ruby_llm 1.13.2 → 1.16.0

data/lib/ruby_llm/providers/gemini/capabilities.rb

@@ -3,13 +3,35 @@
 module RubyLLM
   module Providers
     class Gemini
-      # Determines capabilities and pricing for Google Gemini models
+      # Provider-level capability checks and narrow registry fallbacks.
       module Capabilities
         module_function
 
+        PRICES = {
+          flash_2: { input: 0.10, output: 0.40 }, # rubocop:disable Naming/VariableNumber
+          flash_lite_2: { input: 0.075, output: 0.30 }, # rubocop:disable Naming/VariableNumber
+          flash: { input: 0.075, output: 0.30 },
+          flash_8b: { input: 0.0375, output: 0.15 },
+          pro: { input: 1.25, output: 5.0 },
+          pro_2_5: { input: 0.12, output: 0.50 }, # rubocop:disable Naming/VariableNumber
+          gemini_embedding: { input: 0.002, output: 0.004 },
+          embedding: { input: 0.00, output: 0.00 },
+          imagen: { price: 0.03 },
+          aqa: { input: 0.00, output: 0.00 }
+        }.freeze
+
+        def supports_tool_choice?(_model_id)
+          true
+        end
+
+        def supports_tool_parallel_control?(_model_id)
+          false
+        end
+
         def context_window_for(model_id)
           case model_id
-          when /gemini-2\.5-pro-exp-03-25/, /gemini-2\.0-flash/, /gemini-2\.0-flash-lite/, /gemini-1\.5-flash/, /gemini-1\.5-flash-8b/ # rubocop:disable Layout/LineLength
+          when /gemini-2\.5-pro-exp-03-25/, /gemini-2\.0-flash/, /gemini-2\.0-flash-lite/, /gemini-1\.5-flash/,
+               /gemini-1\.5-flash-8b/
             1_048_576
           when /gemini-1\.5-pro/ then 2_097_152
           when /gemini-embedding-exp/ then 8_192
@@ -23,7 +45,8 @@ module RubyLLM
         def max_tokens_for(model_id)
           case model_id
           when /gemini-2\.5-pro-exp-03-25/ then 64_000
-          when /gemini-2\.0-flash/, /gemini-2\.0-flash-lite/, /gemini-1\.5-flash/, /gemini-1\.5-flash-8b/, /gemini-1\.5-pro/ # rubocop:disable Layout/LineLength
+          when /gemini-2\.0-flash/, /gemini-2\.0-flash-lite/, /gemini-1\.5-flash/, /gemini-1\.5-flash-8b/,
+               /gemini-1\.5-pro/
             8_192
           when /gemini-embedding-exp/ then nil
           when /text-embedding-004/, /embedding-001/ then 768
@@ -32,18 +55,24 @@ module RubyLLM
           end
         end
 
-        def input_price_for(model_id)
-          base_price = PRICES.dig(pricing_family(model_id), :input) || default_input_price
-          return base_price unless long_context_model?(model_id)
-
-          context_window_for(model_id) > 128_000 ? base_price * 2 : base_price
+        def critical_capabilities_for(model_id)
+          capabilities = []
+          capabilities << 'function_calling' if supports_functions?(model_id)
+          capabilities << 'structured_output' if supports_structured_output?(model_id)
+          capabilities << 'vision' if supports_vision?(model_id)
+          capabilities
         end
 
-        def output_price_for(model_id)
-          base_price = PRICES.dig(pricing_family(model_id), :output) || default_output_price
-          return base_price unless long_context_model?(model_id)
-
-          context_window_for(model_id) > 128_000 ? base_price * 2 : base_price
+        def pricing_for(model_id)
+          prices = PRICES.fetch(pricing_family(model_id), { input: 0.075, output: 0.30 })
+          {
+            text_tokens: {
+              standard: {
+                input_per_million: prices[:input] || prices[:price] || 0.075,
+                output_per_million: prices[:output] || prices[:price] || 0.30
+              }
+            }
+          }
         end
 
         def supports_vision?(model_id)
@@ -52,25 +81,13 @@ module RubyLLM
           model_id.match?(/gemini|flash|pro|imagen/)
         end
 
