feat: Universal Local LLM & Custom Endpoint Support (Ollama, LM Studio, vLLM, etc.)

[sreenivas134]

This PR introduces comprehensive support for local LLMs and custom API endpoints. While existing proposals often focus on specific providers (like Ollama), this implementation leverages LiteLLM's full capabilities to provide a universal, config-driven interface that works with any OpenAI-compatible or local provider (LM Studio, vLLM, LocalAI, Ollama, etc.) with minimal configuration.

✨ Key Features

• Universal api_base Support: Adds a global api_base attribute that can be configured via config.yaml, environment variables, or directly through the CLI.
• Automated Provider Normalization: Implements intelligent model-name routing. If a custom endpoint is detected (like LM Studio at localhost:1234), the system automatically ensures the correct openai/ prefix is applied, removing the need for users to manually prefix their models for most local inference servers.
• Robust Direct Parameter Passing: Unlike implementations that rely solely on environment variables, this PR passes the api_base directly to LiteLLM's completion calls, ensuring reliability across multi-provider sessions and preventing environment pollution.
• Centralized Configuration: All LLM setup logic is consolidated into a new setup_litellm helper in utils.py, making the codebase cleaner and easier to maintain.

Verification

• Tested with LM Studio and Ollama local endpoints.
• Verified environment variable isolation (no cross-contamination between different base_url settings).
• Confirmed recursive support for both PDF and Markdown processing flows.

[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.

We are going to close this as redundant with LiteLLM's existing functionality. PageIndex already routes LLM calls through LiteLLM, and LiteLLM already
supports provider selection via model strings like openai/..., anthropic/..., gemini/..., ollama/..., etc., as well as custom endpoints via its built-
in api_base parameter.

We do not want to duplicate LiteLLM provider/endpoint routing inside PageIndex with ad-hoc CLI flags, environment-variable mutation, global state, or model-
name rewriting. That would create inconsistent behavior across the CLI, SDK, PDF, and Markdown paths.

A future contribution would be more useful if it documents the recommended LiteLLM usage, or adds a minimal PageIndex config pass-through for api_base that
is scoped, tested, and free of global side effects.