ruby_llm 1.14.1 → 1.16.0

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

@@ -1,5 +1,9 @@
 # frozen_string_literal: true
 
+require 'base64'
+require 'faraday'
+require 'json'
+
 module RubyLLM
   module Providers
     class Bedrock
@@ -158,7 +162,10 @@ module RubyLLM
         end
 
         def extract_input_tokens(metadata_usage, usage, message_usage)
-          metadata_usage['inputTokens'] || usage['inputTokens'] || message_usage['input_tokens']
+          bedrock_usage = metadata_usage['inputTokens'] ? metadata_usage : usage
+          return Bedrock::Chat.input_tokens(bedrock_usage) if bedrock_usage['inputTokens']
+
+          message_usage['input_tokens']
         end
 
         def extract_output_tokens(metadata_usage, usage)
@@ -221,6 +228,7 @@ module RubyLLM
 
           reasoning_text = reasoning_content['reasoningText'] || {}
           return reasoning_text['text'] if reasoning_text['text']
+          return reasoning_content['text'] if reasoning_content['text']
           return event.dig('delta', 'thinking') if event.dig('delta', 'type') == 'thinking_delta'
 
           nil
@@ -241,6 +249,7 @@ module RubyLLM
           reasoning_content = delta['reasoningContent'] || {}
           reasoning_text = reasoning_content['reasoningText'] || {}
           return reasoning_text['signature'] if reasoning_text['signature']
+          return reasoning_content['signature'] if reasoning_content['signature']
           return event.dig('delta', 'signature') if event.dig('delta', 'type') == 'signature_delta'
 
           nil

data/lib/ruby_llm/providers/bedrock.rb

@@ -11,7 +11,7 @@ module RubyLLM
       include Bedrock::Streaming
 
       def api_base
-        "https://bedrock-runtime.#{bedrock_region}.amazonaws.com"
+        @config.bedrock_api_base || "https://bedrock-runtime.#{bedrock_region}.amazonaws.com"
       end
 
       def headers
@@ -54,7 +54,7 @@ module RubyLLM
 
       class << self
         def configuration_options
-          %i[bedrock_api_key bedrock_secret_key bedrock_region bedrock_session_token]
+          %i[bedrock_api_key bedrock_secret_key bedrock_region bedrock_session_token bedrock_api_base]
         end
 
         def configuration_requirements

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

@@ -7,6 +7,19 @@ module RubyLLM
       module Capabilities
         module_function
 
+        DEFAULT_CONTEXT_WINDOW = 1_000_000
+        DEFAULT_MAX_OUTPUT_TOKENS = 384_000
+        DEFAULT_PRICES = {
+          input: 0.14,
+          output: 0.28,
+          cache_read: 0.0028
+        }.freeze
+        PRO_PRICES = {
+          input: 0.435,
+          output: 0.87,
+          cache_read: 0.003625
+        }.freeze
+
         def supports_tool_choice?(_model_id)
           true
         end
@@ -14,6 +27,36 @@ module RubyLLM
         def supports_tool_parallel_control?(_model_id)
           false
         end
+
+        def context_window_for(_model_id)
+          DEFAULT_CONTEXT_WINDOW
+        end
+
+        def max_tokens_for(_model_id)
+          DEFAULT_MAX_OUTPUT_TOKENS
+        end
+
+        def critical_capabilities_for(model_id)
+          v4_model = model_id.start_with?('deepseek-v4-')
+          capabilities = ['function_calling']
+          capabilities << 'structured_output' if v4_model
+          capabilities << 'reasoning' if model_id == 'deepseek-reasoner' || v4_model
+          capabilities
+        end
+
+        def pricing_for(model_id)
+          prices = model_id == 'deepseek-v4-pro' ? PRO_PRICES : DEFAULT_PRICES
+
+          {
+            text_tokens: {
+              standard: {
+                input_per_million: prices[:input],
+                output_per_million: prices[:output],
+                cache_read_input_per_million: prices[:cache_read]
+              }
+            }
+          }
+        end
       end
     end
   end

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

@@ -10,6 +10,15 @@ module RubyLLM
         def format_role(role)
           role.to_s
         end
+
+        def format_content(content)
+          OpenAI::Media.format_content(
+            content,
+            document_attachments: :none,
+            image_attachments: false,
+            audio_attachments: false
+          )
+        end
       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/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/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/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
 

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

@@ -0,0 +1,55 @@
+# frozen_string_literal: true
+
+module RubyLLM
+  module Providers
+    class Mistral
+      # Handles media content for Mistral Chat Completions.
+      module Media
+        module_function
+
+        def format_content(content) # rubocop:disable Metrics/PerceivedComplexity
+          if content.is_a?(RubyLLM::Content::Raw)
+            value = content.value
+            return value.is_a?(Hash) ? value.to_json : value
+          end
+          return content.to_json if content.is_a?(Hash) || content.is_a?(Array)
+          return content unless content.is_a?(Content)
+
+          parts = []
+          parts << OpenAI::Media.format_text(content.text) if content.text
+
+          content.attachments.each do |attachment|
+            case attachment.type
+            when :image
+              parts << format_image(attachment)
+            when :audio
+              parts << OpenAI::Media.format_audio(attachment)
+            when :pdf, :document
+              parts << format_document(attachment)
+            when :text
+              parts << OpenAI::Media.format_text_file(attachment)
+            else
+              raise UnsupportedAttachmentError, attachment.mime_type
+            end
+          end
+
+          parts
+        end
+
+        def format_image(image)
+          {
+            type: 'image_url',
+            image_url: image.url? ? image.source.to_s : image.for_llm
+          }
+        end
+
+        def format_document(document)
+          {
+            type: 'document_url',
+            document_url: document.url? ? document.source.to_s : document.for_llm
+          }
+        end
+      end
+    end
+  end
+end

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

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

data/lib/ruby_llm/providers/mistral.rb

@@ -9,7 +9,7 @@ module RubyLLM
       include Mistral::Embeddings
 
       def api_base
-        'https://api.mistral.ai/v1'
+        @config.mistral_api_base || 'https://api.mistral.ai/v1'
       end
 
       def headers
@@ -24,7 +24,7 @@ module RubyLLM
         end
 
         def configuration_options
-          %i[mistral_api_key]
+          %i[mistral_api_key mistral_api_base]
         end
 
         def configuration_requirements

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

@@ -11,13 +11,20 @@ module RubyLLM
           messages.map do |msg|
             {
               role: format_role(msg.role),
-              content: Ollama::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 = Ollama::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