-        def supports_video?(model_id)
-          model_id.match?(/gemini/)
-        end
-
         def supports_functions?(model_id)
           return false if model_id.match?(/text-embedding|embedding-001|aqa|flash-lite|imagen|gemini-2\.0-flash-lite/)
 
           model_id.match?(/gemini|pro|flash/)
         end
 
-        def supports_tool_choice?(_model_id)
-          true
-        end
-
-        def supports_tool_parallel_control?(_model_id)
-          false
-        end
-
-        def supports_json_mode?(model_id)
+        def supports_structured_output?(model_id)
           if model_id.match?(/text-embedding|embedding-001|aqa|imagen|gemini-2\.0-flash-lite|gemini-2\.5-pro-exp-03-25/)
             return false
           end
@@ -78,59 +95,6 @@ module RubyLLM
           model_id.match?(/gemini|pro|flash/)
         end
 
-        def format_display_name(model_id)
-          model_id
-            .delete_prefix('models/')
-            .split('-')
-            .map(&:capitalize)
-            .join(' ')
-            .gsub(/(\d+\.\d+)/, ' \1')
-            .gsub(/\s+/, ' ')
-            .gsub('Aqa', 'AQA')
-            .strip
-        end
-
-        def supports_caching?(model_id)
-          if model_id.match?(/flash-lite|gemini-2\.5-pro-exp-03-25|aqa|imagen|text-embedding|embedding-001/)
-            return false
-          end
-
-          model_id.match?(/gemini|pro|flash/)
-        end
-
-        def supports_tuning?(model_id)
-          model_id.match?(/gemini-1\.5-flash|gemini-1\.5-flash-8b/)
-        end
-
-        def supports_audio?(model_id)
-          model_id.match?(/gemini|pro|flash/)
-        end
-
-        def model_type(model_id)
-          case model_id
-          when /text-embedding|embedding|gemini-embedding/ then 'embedding'
-          when /imagen/ then 'image'
-          else 'chat'
-          end
-        end
-
-        def model_family(model_id)
-          case model_id
-          when /gemini-2\.5-pro-exp-03-25/ then 'gemini25_pro_exp'
-          when /gemini-2\.0-flash-lite/ then 'gemini20_flash_lite'
-          when /gemini-2\.0-flash/ then 'gemini20_flash'
-          when /gemini-1\.5-flash-8b/ then 'gemini15_flash_8b'
-          when /gemini-1\.5-flash/ then 'gemini15_flash'
-          when /gemini-1\.5-pro/ then 'gemini15_pro'
-          when /gemini-embedding-exp/ then 'gemini_embedding_exp'
-          when /text-embedding-004/ then 'embedding4'
-          when /embedding-001/ then 'embedding1'
-          when /aqa/ then 'aqa'
-          when /imagen-3/ then 'imagen3'
-          else 'other'
-          end
-        end
-
         def pricing_family(model_id)
           case model_id
           when /gemini-2\.5-pro-exp-03-25/ then :pro_2_5 # rubocop:disable Naming/VariableNumber
@@ -147,142 +111,8 @@ module RubyLLM
           end
         end
 
