feat:allow api_base (base_url) in .env

[saccharin98]

• ## Summary

Adds lightweight API base URL support for OpenAI-compatible providers through LiteLLM.

Changes

• Pass OPENAI_BASE_URL / OPENAI_API_BASE / CHATGPT_API_BASE to LiteLLM as api_base
• Add --base-url / --api-base CLI options in run_pageindex.py
• Document OpenAI-compatible/Ollama usage in README
• Add mock tests for sync and async LiteLLM API base forwarding

Test Plan

• python3 -m pytest tests/test_litellm_api_base.py
• python3 -m pytest tests/test_config.py tests/test_litellm_api_base.py
• python3 run_pageindex.py --help
• git diff --check

Closes #229

[rejojer]

Code review

No issues found. Checked for bugs and CLAUDE.md compliance.

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}

[saccharin98]

Code review

Found 1 issue:

1. Unconditional api_base applied to all LiteLLM providers — _litellm_api_base_kwargs() passes the api_base parameter to every litellm.completion() / litellm.acompletion() call regardless of the model or provider. When OPENAI_BASE_URL is set (e.g. for Ollama or a local endpoint), non-OpenAI models like anthropic/claude-sonnet or gemini/... will also be routed to that endpoint, causing failures. Before this PR, no api_base was passed, so multi-provider usage worked correctly. The fix should conditionally apply api_base only for OpenAI-compatible models/providers.

PageIndex/pageindex/index/utils.py

Lines 13 to 21 in 11b8cd7

	def _litellm_api_base_kwargs():
	api_base = (
	os.getenv("OPENAI_BASE_URL")
	or os.getenv("OPENAI_API_BASE")
	or os.getenv("CHATGPT_API_BASE")
	)
	return {"api_base": api_base} if api_base else {}

PageIndex/pageindex/utils.py

Lines 26 to 34 in 11b8cd7

	def _litellm_api_base_kwargs():
	api_base = (
	os.getenv("OPENAI_BASE_URL")
	or os.getenv("OPENAI_API_BASE")
	or os.getenv("CHATGPT_API_BASE")
	)
	return {"api_base": api_base} if api_base else {}

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}

[saccharin98]

Review

Reviewed the current head (60d3184). No blocking issues found.

The earlier provider-routing concern looks addressed now: _litellm_api_base_kwargs(model) scopes OPENAI_BASE_URL / OPENAI_API_BASE / CHATGPT_API_BASE to OpenAI-compatible models instead of passing api_base to every LiteLLM provider. That should avoid incorrectly routing explicit non-OpenAI models such as anthropic/... and gemini/... through the OpenAI-compatible host.

The added regression coverage also checks the important cases:

• bare OpenAI model still receives api_base
• openai/... model still receives api_base
• anthropic/... and gemini/... do not receive api_base
• the legacy pageindex.utils path is covered

I also ran python3 -m pytest tests/test_litellm_api_base.py locally and it passed (8 passed).

[rejojer]

Code review

Found 1 issue:

1. PageIndexClient.__init__ does not accept a base_url parameter, so SDK users (local mode) must still set OPENAI_BASE_URL in their process env manually — while the CLI path mutates os.environ["OPENAI_BASE_URL"] on the user's behalf in run_pageindex.py. This configuration-surface inconsistency was concern initial run, key error #1 in the blocking review on predecessor PR fix: add --base_url CLI option for custom LLM endpoints #229 ("does not cover the Python SDK / PageIndexClient local-mode path, so the configuration surface would be inconsistent", fix: add --base_url CLI option for custom LLM endpoints #229 (comment)) and remains unaddressed here.

PageIndex/pageindex/client.py

Lines 41 to 49 in 60d3184

	def __init__(self, api_key: str = None, model: str = None,
	retrieve_model: str = None, storage_path: str = None,
	storage=None, index_config: IndexConfig | dict = None):
	if api_key:
	self._init_cloud(api_key)
	else:
	self._init_local(model, retrieve_model, storage_path, storage, index_config)

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}