From bdb6ba77536b6242b8728e044415f4ca43921223 Mon Sep 17 00:00:00 2001
From: Justyna Wojtczak <justine84@gmail.com>
Date: Wed, 18 Mar 2026 00:21:17 +0100
Subject: [PATCH] fix: use Ollama OpenAI-compatible endpoint (/v1) by default
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The Ollama provider inherits from OpenAI but used the native API
base (e.g. http://localhost:11434) which doesn't serve endpoints
like /chat/completions or /models. This caused:

- Model listing to fail with 404 (tried /models instead of /api/tags)
- with_schema to silently not work (models not in registry, no
  structured_output capability detected)

Fix: automatically append /v1 to ollama_api_base so the provider
uses Ollama's OpenAI-compatible endpoint. Idempotent — does not
double-append if /v1 is already present. This gives:

- Model listing works (/v1/models)
- Chat works (/v1/chat/completions)
- with_schema works (response_format passed correctly)
- All OpenAI provider logic inherited without changes

Tested with Ollama 0.17.5, gemma:latest (9B), llama3.2:3b (3B).

Note: Ollama's native /api/chat endpoint with `format` param
provides better nested JSON Schema enforcement than the OpenAI-
compatible endpoint. A future enhancement could use the native
endpoint for structured output.
---
 lib/ruby_llm/providers/ollama.rb       |  8 ++-
 spec/ruby_llm/providers/ollama_spec.rb | 79 +++++++++++++++++++++++---
 2 files changed, 77 insertions(+), 10 deletions(-)

diff --git a/lib/ruby_llm/providers/ollama.rb b/lib/ruby_llm/providers/ollama.rb
index 3c0d676a2..45c6e2fd7 100644
--- a/lib/ruby_llm/providers/ollama.rb
+++ b/lib/ruby_llm/providers/ollama.rb
@@ -8,8 +8,14 @@ class Ollama < OpenAI
       include Ollama::Media
       include Ollama::Models
 
+      # Ollama exposes two API surfaces:
+      #   - Native API at /api/* (different request/response format)
+      #   - OpenAI-compatible API at /v1/* (same format as OpenAI)
+      # Since this provider inherits from OpenAI, we use the /v1 endpoint
+      # so all OpenAI logic (chat, models, schemas) works without changes.
       def api_base
-        @config.ollama_api_base
+        base = @config.ollama_api_base.to_s.chomp('/')
+        base.end_with?('/v1') ? base : "#{base}/v1"
       end
 
       def headers
diff --git a/spec/ruby_llm/providers/ollama_spec.rb b/spec/ruby_llm/providers/ollama_spec.rb
index 9910e08c1..eae77e17a 100644
--- a/spec/ruby_llm/providers/ollama_spec.rb
+++ b/spec/ruby_llm/providers/ollama_spec.rb
@@ -3,21 +3,82 @@
 require 'spec_helper'
 
 RSpec.describe RubyLLM::Providers::Ollama do
-  include_context 'with configured RubyLLM'
+  subject(:provider) { described_class.new(config) }
+
+  let(:config) do
+    instance_double(
+      RubyLLM::Configuration,
+      request_timeout: 300,
+      max_retries: 3,
+      retry_interval: 0.1,
+      retry_interval_randomness: 0.5,
+      retry_backoff_factor: 2,
+      http_proxy: nil,
+      ollama_api_base: ollama_api_base,
+      ollama_api_key: ollama_api_key
+    )
+  end
+
+  let(:ollama_api_base) { 'http://localhost:11434' }
+  let(:ollama_api_key) { nil }
 
   describe '#headers' do
-    it 'returns empty headers when no API key is configured' do
-      RubyLLM.configure { |config| config.ollama_api_key = nil }
-      provider = described_class.new(RubyLLM.config)
+    context 'when no API key is configured' do
+      let(:ollama_api_key) { nil }
+
+      it 'returns empty headers' do
+        expect(provider.headers).to eq({})
+      end
+    end
+
+    context 'when API key is configured' do
+      let(:ollama_api_key) { 'test-ollama-key' }
+
+      it 'returns Authorization header' do
+        expect(provider.headers).to eq({ 'Authorization' => 'Bearer test-ollama-key' })
+      end
+    end
+  end
+
+  describe '#api_base' do
+    context 'when base URL does not include /v1' do
+      let(:ollama_api_base) { 'http://localhost:11434' }
+
+      it 'appends /v1 for OpenAI-compatible endpoint' do
+        expect(provider.api_base).to eq('http://localhost:11434/v1')
+      end
+    end
+
+    context 'when base URL already includes /v1' do
+      let(:ollama_api_base) { 'http://localhost:11434/v1' }
+
+      it 'does not double-append /v1' do
+        expect(provider.api_base).to eq('http://localhost:11434/v1')
+      end
+    end
+
+    context 'when base URL has a trailing slash' do
+      let(:ollama_api_base) { 'http://localhost:11434/' }
+
+      it 'strips trailing slash and appends /v1' do
+        expect(provider.api_base).to eq('http://localhost:11434/v1')
+      end
+    end
+
+    context 'when base URL has /v1/ with trailing slash' do
+      let(:ollama_api_base) { 'http://localhost:11434/v1/' }
 
-      expect(provider.headers).to eq({})
+      it 'normalizes to /v1 without trailing slash' do
+        expect(provider.api_base).to eq('http://localhost:11434/v1')
+      end
     end
 
-    it 'returns Authorization header when API key is configured' do
-      RubyLLM.configure { |config| config.ollama_api_key = 'test-ollama-key' }
-      provider = described_class.new(RubyLLM.config)
+    context 'when using a custom host and port' do
+      let(:ollama_api_base) { 'https://my-ollama.com:8080' }
 
-      expect(provider.headers).to eq({ 'Authorization' => 'Bearer test-ollama-key' })
+      it 'appends /v1 to the custom base' do
+        expect(provider.api_base).to eq('https://my-ollama.com:8080/v1')
+      end
     end
   end
 end