Add instructions for embedding

[berengerdoneux]

Hi, maybe a feature request but lots of opensource embedding models require additionnal instructions for best performance (ex. nomic-embed-text-v1.5, multilingual-e5-large-instruct).

Is there a way to provide theses instructions ?

Regards.

[prasmussen15]

Hey, so graphiti has a generic abstract EmbedderClient that is used for our actual calls. It implements embedderClient.create and embedderClient.create_batch methods. If you want to use an embedder that isn't currently supported you can define a new class that inherits from EmbedderClient and implements those methods.

For example, you could make a NomicEmbed(EmbedderClient) class and either have it instantiate the embedding instructions or use some defaults on the nomicEmbed.create call. I can help advise on this and review if you need help implementing this.

Also if you create a new working embedder client for one of these models we would also be happy to merge in your PR so others in the community can use it as well!

[berengerdoneux]

Hi, thanks for your answer.

I looked quickly at the code but I don't think it will be a quick and easy implementation. Thoses embedder generally require different prompt for document embedding vs retrieval. Actually a lot of different prompts, depending on the type of source for multilingual-e5-large-instruct https://github.com/microsoft/unilm/blob/9c0f1ff7ca53431fe47d2637dfe253643d94185b/e5/utils.py#L106

I'm not sure how to implement this but it shoud'nt be infeasible I gues.

Regards.

[claude[bot]]

@berengerdoneux Is this still relevant? Please confirm within 14 days or this issue will be closed.

[claude[bot]]

@berengerdoneux Is this still an issue? Please confirm within 14 days or this issue will be closed.

[claude[bot]]

@berengerdoneux Is this still an issue? Please confirm within 14 days or this issue will be closed.