-        def long_context_model?(model_id)
-          model_id.match?(/gemini-1\.5-(?:pro|flash)|gemini-1\.5-flash-8b/)
-        end
-
-        def context_length(model_id)
-          context_window_for(model_id)
-        end
-
-        PRICES = {
-          flash_2: { # rubocop:disable Naming/VariableNumber
-            input: 0.10,
-            output: 0.40,
-            audio_input: 0.70,
-            cache: 0.025,
-            cache_storage: 1.00,
-            grounding_search: 35.00
-          },
-          flash_lite_2: { # rubocop:disable Naming/VariableNumber
-            input: 0.075,
-            output: 0.30
-          },
-          flash: {
-            input: 0.075,
-            output: 0.30,
-            cache: 0.01875,
-            cache_storage: 1.00,
-            grounding_search: 35.00
-          },
-          flash_8b: {
-            input: 0.0375,
-            output: 0.15,
-            cache: 0.01,
-            cache_storage: 0.25,
-            grounding_search: 35.00
-          },
-          pro: {
-            input: 1.25,
-            output: 5.0,
-            cache: 0.3125,
-            cache_storage: 4.50,
-            grounding_search: 35.00
-          },
-          pro_2_5: { # rubocop:disable Naming/VariableNumber
-            input: 0.12,
-            output: 0.50
-          },
-          gemini_embedding: {
-            input: 0.002,
-            output: 0.004
-          },
-          embedding: {
-            input: 0.00,
-            output: 0.00
-          },
-          imagen: {
-            price: 0.03
-          },
-          aqa: {
-            input: 0.00,
-            output: 0.00
-          }
-        }.freeze
-
-        def default_input_price
-          0.075
-        end
-
-        def default_output_price
-          0.30
-        end
-
-        def modalities_for(model_id)
-          modalities = {
-            input: ['text'],
-            output: ['text']
-          }
-
-          if supports_vision?(model_id)
-            modalities[:input] << 'image'
-            modalities[:input] << 'pdf'
-          end
-
-          modalities[:input] << 'video' if supports_video?(model_id)
-          modalities[:input] << 'audio' if model_id.match?(/audio/)
-          modalities[:output] << 'embeddings' if model_id.match?(/embedding|gemini-embedding/)
-          modalities[:output] = ['image'] if model_id.match?(/imagen/)
-
-          modalities
-        end
-
-        def capabilities_for(model_id)
-          capabilities = ['streaming']
-
-          capabilities << 'function_calling' if supports_functions?(model_id)
-          capabilities << 'structured_output' if supports_json_mode?(model_id)
-          capabilities << 'batch' if model_id.match?(/embedding|flash/)
-          capabilities << 'caching' if supports_caching?(model_id)
-          capabilities << 'fine_tuning' if supports_tuning?(model_id)
-          capabilities
-        end
-
-        def pricing_for(model_id)
-          family = pricing_family(model_id)
-          prices = PRICES.fetch(family, { input: default_input_price, output: default_output_price })
-
-          standard_pricing = {
-            input_per_million: prices[:input],
-            output_per_million: prices[:output]
-          }
-
-          standard_pricing[:cached_input_per_million] = prices[:input_hit] if prices[:input_hit]
-
-          batch_pricing = {
-            input_per_million: (standard_pricing[:input_per_million] || 0) * 0.5,
-            output_per_million: (standard_pricing[:output_per_million] || 0) * 0.5
-          }
-
-          if standard_pricing[:cached_input_per_million]
-            batch_pricing[:cached_input_per_million] = standard_pricing[:cached_input_per_million] * 0.5
-          end
-
-          pricing = {
-            text_tokens: {
-              standard: standard_pricing,
-              batch: batch_pricing
-            }
-          }
-
-          if model_id.match?(/embedding|gemini-embedding/)
-            pricing[:embeddings] = {
-              standard: { input_per_million: prices[:price] || 0.002 }
-            }
-          end
-
-          pricing
-        end
+        module_function :context_window_for, :max_tokens_for, :critical_capabilities_for, :pricing_for,
+                        :supports_vision?, :supports_functions?, :supports_structured_output?, :pricing_family
       end
     end
   end

data/lib/ruby_llm/providers/gemini/chat.rb

@@ -72,8 +72,7 @@ module RubyLLM
         def format_role(role)
           case role
           when :assistant then 'model'
-          when :system then 'user'
-          when :tool then 'function'
+          when :system, :tool then 'user'
           else role.to_s
           end
         end
@@ -118,7 +117,7 @@ module RubyLLM
               signature: extract_thought_signature(parts)
             ),
             tool_calls: tool_calls,
-            input_tokens: data.dig('usageMetadata', 'promptTokenCount'),
+            input_tokens: input_tokens(data),
             output_tokens: calculate_output_tokens(data),
             cached_tokens: data.dig('usageMetadata', 'cachedContentTokenCount'),
             thinking_tokens: data.dig('usageMetadata', 'thoughtsTokenCount'),
@@ -127,6 +126,13 @@ module RubyLLM
           )
         end
 
