[inference] Add DeepInfra sentence-similarity provider helper

[ovuruska]

What

Adds a sentence-similarity task helper for the DeepInfra inference provider.

DeepInfra serves its embedding models (bge, e5, gte, sentence-transformers, nvidia Nemotron-Embed, etc.) through an OpenAI-compatible POST /v1/openai/embeddings endpoint. Many of these repos carry pipeline_tag: sentence-similarity on the Hub, so a DeepInfra mapping can only be registered under the sentence-similarity task. Until now there was no sentence-similarity helper for any third-party provider (only hf-inference), so those mappings could not route.

How

DeepInfraSentenceSimilarityTask:

• route → v1/openai/embeddings
• payload → embeds [source_sentence, ...sentences] in a single request (source first); model is applied last so caller parameters cannot override the mapped provider model.
• response → the embeddings preserve request order via index; item 0 is the source, items 1.. are the comparison sentences. We compute the cosine similarity of each sentence against the source and return number[], matching SentenceSimilarityOutput. This mirrors the task definition ("by comparing their embeddings").

Registered under the deepinfra provider in getProviderHelper.

Tests

• packages/inference/test/deepinfra-sentence-similarity.spec.ts — offline unit tests: route, payload ordering, cosine computation, response index ordering, malformed-response handling (5 tests, all passing).
• An integration test in the existing DeepInfra block of InferenceClient.spec.ts.

Validation: vitest run (unit) ✅, tsc --noEmit ✅, prettier --check ✅, eslint ✅.

---

Note

Low Risk
Additive provider task with local similarity math and tests; no changes to auth or existing DeepInfra task behavior beyond a minor formatting tweak.

Overview
Adds DeepInfra routing for the sentence-similarity task so Hub embedding models mapped to DeepInfra can be used with InferenceClient.sentenceSimilarity.

DeepInfraSentenceSimilarityTask calls v1/openai/embeddings with [source_sentence, ...sentences] in one request, sorts embeddings by API index, and returns cosine similarities vs. the source as number[] (same shape as SentenceSimilarityOutput). The helper is registered on the deepinfra provider in getProviderHelper.

Coverage includes a dedicated unit spec (deepinfra-sentence-similarity.spec.ts) and a DeepInfra integration case in InferenceClient.spec.ts (e.g. BAAI/bge-m3).

^{Reviewed by Cursor Bugbot for commit 999bc5e. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want higher recall? High effort reviews run extra passes and find more bugs. A team admin can switch effort levels in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit 999bc5e. Configure here.}

[cursor]

Snippet payload crashes on string inputs

Medium Severity

preparePayload destructures inputs as { source_sentence, sentences } and spreads sentences. Default snippet generation passes getModelInputSnippet as a JSON string, so sentences is undefined and spreading it throws. That breaks DeepInfra sentence-similarity snippet generation (and any caller that does not pass a structured object).

 

^{Reviewed by Cursor Bugbot for commit 999bc5e. Configure here.}