Add Atlas Cloud LiteLLM provider alias

[binyangzhu000-sudo]

Summary

• add an atlascloud/ model alias that routes PageIndex LiteLLM calls through Atlas Cloud's OpenAI-compatible endpoint
• read ATLASCLOUD_API_KEY and optional ATLASCLOUD_API_BASE only when Atlas Cloud models are selected
• document a minimal Atlas Cloud config example and add focused unit coverage for the model mapping

Validation

• .venv/bin/python -m compileall pageindex
• .venv/bin/python -m pytest tests/test_atlascloud_litellm.py tests/test_issue_163.py
• git diff --check
• verified Atlas Cloud public model catalog contains qwen/qwen3.5-flash and deepseek-ai/deepseek-v4-pro

No sponsor, logo, credits, or partner sections were added.

[binyangzhu000-sudo]

Resolved the merge conflicts with the current main in 9d6af7c. The resolution preserves the new direct OpenAI SDK routing while keeping atlascloud/... models on LiteLLM with the Atlas Cloud endpoint and API key.

Validation:

• python 3.11 -m pytest -q tests (25 passed)
• git diff upstream/main --check

[binyangzhu000-sudo]

Synced with main (56 commits behind) and rebuilt this on the reworked LLM lane rather than restoring the old shape.

What upstream changed, and how this adapts

• The direct OpenAI-SDK route (_is_openai_model, use_openai_sdk) is gone and everything goes through LiteLLM now. I took that deletion as-is — the Atlas support no longer rides on that path. (Checked it was yours to delete, not mine to keep: _is_openai_model was already in this PR's merge base, while prepare_litellm_call was the part I'd added.)
• prepare_litellm_call() returned (model, kwargs), which no longer matches anything. Split it to fit your shape instead: _litellm_model() gets an atlascloud/ branch that normalizes to the openai/ wire form, and a new _atlascloud_kwargs() supplies api_base/api_key. llm_completion and llm_acompletion merge those before _llm_backend, so an explicit backend override still wins.
• Two test assertions updated for upstream behaviour changes, not for my code: bare model names now get the openai/ prefix, and completion kwargs carry max_retries: 0 instead of temperature: 0.

README: taken wholesale from upstream. My old snippet used the retired model:/retrieve_model: YAML keys and the page is now SDK-oriented with no provider-configuration section — dropping it in anyway would have been noise. The runnable example stays in pageindex/config.yaml, which auto-merged onto your current index_model/chat_model naming.

Verification — pytest tests/: 330 passed, 162 skipped, including this PR's 8 Atlas cases.

Disclosure: I work at Atlas Cloud.