+        def input_tokens(data)
+          prompt_tokens = data.dig('usageMetadata', 'promptTokenCount')
+          return unless prompt_tokens
+
+          [prompt_tokens.to_i - data.dig('usageMetadata', 'cachedContentTokenCount').to_i, 0].max
+        end
+
         def convert_schema_to_gemini(schema)
           return nil unless schema
 
@@ -307,7 +313,7 @@ module RubyLLM
           end
 
           def build_tool_response(parts)
-            { role: 'function', parts: parts }
+            { role: 'user', parts: parts }
           end
 
           def remember_tool_calls

data/lib/ruby_llm/providers/gemini/images.rb

@@ -5,11 +5,11 @@ module RubyLLM
     class Gemini
       # Image generation methods for the Gemini API implementation
       module Images
-        def images_url
+        def images_url(with: nil, mask: nil) # rubocop:disable Lint/UnusedMethodArgument
           "models/#{@model}:predict"
         end
 
-        def render_image_payload(prompt, model:, size:)
+        def render_image_payload(prompt, model:, size:, with: nil, mask: nil, params: {}) # rubocop:disable Lint/UnusedMethodArgument,Metrics/ParameterLists
           RubyLLM.logger.debug { "Ignoring size #{size}. Gemini does not support image size customization." }
           @model = model
           {

data/lib/ruby_llm/providers/gemini/media.rb

@@ -1,5 +1,8 @@
 # frozen_string_literal: true
 
+require 'base64'
+require 'stringio'
+
 module RubyLLM
   module Providers
     class Gemini # rubocop:disable Style/Documentation
@@ -16,19 +19,23 @@ module RubyLLM
           parts << format_text(content.text) if content.text
 
           content.attachments.each do |attachment|
-            case attachment.type
-            when :text
-              parts << format_text_file(attachment)
-            when :unknown
-              raise UnsupportedAttachmentError, attachment.mime_type
-            else
-              parts << format_attachment(attachment)
-            end
+            parts << format_content_attachment(attachment)
           end
 
           parts
         end
 
+        def format_content_attachment(attachment)
+          case attachment.type
+          when :text
+            format_text_file(attachment)
+          when :document, :unknown
+            raise UnsupportedAttachmentError, attachment.mime_type
+          else
+            format_attachment(attachment)
+          end
+        end
+
         def format_attachment(attachment)
           {
             inline_data: {
@@ -71,7 +78,7 @@ module RubyLLM
         text = nil if text.empty?
         return text if attachments.empty?
 
-        Content.new(text:, attachments:)
+        Content.new(text, attachments)
       end
 
       def build_inline_attachment(inline_data, index)

data/lib/ruby_llm/providers/gemini/models.rb

@@ -17,14 +17,12 @@ module RubyLLM
 
             Model::Info.new(
               id: model_id,
-              name: model_data['displayName'],
+              name: model_data['displayName'] || model_id,
               provider: slug,
-              family: capabilities.model_family(model_id),
               created_at: nil,
               context_window: model_data['inputTokenLimit'] || capabilities.context_window_for(model_id),
               max_output_tokens: model_data['outputTokenLimit'] || capabilities.max_tokens_for(model_id),
-              modalities: capabilities.modalities_for(model_id),
-              capabilities: capabilities.capabilities_for(model_id),
+              capabilities: capabilities.critical_capabilities_for(model_id),
               pricing: capabilities.pricing_for(model_id),
               metadata: {
                 version: model_data['version'],

