fix: use provider-specific default base_url in BaseLLMReranker (Ollama gets localhost:11434, not OpenAI endpoint)

[amethystani]

Summary

Fixes #11952 — BaseLLMReranker uses DEFAULT_LLM_ENDPOINT (https://api.openai.com/v1) as the fallback base_url for all OpenAI-compatible providers, including Ollama. This means any user who configures an Ollama reranker without explicitly setting base_url gets their requests silently routed to the OpenAI endpoint, which fails with 404.

Root Cause

In BaseLLMReranker._init_client() the branch that handles openai and ollama providers always fell back to the same default:

# Before — same OpenAI URL used as fallback for both providers
base_url = self.base_url or DEFAULT_LLM_ENDPOINT   # https://api.openai.com/v1

Fix

Use a provider-specific default so Ollama falls back to http://localhost:11434/v1 (the standard local Ollama endpoint), while the openai provider retains its existing DEFAULT_LLM_ENDPOINT default:

if self.provider == 'ollama':
    default_base_url = 'http://localhost:11434/v1'
else:
    default_base_url = DEFAULT_LLM_ENDPOINT
base_url = self.base_url or default_base_url

An explicit base_url is still respected in all cases.

Tests Added

tests/unit/interfaces/knowledge_base/test_reranker_base_url.py — 5 parametrized cases covering:

• Ollama with no base_url → gets Ollama default ✓
• Ollama with explicit base_url → respected ✓
• OpenAI with no base_url → gets DEFAULT_LLM_ENDPOINT ✓
• OpenAI with explicit base_url → respected ✓
• Regression: Ollama default ≠ OpenAI endpoint ✓

cc @MinuraPunchihewa @paxmaxei @kumpelblase2 — this is the remaining bug after the settings.py default was corrected. Happy to address any feedback.

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[amethystani]

I have read the CLA Document and I hereby sign the CLA

[entelligence-ai-pr-reviews]

EntelligenceAI PR Summary

Fixes gh-11952 by ensuring BaseLLMReranker uses the correct default base URL based on the configured provider.

• Modified _init_client in base_reranker.py to check if provider is 'ollama' and apply http://localhost:11434/v1 as the fallback base URL
• All other providers continue to use DEFAULT_LLM_ENDPOINT when no base_url is specified
• Added test_reranker_base_url.py with unit tests using unittest.mock to bypass _init_client during construction and inspect AsyncOpenAI call kwargs for URL correctness

---

Confidence Score: 4/5 - Mostly Safe

Safe to merge — this PR correctly addresses the provider-specific default base URL bug in BaseLLMReranker._init_client, ensuring Ollama uses http://localhost:11434/v1 instead of the OpenAI endpoint. The fix is surgical and well-scoped, touching only the relevant conditional logic in base_reranker.py and backing the change with unit tests in test_reranker_base_url.py. No review comments were generated and no critical or significant issues were identified by static analysis.

Key Findings:

• The conditional check in _init_client correctly isolates Ollama's default base URL (http://localhost:11434/v1) without affecting any other provider's fallback to DEFAULT_LLM_ENDPOINT, making the logic change minimal and low-risk.
• Unit tests in test_reranker_base_url.py use unittest.mock to bypass _init_client during construction and inspect AsyncOpenAI call arguments, which is an appropriate testing pattern for this kind of initialization-time behavior.
• Only 1 of 2 changed files was reviewed by the automated analysis (coverage gap on either the test file or the implementation), but given the simplicity of the change and zero flagged issues, this is a minor concern rather than a blocking one.

Files requiring special attention

• base_reranker.py
• test_reranker_base_url.py

[torrmal]

thank you!! can you please create a pull request in github.com/mindsdb/engine

[amethystani]

Sounds good! I'll port this over to the engine repo right now and link the new PR here once it's up.

[amethystani]

The PR is up on the engine repo here: mindsdb/engine#2

Closing this one out. Thanks again!