text-multilingual-embedding-002 Support prompt ?

[rosh11090]

Can we provide structured SKU text (e.g., "brand: xyz title: abc : category: sdfa description: hello..") to text-multilingual-embedding-002 and then instruct the model to prioritize certain fields (brand, title, model) over others (description) for e-commerce retrieval with prompt ?? 

Gemini says its possible, but jsut want to double check if anyone has tried and knows the answer like this.

Represent this query for retrieving relevant documents: "The cat sat on the mat."

[a_aleinikov]

Hi @rosh11090 ,

Great question — and yes, text-multilingual-embedding-002 does not support prompts or instruction tuning in the way Gemini models do.

This model is not instruction-following — it's a pure embedding model, so:

• You cannot guide it to prioritize certain fields like brand or title.

• It simply embeds the full input text as-is into a vector space.

Preprocessing and field prioritization must be handled before embedding, e.g., by reordering or weighting key fields in the input string.


You can format your input like this to emphasize important fields: [BRAND] xyz [TITLE] abc [MODEL] 123 [DESCRIPTION] hello world

This gives consistent signals to the embedding model about structure, though it's not true instruction following.