data/lib/ruby_llm/providers/gemini/streaming.rb

@@ -1,5 +1,7 @@
 # frozen_string_literal: true
 
+require 'json'
+
 module RubyLLM
   module Providers
     class Gemini
@@ -70,7 +72,10 @@ module RubyLLM
         end
 
         def extract_input_tokens(data)
-          data.dig('usageMetadata', 'promptTokenCount')
+          prompt_tokens = data.dig('usageMetadata', 'promptTokenCount')
+          return unless prompt_tokens
+
+          [prompt_tokens.to_i - data.dig('usageMetadata', 'cachedContentTokenCount').to_i, 0].max
         end
 
         def extract_output_tokens(data)

data/lib/ruby_llm/providers/gemini/tools.rb

@@ -1,5 +1,7 @@
 # frozen_string_literal: true
 
+require 'securerandom'
+
 module RubyLLM
   module Providers
     class Gemini
@@ -46,13 +48,15 @@ module RubyLLM
 
         def format_tool_result(msg, function_name = nil)
           function_name ||= msg.tool_call_id
+          content = msg.content
+          content = '(no output)' if content.nil? || (content.respond_to?(:empty?) && content.empty?)
 
           [{
             functionResponse: {
               name: function_name,
               response: {
                 name: function_name,
-                content: Media.format_content(msg.content)
+                content: Media.format_content(content)
               }
             }
           }]

data/lib/ruby_llm/providers/gemini.rb

@@ -28,6 +28,10 @@ module RubyLLM
           Gemini::Capabilities
         end
 
+        def configuration_options
+          %i[gemini_api_key gemini_api_base]
+        end
+
         def configuration_requirements
           %i[gemini_api_key]
         end

data/lib/ruby_llm/providers/gpustack/chat.rb

@@ -11,13 +11,20 @@ module RubyLLM
           messages.map do |msg|
             {
               role: format_role(msg.role),
-              content: GPUStack::Media.format_content(msg.content),
+              content: format_message_content(msg),
               tool_calls: format_tool_calls(msg.tool_calls),
               tool_call_id: msg.tool_call_id
             }.compact.merge(OpenAI::Chat.format_thinking(msg))
           end
         end
 
+        def format_message_content(msg)
+          content = GPUStack::Media.format_content(msg.content)
+          return '' if content.nil? && OpenAI::Chat.thinking_only_assistant_message?(msg)
+
+          content
+        end
+
         def format_role(role)
           role.to_s
         end

data/lib/ruby_llm/providers/gpustack/models.rb

@@ -1,5 +1,7 @@
 # frozen_string_literal: true
 
+require 'time'
+
 module RubyLLM
   module Providers
     class GPUStack

data/lib/ruby_llm/providers/gpustack.rb

@@ -21,6 +21,10 @@ module RubyLLM
       end
 
       class << self
+        def configuration_options
+          %i[gpustack_api_base gpustack_api_key]
+        end
+
         def local?
           true
         end

data/lib/ruby_llm/providers/mistral/capabilities.rb

@@ -31,6 +31,11 @@ module RubyLLM
           !model_id.match?(/embed|moderation|ocr|voxtral|transcriptions/) && supports_tools?(model_id)
         end
 
+        def supports_reasoning?(model_id)
+          model_id.match?(/magistral/) ||
+            model_id.match?(/\Amistral-(?:small-latest|medium-(?:3(?:[.-]5)?|latest))\z/)
+        end
+
         def format_display_name(model_id)
           case model_id
           when /mistral-large/ then 'Mistral Large'
@@ -101,7 +106,7 @@ module RubyLLM
             capabilities << 'structured_output' if supports_json_mode?(model_id)
             capabilities << 'vision' if supports_vision?(model_id)
 
