How to choose an embedding model?
Embeddings are usually called at far higher volume than chat — the whole knowledge base gets vectorized once, then every query is encoded too, so a small unit-price gap balloons into a big monthly bill. If budget matters pick a cheap one; to save the most, self-host open BGE / Qwen3-Embedding.
Beyond price, watch two numbers: dimensions set how fine the vector is and how much your vector DB costs to store and search (1024–1536 is plenty; go 3072 for accuracy); max input caps how much fits per call — most sit near 8K so long docs need chunking, while a few like Cohere embed v4 reach 128K.
Chinese models (Tongyi, BGE) bill in CNY and handle Chinese well; for multilingual or long context, Cohere, Voyage and Jina are common. To turn this into a monthly number, plug your volume into the Token cost calculator.