BGE-M3
Free.ai (self-hosted)
·
embeddings
·
~100 Whakahauhau call
Ko te BGE-M3 he a tauira whakawaitara i hangaia e te BAAI. He kaha rawa i te Text embeddings, semantic search. I whakanohoia ki ngā GPUs Free.ai - e haere ana i te wātea i runga i tō tātau pūpū tohu i ia rā (100 ngā tohu i ia whakarongo). I tukua i raro i te MIT - ka whakaaetia te whakamahinga hokohoko i runga i te Free.ai.
Ka whakamahia mā te API
API OpenAI-hōu REST. Whakanao i tētahi kī, ā, ka karangatia tēnei tauira i roto i ngā waeine.
curl -X POST https://api.free.ai/v1/image/generate/ \
-H "Authorization: Bearer sk-free-..." \
-H "Content-Type: application/json" \
-d '{"model":"bge-m3","prompt":"your prompt here"}'He pātai pānui
Ka tahuri te BGE-M3 i te kupu ki tētahi rarangi mārō (he rārangi o ngā pūrere) e mau ai te tikanga. Ka whakamahia mō te rapu ā-waha, te whakarōpū, te whakarongo, te whakawhanaketanga whakarei-whakarei (RAG), me ētahi atu mahi e ōrite ana tēnei kupu ki taua kupu" e pā ana.
Ko ngā ahu tūturu ko te 384, 768, 1024, 1536 rānei e ai ki te tauira. BGE-M3 e tuku ana i te 1024-dim; OpenAI Ada e tuku ana i te 1536. Kei roto i te urupare API te ahu kia kōwhiria ai e tōna raupapa DB te taupū tika.
Ko ngā tauira whakahuatanga o nāianei (tae atu ki te nuinga o ngā kōwhiringa i runga i te Free.ai) e whakaakona ana i ngā reo 100+ me ngā mahi whakaora reo - rapu i te reo Ingarihi, ōrite ngā tuhinga ki te reo Pāniora.
512 ki te 8,192 ngā tohu i runga anō i te tauira. Ka whakaitihia ngā tāuru roa - ka whakaitihia ngā tuhinga roa ki ngā wāhanga i mua i te whakatūnga.
Ka haere a BGE-M3 ki a tātau ake GPUs, ā, ko tētahi o ngā utauta tino māmā - tata ki ngā tohu ~100 i ia whakarongo i tangohia mai i tō tātau pūrere wātea i ia rā. $5 = 200K ngā tohu.
He - POST he rārangi aho ki /v1/embeddings/ me te BGE-M3 e hoki ana ki tētahi rārangi rarangi i roto i te raupapa ōrite. Te rahi o te rōpū tae atu ki te 2,048 i ia tono.
L2-whakatikahia e te pūmau - ōritetanga cosine = hua pito. Whakapā atu 'whakatikahia = hē' mēnā e hiahiatia ana e koe ngā rarangi pūrua mō tētahi rerekētanga tawhiti.
He aha - Pinecone, Weaviate, Qdrant, Chroma, pgvector, FAISS, LanceDB. BGE-M3 e hoki ana ki ngā pūrere JSON noa iho; Kāore anō te DB kia kitea te tauira.
He - POST ki te /v1/embeddings/ me te tauira "BGE-M3". OpenAI-hōu te āhua o te urupare, nā reira ko ngā pukapuka kaimārō tīariari e mahi ana i te kore huri. He tohutoro katoa te /api/.
Ko ngā tauira ā-whāinga e pupuri ana i tō tātau kupu i runga i a tātau GPU, ā, ka whakakoretia i muri i te hokinga o te whakarongo. Ka haere te utu mā te DPA. Kāore e whakaakona e tātau i o tātau tāuru.
Sub-100ms mō te kupu poto i runga i te whakawhanaketanga, 100-500ms i runga i te utu. Ka āhua ōrite te tauine o ngā whakarongo whakatōpū - 1,000 ngā kōwae i oti i roto i te 2-10 sekona.
He - ka whakaaetia e te Free.ai te whakamahinga hokohoko o ngā whakatūtū. Ka hanga te rapu whakanao, ngā pūtautau RAG, ngā pūnaha whakarongo me te kore moni whakawhāiti.