HuggingFace: https://huggingface.co/google/embeddinggemma-2
Iphepha lobuchwepheshe: https://ai.google.dev/gemma/docs/embeddinggemma/model_card_2

Ukusesha kwencazelo enemodi eziningi ocingweni: Sesha imidiya yasendaweni ku-albhamu yocingo usebenzisa umbhalo noma izithombe ozithathe ngekhamera, konke kwenziwa ungaxhunyiwe ku-inthanethi ngaphandle kokulayisha idatha eyimfihlo.
Ukusesha okuqukethwe kwamavidiyo nokuthola izikhathi: Kumavidiyo okuqapha, ezifundweni eziku-inthanethi noma ekuqopheni imihlangano, sebenzisa umbhalo noma umsindo ukuze uthole ngokushesha izingxenye ezithile, njengendawo “yofudu ludla”, ngaphandle kokusesha uhlaka ngalunye ngesandla.
Isizindalwazi solwazi se-RAG yasendaweni: Hlanganisa ne-Gemma 4 ukuze wakhe isistimu ye-RAG esebenza ungaxhunyiwe ku-inthanethi, ukhombe amadokhumenti, izithombe nomsindo wasendaweni ndawonye. Ilungele amanethiwekhi angaphakathi ezinkampanini nezindawo eziqukethe ulwazi oluyimfihlo ezingavunyelwe ukulayishwa efwini.
Ukukhomba ikhodi nokusesha kwe-Agent: Yakha inkomba yamavekhtha yekhodi yasendaweni ukuze usekele ukusesha kwekhodi ngencazelo ku-coding agent. Isikolo sokusesha ikhodi sikhuphuke cishe ngamaphuzu angu-10 uma kuqhathaniswa nesizukulwane esedlule.