-            capabilities << 'reasoning' if model_id.match?(/magistral/)
+            capabilities << 'reasoning' if supports_reasoning?(model_id)
             capabilities << 'batch' unless model_id.match?(/voxtral|ocr|embed|moderation/)
             capabilities << 'fine_tuning' if model_id.match?(/mistral-(small|medium|large)|devstral/)
             capabilities << 'distillation' if model_id.match?(/ministral/)
@@ -118,7 +123,7 @@ module RubyLLM
           }
         end
 
-        def release_date_for(model_id)
+        def release_date_for(model_id) # rubocop:disable Metrics/CyclomaticComplexity
           case model_id
           when 'open-mistral-7b', 'mistral-tiny' then '2023-09-27'
           when 'mistral-medium-2312', 'mistral-small-2312', 'mistral-small',

data/lib/ruby_llm/providers/mistral/chat.rb

@@ -27,14 +27,32 @@ module RubyLLM
                            schema: nil, thinking: nil, tool_prefs: nil)
           payload = super
           payload.delete(:stream_options)
-          payload.delete(:reasoning_effort)
-          warn_on_unsupported_thinking(model, thinking)
+          configure_thinking_payload(payload, model, thinking)
+          normalize_required_tool_choice(payload)
           payload
         end
         # rubocop:enable Metrics/ParameterLists
 
+        def build_tool_choice(tool_choice)
+          return 'any' if tool_choice == :required
+
+          OpenAI::Tools.build_tool_choice(tool_choice)
+        end
+
+        def normalize_required_tool_choice(payload)
+          return unless payload[:tool_choice] == 'any' && Array(payload[:tools]).one?
+
+          function_name = payload.dig(:tools, 0, :function, :name)
+          return unless function_name
+
+          payload[:tool_choice] = {
+            type: 'function',
+            function: { name: function_name }
+          }
+        end
+
         def format_content_with_thinking(msg)
-          formatted_content = OpenAI::Media.format_content(msg.content)
+          formatted_content = Mistral::Media.format_content(msg.content)
           return formatted_content unless msg.role == :assistant && msg.thinking
 
           content_blocks = build_thinking_blocks(msg.thinking)
@@ -45,14 +63,47 @@ module RubyLLM
 
         def warn_on_unsupported_thinking(model, thinking)
           return unless thinking&.enabled?
-          return if model.id.to_s.include?('magistral')
+          return if native_reasoning_model?(model.id) || adjustable_reasoning_model?(model.id)
 
           RubyLLM.logger.warn(
-            'Mistral thinking is only supported on Magistral models. ' \
+            'Mistral thinking is only supported on Magistral and adjustable-reasoning models. ' \
             "Ignoring thinking settings for #{model.id}."
           )
         end
 
+        def configure_thinking_payload(payload, model, thinking)
+          return unless thinking&.enabled?
+
+          if native_reasoning_model?(model.id)
+            configure_native_reasoning_payload(payload, thinking)
+          elsif adjustable_reasoning_model?(model.id)
+            payload[:reasoning_effort] = reasoning_effort_for(thinking)
+          else
+            payload.delete(:reasoning_effort)
+            warn_on_unsupported_thinking(model, thinking)
+          end
+        end
+
+        def configure_native_reasoning_payload(payload, thinking)
+          payload.delete(:reasoning_effort)
+          payload[:prompt_mode] = thinking.effort == 'none' ? nil : 'reasoning'
+        end
+
+        def reasoning_effort_for(thinking)
+          effort = thinking.respond_to?(:effort) ? thinking.effort : nil
+          return effort if %w[high none].include?(effort)
+
+          'high'
+        end
+
+        def native_reasoning_model?(model_id)
+          model_id.to_s.include?('magistral')
+        end
+
+        def adjustable_reasoning_model?(model_id)
+          model_id.to_s.match?(/\Amistral-(?:small-latest|medium-(?:3(?:[.-]5)?|latest))\z/)
+        end
+
         def build_thinking_blocks(thinking)
           return [] unless thinking