Allow local models to fetch model list

This prevents assume local model warning from happening while keeping local model registry.

Both GPUStack and Ollama can now fetch model lists, so in theory running RubyLLM.models.refresh! is all that's needed to verify's model availability.

Author: khasinski

lib/ruby_llm/models.rb

@@ -37,7 +37,7 @@ def refresh!
       end
 
       def fetch_from_providers
-        configured = Provider.configured_providers(RubyLLM.config).filter(&:remote?)
+        configured = Provider.configured_providers(RubyLLM.config)
 
         RubyLLM.logger.info "Fetching models from providers: #{configured.map(&:slug).join(', ')}"
 
@@ -47,23 +47,24 @@ def fetch_from_providers
       end
 
       def resolve(model_id, provider: nil, assume_exists: false) # rubocop:disable Metrics/PerceivedComplexity
-        assume_exists = true if provider && Provider.providers[provider.to_sym].local?
-
-        if assume_exists
+        if assume_exists || local_provider?(provider)
+          model = find_local_model(model_id, provider) if local_provider?(provider)
           raise ArgumentError, 'Provider must be specified if assume_exists is true' unless provider
 
           provider = Provider.providers[provider.to_sym] || raise(Error, "Unknown provider: #{provider.to_sym}")
-          model = Model::Info.new(
-            id: model_id,
-            name: model_id.gsub('-', ' ').capitalize,
-            provider: provider.slug,
-            capabilities: %w[function_calling streaming],
-            modalities: { input: %w[text image], output: %w[text] },
-            metadata: { warning: 'Assuming model exists, capabilities may not be accurate' }
-          )
-          if RubyLLM.config.log_assume_model_exists
-            RubyLLM.logger.warn "Assuming model '#{model_id}' exists for provider '#{provider}'. " \
-                                'Capabilities may not be accurately reflected.'
+          unless model
+            model = Model::Info.new(
+              id: model_id,
+              name: model_id.gsub('-', ' ').capitalize,
+              provider: provider.slug,
+              capabilities: %w[function_calling streaming],
+              modalities: { input: %w[text image], output: %w[text] },
+              metadata: { warning: 'Assuming model exists, capabilities may not be accurate' }
+            )
+            if RubyLLM.config.log_assume_model_exists
+              RubyLLM.logger.warn "Assuming model '#{model_id}' exists for provider '#{provider}'. " \
+                                  'Capabilities may not be accurately reflected.'
+            end
           end
         else
           model = Models.find model_id, provider
@@ -223,5 +224,15 @@ def find_without_provider(model_id)
         all.find { |m| m.id == Aliases.resolve(model_id) } ||
         raise(ModelNotFoundError, "Unknown model: #{model_id}")
     end
+
+    def self.local_provider?(provider)
+      provider && Provider.providers[provider.to_sym]&.local?
+    end
+
+    def self.find_local_model(model_id, provider)
+      Models.find(model_id, provider)
+    rescue ModelNotFoundError
+      nil
+    end
   end
 end

lib/ruby_llm/providers/ollama.rb

@@ -7,6 +7,7 @@ module Ollama
       extend OpenAI
       extend Ollama::Chat
       extend Ollama::Media
+      extend Ollama::Models
 
       module_function
 

lib/ruby_llm/providers/ollama/models.rb

@@ -0,0 +1,35 @@
+# frozen_string_literal: true
+
+module RubyLLM
+  module Providers
+    module Ollama
+      # Models methods of the Ollama API integration
+      module Models
+        module_function
+
+        def models_url
+          'models'
+        end
+
+        def parse_list_models_response(response, slug, _capabilities)
+          Array(response.body['data']).map do |model_data|
+            Model::Info.new(
+              id: model_data['id'],
+              name: model_data['id'],
+              provider: slug,
+              family: 'ollama',
+              created_at: model_data['created'] ? Time.at(model_data['created']) : nil,
+              modalities: {
+                input: %w[text image], # Ollama models don't advertise input modalities, so we assume text and image
+                output: ['text'] # Ollama models don't expose output modalities, so we assume text
+              },
+              capabilities: %w[streaming function_calling structured_output],
+              pricing: {}, # Ollama does not provide pricing details
+              metadata: {}
+            )
+          end
+        end
+      end
+    end
+  end
+end