ruby_llm 1.12.0 → 1.14.1

data/lib/ruby_llm/providers/bedrock.rb

@@ -18,12 +18,15 @@ module RubyLLM
         {}
       end
 
-      def complete(messages, tools:, temperature:, model:, params: {}, headers: {}, schema: nil, thinking: nil, &) # rubocop:disable Metrics/ParameterLists
+      # rubocop:disable Metrics/ParameterLists
+      def complete(messages, tools:, temperature:, model:, params: {}, headers: {}, schema: nil, thinking: nil,
+                   tool_prefs: nil, &)
         normalized_params = normalize_params(params, model:)
 
         super(
           messages,
           tools: tools,
+          tool_prefs: tool_prefs,
           temperature: temperature,
           model: model,
           params: normalized_params,
@@ -33,6 +36,7 @@ module RubyLLM
           &
         )
       end
+      # rubocop:enable Metrics/ParameterLists
 
       def parse_error(response)
         return if response.body.nil? || response.body.empty?
@@ -49,6 +53,10 @@ module RubyLLM
       end
 
       class << self
+        def configuration_options
+          %i[bedrock_api_key bedrock_secret_key bedrock_region bedrock_session_token]
+        end
+
         def configuration_requirements
           %i[bedrock_api_key bedrock_secret_key bedrock_region]
         end

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

@@ -3,127 +3,17 @@
 module RubyLLM
   module Providers
     class DeepSeek
-      # Determines capabilities and pricing for DeepSeek models
+      # Provider-level capability checks used outside the model registry.
       module Capabilities
         module_function
 
-        def context_window_for(model_id)
-          case model_id
-          when /deepseek-(?:chat|reasoner)/ then 64_000
-          else 32_768
-          end
+        def supports_tool_choice?(_model_id)
+          true
         end
 
-        def max_tokens_for(model_id)
-          case model_id
-          when /deepseek-(?:chat|reasoner)/ then 8_192
-          else 4_096
-          end
-        end
-
-        def input_price_for(model_id)
-          PRICES.dig(model_family(model_id), :input_miss) || default_input_price
-        end
-
-        def output_price_for(model_id)
-          PRICES.dig(model_family(model_id), :output) || default_output_price
-        end
-
-        def cache_hit_price_for(model_id)
-          PRICES.dig(model_family(model_id), :input_hit) || default_cache_hit_price
-        end
-
-        def supports_vision?(_model_id)
+        def supports_tool_parallel_control?(_model_id)
           false
         end
-
-        def supports_functions?(model_id)
-          model_id.match?(/deepseek-chat/)
-        end
-
-        def supports_json_mode?(_model_id)
-          false
-        end
-
-        def format_display_name(model_id)
-          case model_id
-          when 'deepseek-chat' then 'DeepSeek V3'
-          when 'deepseek-reasoner' then 'DeepSeek R1'
-          else
-            model_id.split('-')
-                    .map(&:capitalize)
-                    .join(' ')
-          end
-        end
-
-        def model_type(_model_id)
-          'chat'
-        end
-
-        def model_family(model_id)
-          case model_id
-          when /deepseek-reasoner/ then :reasoner
-          else :chat
-          end
-        end
-
-        PRICES = {
-          chat: {
-            input_hit: 0.07,
-            input_miss: 0.27,
-            output: 1.10
-          },
-          reasoner: {
-            input_hit: 0.14,
-            input_miss: 0.55,
-            output: 2.19
-          }
-        }.freeze
-
-        def default_input_price
-          0.27
-        end
-
-        def default_output_price
-          1.10
-        end
-
-        def default_cache_hit_price
-          0.07
-        end
-
-        def modalities_for(_model_id)
-          {
-            input: ['text'],
-            output: ['text']
-          }
-        end
-
-        def capabilities_for(model_id)
-          capabilities = ['streaming']
-
-          capabilities << 'function_calling' if model_id.match?(/deepseek-chat/)
-
-          capabilities
-        end
-
-        def pricing_for(model_id)
-          family = model_family(model_id)
-          prices = PRICES.fetch(family, { input_miss: default_input_price, output: default_output_price })
-
-          standard_pricing = {
-            input_per_million: prices[:input_miss],
-            output_per_million: prices[:output]
-          }
-
-          standard_pricing[:cached_input_per_million] = prices[:input_hit] if prices[:input_hit]
-
-          {
-            text_tokens: {
-              standard: standard_pricing
-            }
-          }
-        end
       end
     end
   end

data/lib/ruby_llm/providers/deepseek.rb

@@ -7,7 +7,7 @@ module RubyLLM
       include DeepSeek::Chat
 
       def api_base
-        'https://api.deepseek.com'
+        @config.deepseek_api_base || 'https://api.deepseek.com'
       end
 
       def headers
@@ -21,6 +21,10 @@ module RubyLLM
           DeepSeek::Capabilities
         end
 
+        def configuration_options
+          %i[deepseek_api_key deepseek_api_base]
+        end
+
         def configuration_requirements
           %i[deepseek_api_key]
         end

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)
@@ -52,17 +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_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
@@ -70,59 +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
@@ -139,142 +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

@@ -14,7 +14,10 @@ module RubyLLM
           "models/#{@model}:generateContent"
         end
 
-        def render_payload(messages, tools:, temperature:, model:, stream: false, schema: nil, thinking: nil) # rubocop:disable Metrics/ParameterLists,Lint/UnusedMethodArgument
+        # rubocop:disable Metrics/ParameterLists,Lint/UnusedMethodArgument
+        def render_payload(messages, tools:, temperature:, model:, stream: false, schema: nil,
+                           thinking: nil, tool_prefs: nil)
+          tool_prefs ||= {}
           @model = model.id
           payload = {
             contents: format_messages(messages),
@@ -26,9 +29,15 @@ module RubyLLM
           payload[:generationConfig].merge!(structured_output_config(schema, model)) if schema
           payload[:generationConfig][:thinkingConfig] = build_thinking_config(model, thinking) if thinking&.enabled?
 
