Thank you for your comment.
I’m looking forward to seeing more ways to run LLMs.
I usually use Japanese (a non-Latin language), so I use the EMBED model, which is also excellent for non-Latin languages.
As you can see from the tables below, Cohere:embed-multilingual-v3.0 is a better choice for external services, and Hugging Face:multilingual-e5-large-embedding is a better choice for local. So, Hugging Face:multilingual-e5-large-embedding is the best choice in terms of cost for me.
The table below shows the rank at which the three required articles were found by the vector search among 103 chunks.
in English
Out[]=
embed Model | service | Article 45 | Article 46 | Article 102 |
Wolfram: SentenceBERT | local | 3 | 18 | 4 |
Hugging Face: multilingual-e5-large-embedding | local | 1 | 4 | 3 |
Cohere: embed-multilingual-v3.0 | external | 1 | 3 | 2 |
OpenAI: text-embedding-ada-002 | external | 2 | 3 | 1 |
in Japanese
Out[]=
embed Model | service | Article 45 | Article 46 | Article 102 |
Wolfram: SentenceBERT | local | 80 | 86 | 17 |
Hugging Face: multilingual-e5-large-embedding | local | 3 | 2 | 1 |
Cohere: embed-multilingual-v3.0 | external | 3 | 2 | 1 |
OpenAI: text-embedding-ada-002 | external | 9 | 4 | 2 |