Add Hugging Face Inference provider

[Copilot]

Adds the Hugging Face Inference API provider following the existing OpenAI architecture pattern.

Provider

HF Inference (provider-hf-inference/)

• Uses @huggingface/inference SDK (InferenceClient)
• Supports: TextGeneration, TextEmbedding, TextRewriter, TextSummary
• Optional provider field routes to specific HF inference backends (fal-ai, fireworks-ai)
• API key: provider_config.api_key or HF_TOKEN env var

Structure

The provider follows the 6-file pattern:

provider-hf-inference/
├── common/
│   ├── HFI_Constants.ts
│   ├── HFI_ModelSchema.ts
│   └── HFI_JobRunFns.ts
├── HfInferenceProvider.ts
├── HFI_Worker.ts
└── index.ts

Integration

• Added exports to src/index.ts (constants, schemas, provider classes)
• Added ./hf-inference to package.json exports
• Added @huggingface/inference as optional peer dependency
• Updated build scripts for new entry point

Usage

import { HF_INFERENCE, HfInferenceProvider } from "@workglow/ai-provider";
import { HFI_TASKS } from "@workglow/ai-provider/hf-inference";

const model = {
  provider: "HF_INFERENCE",
  provider_config: {
    model_name: "meta-llama/Llama-3.3-70B-Instruct",
    provider: "fireworks-ai" // optional routing
  }
};

await new HfInferenceProvider(HFI_TASKS).register({ mode: "inline" });

Original prompt

---

This section details on the original issue you should resolve

<issue_title>Add HF and featherless inference</issue_title>
<issue_description>---
name: Add AI Providers
overview: Create two new AI providers (Hugging Face Inference API and Featherless.ai) following the established OpenAI provider pattern, each with constants, model schema, task implementations, provider class, worker, and barrel exports.
todos:

• id: hf-inference-provider
  content: "Create the Hugging Face Inference provider (6 files: constants, model schema, job run fns, provider class, worker, index)"
  status: pending
• id: featherless-provider
  content: "Create the Featherless.ai provider (6 files: constants, model schema, job run fns, provider class, worker, index)"
  status: pending
• id: integrate-exports
  content: Update src/index.ts barrel exports and package.json (exports, dependencies, build scripts)
  status: pending
  isProject: false

---

Add Hugging Face Inference and Featherless.ai Providers

Both providers will follow the exact same architecture as the existing OpenAI provider in libs/packages/ai-provider/src/provider-openai/, with each consisting of 6 files organized into a directory with a common/ subdirectory.

Architecture (per provider)

Each provider directory contains:

• common/<PREFIX>_Constants.ts - Provider name constant
• common/<PREFIX>_ModelSchema.ts - Model config schema (provider discriminator, provider_config shape)
• common/<PREFIX>_JobRunFns.ts - Task run/stream functions + task registries
• <Name>Provider.ts - Provider class extending AiProvider
• <Name>_Worker.ts - Worker registration function
• index.ts - Barrel exports

Both providers support the same 4 task types as OpenAI: TextGenerationTask, TextEmbeddingTask, TextRewriterTask, TextSummaryTask.

---

Provider 1: Hugging Face Inference (provider-hf-inference/)

SDK: @huggingface/inference (uses InferenceClient)

• Constant: HF_INFERENCE = "HF_INFERENCE"
• Model config: { provider: "HF_INFERENCE", provider_config: { model_name, api_key?, provider? } }
  • provider is an optional field to route to specific HF inference providers (e.g. "fal-ai", "fireworks-ai")
• Task implementations use lazy import("@huggingface/inference") and call:
  • client.chatCompletion(...) for text generation, rewriting, summarization
  • client.featureExtraction(...) for embeddings
• Streaming uses client.chatCompletionStream(...) yielding text-delta events

Key file: common/HFI_JobRunFns.ts -- pattern mirrors provider-openai/common/OpenAI_JobRunFns.ts but swaps the OpenAI SDK calls for HF Inference SDK calls.

---

Provider 2: Featherless.ai (provider-featherless/)

SDK: openai (Featherless is OpenAI-compatible, base URL: https://api.featherless.ai/v1)

• Constant: FEATHERLESS_AI = "FEATHERLESS_AI"
• Model config: { provider: "FEATHERLESS_AI", provider_config: { model_name, api_key? } }
• Task implementations are nearly identical to OpenAI but:
  • Use FEATHERLESS_API_KEY env var fallback
  • Default baseURL to https://api.featherless.ai/v1
  • No organization field
  • No embedding support (Featherless doesn't support /v1/embeddings), so TextEmbeddingTask will be excluded

Key file: common/Featherless_JobRunFns.ts -- closely mirrors provider-openai/common/OpenAI_JobRunFns.ts with adjusted base URL and API key env var.

---

Integration Changes

libs/packages/ai-provider/src/index.ts

Add 6 new export lines (constants, model schemas, provider classes) for both providers.

libs/packages/ai-provider/package.json

• Add exports entries: "./hf-inference" and "./featherless"
• Add @huggingface/inference to peerDependencies (optional), peerDependenciesMeta, and devDependencies
• openai is already a peer dependency (reused by Featherless)
• Update build-code and watch-code scripts to include the two new entry points

---

File Inventory (14 new files)

Hugging Face Inference (7 files):

• src/provider-hf-inference/common/HFI_Constants.ts
• src/provider-hf-inference/common/HFI_ModelSchema.ts
• src/provider-hf-inference/common/HFI_JobRunFns.ts
• src/provider-hf-inference/HfInferenceProvider.ts
• src/provider-hf-inference/HFI_Worker.ts
• src/provider-hf-inference/index.ts

Featherless.ai (7 files):

• src/provider-featherless/common/Featherless_Constants.ts
• src/provider-featherless/common/Featherless_ModelSchema.ts
• src/provider-featherless/common/Featherless_JobRunFns.ts
• src/provider-featherless/FeatherlessProvider.ts
• src/provider-featherless/Featherless_Worker.ts
• `src/provider-feather...

• Fixes Add HF and featherless inference #214

---

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.

[sroussey]

@copilot remove the featherless related code

[Copilot]

@copilot remove the featherless related code

Removed all Featherless-related code. The PR now only includes the HF Inference provider. Changes in commit 8d6a062.