-          payload[:tools] = format_tools(tools) if tools.any?
+          if tools.any?
+            payload[:tools] = format_tools(tools)
+            # Gemini doesn't support controlling parallel tool calls
+            payload[:toolConfig] = build_tool_config(tool_prefs[:choice]) unless tool_prefs[:choice].nil?
+          end
+
           payload
         end
+        # rubocop:enable Metrics/ParameterLists,Lint/UnusedMethodArgument
 
         def build_thinking_config(_model, thinking)
           config = { includeThoughts: true }
@@ -111,6 +120,7 @@ module RubyLLM
             tool_calls: tool_calls,
             input_tokens: data.dig('usageMetadata', 'promptTokenCount'),
             output_tokens: calculate_output_tokens(data),
+            cached_tokens: data.dig('usageMetadata', 'cachedContentTokenCount'),
             thinking_tokens: data.dig('usageMetadata', 'thoughtsTokenCount'),
             model_id: data['modelVersion'] || response.env.url.path.split('/')[3].split(':')[0],
             raw: response
@@ -120,6 +130,9 @@ module RubyLLM
         def convert_schema_to_gemini(schema)
           return nil unless schema
 
+          # Extract inner schema if wrapper format (e.g., from RubyLLM::Schema.to_json_schema)
+          schema = schema[:schema] || schema
+
           GeminiSchema.new(schema).to_h
         end
 
@@ -132,7 +145,10 @@ module RubyLLM
           parts = candidate.dig('content', 'parts')
           return '' unless parts&.any?
 
-          build_response_content(parts)
+          non_thought_parts = parts.reject { |part| part['thought'] }
+          return '' unless non_thought_parts.any?
+
+          build_response_content(non_thought_parts)
         end
 
         def extract_text_parts(parts)
@@ -176,7 +192,7 @@ module RubyLLM
         end
 
         def build_json_schema(schema)
-          normalized = RubyLLM::Utils.deep_dup(schema)
+          normalized = RubyLLM::Utils.deep_dup(schema[:schema])
           normalized.delete(:strict)
           normalized.delete('strict')
           RubyLLM::Utils.deep_stringify_keys(normalized)

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

@@ -10,7 +10,7 @@ module RubyLLM
         end
 
         def render_image_payload(prompt, model:, size:)
-          RubyLLM.logger.debug "Ignoring size #{size}. Gemini does not support image size customization."
+          RubyLLM.logger.debug { "Ignoring size #{size}. Gemini does not support image size customization." }
           @model = model
           {
             instances: [

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

@@ -22,6 +22,7 @@ module RubyLLM
             ),
             input_tokens: extract_input_tokens(data),
             output_tokens: extract_output_tokens(data),
+            cached_tokens: data.dig('usageMetadata', 'cachedContentTokenCount'),
             thinking_tokens: data.dig('usageMetadata', 'thoughtsTokenCount'),
             tool_calls: extract_tool_calls(data)
           )
@@ -83,7 +84,7 @@ module RubyLLM
           error_data = JSON.parse(data)
           [error_data['error']['code'], error_data['error']['message']]
         rescue JSON::ParserError => e
-          RubyLLM.logger.debug "Failed to parse streaming error: #{e.message}"
+          RubyLLM.logger.debug { "Failed to parse streaming error: #{e.message}" }
           [500, "Failed to parse error: #{data}"]
         end
       end

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

@@ -205,6 +205,25 @@ module RubyLLM
           else 'STRING'
           end
         end
+
+        def build_tool_config(tool_choice)
+          {
+            functionCallingConfig: {
+              mode: forced_tool_choice?(tool_choice) ? 'any' : tool_choice
+            }.tap do |config|
+              # Use allowedFunctionNames to simulate specific tool choice
+              config[:allowedFunctionNames] = [tool_choice] if specific_tool_choice?(tool_choice)
+            end
+          }
+        end
+
+        def forced_tool_choice?(tool_choice)
+          tool_choice == :required || specific_tool_choice?(tool_choice)
+        end
+
+        def specific_tool_choice?(tool_choice)
+          !%i[auto none required].include?(tool_choice)
+        end
       end
     end
   end

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/capabilities.rb

@@ -0,0 +1,20 @@
+# frozen_string_literal: true
+
+module RubyLLM
+  module Providers
+    class GPUStack
+      # Determines capabilities for GPUStack models
+      module Capabilities
+        module_function
+
+        def supports_tool_choice?(_model_id)
+          false
+        end
+
+        def supports_tool_parallel_control?(_model_id)
+          false
+        end
+      end
+    end
+  end
+end

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
@@ -28,6 +32,10 @@ module RubyLLM
         def configuration_requirements
           %i[gpustack_api_base]
         end
+
+        def capabilities
+          GPUStack::Capabilities
+        end
       end
     end
   end