fix: add --base_url CLI option for custom LLM endpoints

[octo-patch]

Fixes #224

Problem

Users running local LLM servers (Ollama, vLLM, LM Studio, etc.) have no way to specify a custom API base URL through the CLI. The OPENAI_API_BASE environment variable works when set manually, but there is no --base_url argument, making the workflow unnecessarily awkward.

Solution

• run_pageindex.py: Added --base_url argument that sets os.environ["OPENAI_API_BASE"] before any LLM calls are made.
• pageindex/utils.py: Updated llm_completion and llm_acompletion to read OPENAI_API_BASE from the environment and forward it to litellm as api_base, so all LLM calls in the pipeline respect the custom endpoint.

The change is fully backward-compatible: if --base_url is not provided and OPENAI_API_BASE is not set, behaviour is identical to before.

Usage

# Ollama
python run_pageindex.py --pdf_path doc.pdf \
  --model ollama/llama3 \
  --base_url http://localhost:11434

# vLLM / LM Studio or any OpenAI-compatible server
python run_pageindex.py --pdf_path doc.pdf \
  --model openai/my-model \
  --base_url http://localhost:8000/v1

Testing

Verified that the argument is parsed correctly and OPENAI_API_BASE is set before any LLM call. The api_base kwarg is only injected into litellm when the env var is present, so there is no overhead for standard OpenAI usage.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[saccharin98]

Thanks for the contribution.

I think we should be careful about the scope here. If the goal is to let users choose an LLM provider, this is already supported through LiteLLM model strings.
run_pageindex.py already accepts --model, and the LLM helpers pass that model through to litellm.completion() / litellm.acompletion(), so users can select providers
with values like anthropic/..., gemini/..., ollama/..., openai/..., etc. In that sense, adding another provider-selection path would be redundant.

For issue #224 specifically, the missing piece is not provider selection, but a custom OpenAI-compatible base URL. That is a valid feature request, but I do not think this PR
is ready as-is:

1. The new --base_url only applies to the run_pageindex.py CLI path. It does not cover the Python SDK / PageIndexClient local-mode path, so the configuration surface
   would be inconsistent.
2. Reading OPENAI_API_BASE inside every llm_completion() / llm_acompletion() call and unconditionally forwarding it as api_base affects all LiteLLM providers, not just
   OpenAI-compatible endpoints. That can be surprising when users are using non-OpenAI providers via LiteLLM.
3. There are no tests for the new behavior. At minimum, we should mock litellm.completion and litellm.acompletion and verify that api_base is omitted by default and
   passed only when explicitly configured.

I would suggest either closing this as redundant if the intent is provider selection, or reworking it as a proper custom endpoint feature wired through the shared config/API
layer rather than only through a CLI env-var side effect.