used hugging face model

README.md

@@ -8,14 +8,14 @@ app_port: 7860
 pinned: false
 ---
 
-# Agents service (CrewAI + NVIDIA NIM)
+# Agents service (CrewAI + LiteLLM + Hugging Face)
 
 FastAPI app that exposes:
 
 - `GET /health` — liveness check
 - `POST /generate` — body: `topic`, optional `feedback`, `memory_context`, `tone_instruction`; returns `{"post": "..."}`
 
-Set **`NVIDIA_NIM_API_KEY`** as a [Space secret](https://huggingface.co/docs/hub/spaces-overview#managing-secrets) (required). Optionally set `AGENTS_LLM_MODEL` (default: `nvidia_nim/meta/llama-3.1-70b-instruct`).
+Set **`HF_TOKEN`** (or **`HUGGINGFACE_HUB_TOKEN`**) as a [Space secret](https://huggingface.co/docs/hub/spaces-overview#managing-secrets) (required). Optionally set **`AGENTS_LLM_MODEL`** (default: `huggingface/Qwen/WebWorld-32B`), **`AGENTS_LLM_TEMPERATURE`**, or **`AGENTS_LLM_BASE_URL`** for an OpenAI-compatible endpoint (see `crew.py`).
 
 **Port:** the container listens on `$PORT` when the platform sets it, otherwise **7860** (Hugging Face Spaces default).
 

crew.py

@@ -1,15 +1,18 @@
 """
 Crew definition for the agents service.
 
-This service is locked to NVIDIA NIM as its only model provider. All requests go
-through LiteLLM's `nvidia_nim/...` provider, which requires `NVIDIA_NIM_API_KEY`.
-
-Configuration via env vars:
-
-    AGENTS_LLM_MODEL         e.g. "nvidia_nim/meta/llama-3.1-70b-instruct" (default)
-    AGENTS_LLM_BASE_URL      defaults to NVIDIA's hosted inference endpoint
-    AGENTS_LLM_TEMPERATURE   float, defaults to 0.5
-    NVIDIA_NIM_API_KEY       required (or NVIDIA_API_KEY as an alias)
+LLM is provided via LiteLLM. Default Hugging Face Hub model id:
+    huggingface/Qwen/WebWorld-32B
+
+If that route fails, LiteLLM supports a Hub inference-provider prefix (only if the model's Hub page lists the
+provider for this model — do not guess the provider):
+    huggingface/<provider>/Qwen/WebWorld-32B
+Example shape (valid only when listed on the model card): huggingface/together/Qwen/WebWorld-32B
+
+Set HF_TOKEN (Hugging Face access token). Optional:
+    AGENTS_LLM_MODEL         default: huggingface/Qwen/WebWorld-32B
+    AGENTS_LLM_TEMPERATURE   float, default 0.5
+    AGENTS_LLM_BASE_URL      optional OpenAI-compatible base URL (e.g. HF Inference Endpoint)
 """
 from __future__ import annotations
 
@@ -21,37 +24,33 @@ from crewai.agents.agent_builder.base_agent import BaseAgent
 from crewai.project import CrewBase, agent, crew, task
 
 
-_NVIDIA_DEFAULT_BASE = "https://integrate.api.nvidia.com/v1"
-_NVIDIA_DEFAULT_MODEL = "nvidia_nim/meta/llama-3.1-70b-instruct"
-
-
-def _resolve_nvidia_key() -> str:
-    """Return the NVIDIA NIM key, accepting either env var name. Hard-fail if missing."""
-    key = (os.getenv("NVIDIA_NIM_API_KEY") or os.getenv("NVIDIA_API_KEY") or "").strip()
-    if not key:
+def _resolve_hf_token() -> str:
+    token = (os.getenv("HF_TOKEN") or os.getenv("HUGGINGFACE_HUB_TOKEN") or "").strip()
+    if not token:
         raise RuntimeError(
-            "Missing NVIDIA NIM key. Set NVIDIA_NIM_API_KEY (or NVIDIA_API_KEY) in the "
+            "Missing Hugging Face token. Set HF_TOKEN (or HUGGINGFACE_HUB_TOKEN) in the "
             "agents service environment."
         )
-    return key
+    return token
 
 
 def _build_llm() -> LLM:
-    model = (os.getenv("AGENTS_LLM_MODEL") or _NVIDIA_DEFAULT_MODEL).strip()
-    if not model.startswith("nvidia_nim/"):
-        raise RuntimeError(
-            f"This service only supports NVIDIA NIM models. Got model={model!r}; "
-            f"expected a value starting with 'nvidia_nim/'."
-        )
-
-    base_url = (os.getenv("AGENTS_LLM_BASE_URL") or _NVIDIA_DEFAULT_BASE).strip().rstrip("/")
+    model = (os.getenv("AGENTS_LLM_MODEL") or "huggingface/Qwen/WebWorld-32B").strip()
     temperature = float(os.getenv("AGENTS_LLM_TEMPERATURE", "0.5"))
 
-    nvidia_key = _resolve_nvidia_key()
-    os.environ["NVIDIA_NIM_API_KEY"] = nvidia_key
-    os.environ["NVIDIA_NIM_API_BASE"] = base_url + "/"
+    hf_token = _resolve_hf_token()
+    os.environ["HF_TOKEN"] = hf_token
+
+    base_url = (os.getenv("AGENTS_LLM_BASE_URL") or "").strip().rstrip("/")
+    if base_url:
+        return LLM(
+            model=model,
+            base_url=base_url,
+            api_key=hf_token,
+            temperature=temperature,
+        )
 
-    return LLM(model=model, base_url=base_url, temperature=temperature)
+    return LLM(model=model, api_key=hf_token, temperature=temperature)
 
 
 @CrewBase

requirements.txt

@@ -2,13 +2,14 @@ fastapi
 uvicorn[standard]
 python-dotenv
 
-# NVIDIA NIM is the only supported LLM provider.
-# CrewAI calls NIM through LiteLLM under the `nvidia_nim/...` provider.
+# LLM: set HF_TOKEN (or HUGGINGFACE_HUB_TOKEN). LiteLLM model ids use huggingface/<org>/<model> (see crew.py).
+# If plain huggingface/Qwen/WebWorld-32B fails, try huggingface/<provider>/Qwen/WebWorld-32B only when the Hub
+# model page lists that inference provider for this model (e.g. a listed provider, not guessed).
 crewai
 crewai-tools
 litellm
 
-# Transitively required by litellm/crewai for NIM (OpenAI-compatible client + tokenizer).
+# Transitively required by litellm/crewai (OpenAI-compatible client + tokenizer).
 openai
 tiktoken