new feature brainstorming - embedding API

[jeremychone]

Following this comment #28 about embeddings, here is a proposal for what the API could look like, and this discussion is about refining it.

• It should be as simple as client.embed(model, content).
• It should have the same Adapter Resolver that we have for the other models.
• The Adapter can be extended with this embed(...) method.
• Please note that genai is not in the business (for now) of knowing which model does what, so if the model does not support the API, it will fail.
• Later, we can see if there is something we can do about this.
• We might want to have a root model in src/embed, probably from the root, that will have the return type, for example. The input can just be a string, but we might want to have an input struct.
• In the adapters/... directory, we should have another file called embed.rs or something similar; this way, it's clear what we have.

These are just some initial thoughts. My hope is that we can piggyback off the chat resolution/flow but have another API or object.

Let's start the discussion to refine or redefine these points. We can just have pros/cons for things we think we might want to take a different approach on, and the goal of this discussion is to get a good, simple code example that we can put in our examples/....

CC @GustavoWidman

[GustavoWidman]

might be best if we split adapter into two different traits: completion and embedding. not all providers provide embedding, so a good indicator of whether one does or not could be the traits it implements (CompletionAdapter, EmbeddingAdapter)

[GustavoWidman]

actually nevermind, that would most likely break way too much code.

[jsantanders]

Hi, I'm a happy user of this crate and this feature would benefit me a lot. So here are my 2 cents:

I have a couple of questions about some of the implementation details:

1. What should the embedding input/output types look like?

• Input: Should it be just String or support Vec<String> for batch operations?
• Output: What's the vector format? Vec<f32>, Vec<f64>, or generic Vec<T>? because some providers like AWS Titan offers embeddings in binary format.
• Metadata: Should it include token counts, model info, etc.?

2. How should it handle different embedding dimensions across providers?

• OpenAI: Various dimensions (1536 for text-embedding-ada-002 and text-embedding-3-small, 3072 for text-embedding-3-large)
• Cohere: 1024, 4096 dimensions depending on model.
• Should the API be generic or provider-specific?

3. In mi mind the API implementation looks something like this:

// Single
let embedding = client.embed("text-embedding-3-small", "Hello world").await?;
let vector: Vec<f32> = embedding.vector;

// Batch
let embeddings = client.embed_batch("text-embedding-3-small", vec!["text1", "text2"]).await?;