Add Nvidia inference specification #5794
Conversation
Jan-Kazlouski-elastic
added
specification
ml
skip-backport
|
Following you can find the validation changes against the target branch for the API.
You can validate this API yourself by using the |
# Conflicts: # output/openapi/elasticsearch-openapi.json # output/openapi/elasticsearch-serverless-openapi.json # output/schema/schema.json
package.json
Outdated
| }, | ||
| "dependencies": { | ||
| "@redocly/cli": "^1.34.5" | ||
| "@redocly/cli": "^1.34.6" |
There was a problem hiding this comment.
I don't think this should be getting changed here.
There was a problem hiding this comment.
This is done by make setup & make contrib execution. It is mandatory to execute that before merging. Changing it to ^1.34.5 results in it going back to ^1.34.6 as soon as make setup is called.
| "rerank", | ||
| "text_embedding", | ||
| "completion", | ||
| "chat_completion" |
There was a problem hiding this comment.
Nitpick, but could these be in alphabetical order?
There was a problem hiding this comment.
Done.
| */ | ||
| model_id: string | ||
| /** | ||
| * For a `text_embedding` task, the maximum number of tokens per input before chunking occurs. |
There was a problem hiding this comment.
This should be "For a `text_embedding` task, the maximum number of tokens per input. Inputs exceeding this value are truncated prior to sending to the Nvidia API."
This is wrong almost everywhere in the docs; there's an issue describing some of the problems with max_input_tokens.
There was a problem hiding this comment.
Fixed.
| text_embedding, | ||
| completion, | ||
| chat_completion, | ||
| rerank |
There was a problem hiding this comment.
For consistency, could these be in alphabetical order?
There was a problem hiding this comment.
Fixed.
| */ | ||
| input_type?: NvidiaInputType | ||
| /** | ||
| * For a `text_embedding` task, the method to handle inputs longer than the maximum token length. |
There was a problem hiding this comment.
To help differentiate this from max_input_tokens it might be better to word it like "the method used by the Nvidia model to handle inputs longer than..."
There was a problem hiding this comment.
Good thinking. Fixed.
| /** | ||
| * The URL of the Nvidia model endpoint. | ||
| */ |
There was a problem hiding this comment.
Would it be helpful to include the default URLs for each task type if url isn't specified?
There was a problem hiding this comment.
Good idea. Added.
| text_embedding, | ||
| chat_completion, | ||
| completion, | ||
| rerank |
There was a problem hiding this comment.
For consistency, could these be in alphabetical order?
There was a problem hiding this comment.
Done.
|
Hi @DonalEvans |
| mistral | ||
| } | ||
|
|
||
| export class NvidiaServiceSettings { |
There was a problem hiding this comment.
There should also be a dimensions field documented for the text_embedding task, sorry for missing this in the first review.
There was a problem hiding this comment.
I don't think we should declare the parameters the user can't specify.
According to your suggestion from this comment - dimensions cannot be set during the creation of an endpoint and are not sent to the Nvidia
There was a problem hiding this comment.
Urgh, my mistake, sorry for forgetting about that, I've been up to my eyeballs in the Jina text embedding code lately (which does support specifying dimensions) and the context switch tripped me up.
| * * `ingest`: Mapped to Nvidia's `passage` value in request. Used when generating embeddings during indexing. | ||
| * * `search`: Mapped to Nvidia's `query` value in request. Used when generating embeddings during querying. | ||
| * | ||
| * IMPORTANT: If not specified `input_type` field in request to Nvidia endpoint is set as `query` by default. |
There was a problem hiding this comment.
This would be better as "For Nvidia endpoints, if the `input_type` field is not specified, it defaults to `query`."
There was a problem hiding this comment.
Fixed.
| */ | ||
| max_input_tokens?: integer | ||
| /** | ||
| * For a `text_embedding` task, the similarity measure. One of cosine, dot_product, l2_norm. |
There was a problem hiding this comment.
It would be good to add that if no similarity measure is specified, the default value is dot_product.
There was a problem hiding this comment.
Added.
# Conflicts: # output/openapi/elasticsearch-openapi.json # output/openapi/elasticsearch-serverless-openapi.json # output/schema/schema.json
This PR adds changes to specification caused by elastic/elasticsearch#132388
Additional actions