Register TogetherAI provider and add chat capabilities

lib/ruby_llm.rb

@@ -23,6 +23,7 @@
   'openrouter' => 'OpenRouter',
   'gpustack' => 'GPUStack',
   'mistral' => 'Mistral',
+  'togetherai' => 'TogetherAI',
   'vertexai' => 'VertexAI',
   'pdf' => 'PDF',
   'UI' => 'UI'
@@ -100,6 +101,7 @@ def logger
 RubyLLM::Provider.register :openrouter, RubyLLM::Providers::OpenRouter
 RubyLLM::Provider.register :perplexity, RubyLLM::Providers::Perplexity
 RubyLLM::Provider.register :vertexai, RubyLLM::Providers::VertexAI
+RubyLLM::Provider.register :togetherai, RubyLLM::Providers::TogetherAI
 
 if defined?(Rails::Railtie)
   require 'ruby_llm/railtie'

lib/ruby_llm/configuration.rb

@@ -24,6 +24,7 @@ class Configuration
                   :gpustack_api_base,
                   :gpustack_api_key,
                   :mistral_api_key,
+                  :togetherai_api_key,
                   # Default models
                   :default_model,
                   :default_embedding_model,

lib/ruby_llm/providers/togetherai.rb

@@ -0,0 +1,35 @@
+# frozen_string_literal: true
+
+module RubyLLM
+  module Providers
+    # Together.ai API integration.
+    class TogetherAI < Provider
+      include TogetherAI::Chat
+      include TogetherAI::Models
+
+      def api_base
+        'https://api.together.xyz/v1'
+      end
+
+      def headers
+        headers_hash = { 'Content-Type' => 'application/json' }
+
+        if @config.togetherai_api_key && [email protected]_api_key.empty?
+          headers_hash['Authorization'] = "Bearer #{@config.togetherai_api_key}"
+        end
+
+        headers_hash
+      end
+
+      class << self
+        def capabilities
+          TogetherAI::Capabilities
+        end
+
+        def configuration_requirements
+          %i[togetherai_api_key]
+        end
+      end
+    end
+  end
+end

lib/ruby_llm/providers/togetherai/capabilities.rb

@@ -0,0 +1,273 @@
+# frozen_string_literal: true
+
+module RubyLLM
+  module Providers
+    class TogetherAI
+      # Capabilities for the Together.ai provider
+      module Capabilities
+        def self.supports_streaming?(model_id)
+          # Most chat models support streaming, exclude specialized non-chat models
+          supports_chat_for?(model_id)
+        end
+
+        def self.supports_vision?(model_id)
+          supports_vision_for?(model_id)
+        end
+
+        def self.supports_functions?(model_id)
+          supports_tools_for?(model_id)
+        end
+
+        def self.supports_json_mode?(model_id)
+          # Most chat models support JSON mode, exclude specialized models
+          supports_chat_for?(model_id) && !model_id.match?(/whisper|voxtral/i)
+        end
+
+        def self.model_type(model_id)
+          return 'embedding' if supports_embeddings_for?(model_id)
+          return 'image' if supports_images_for?(model_id)
+          return 'audio' if supports_audio_for?(model_id)
+          return 'moderation' if supports_moderation_for?(model_id)
+
+          'chat'
+        end
+
+        def self.normalize_temperature(temperature, _model)
+          # Together.ai accepts temperature values between 0.0 and 2.0
+          return temperature if temperature.nil?
+
+          temperature.clamp(0.0, 2.0)
+        end
+
+        def self.max_tokens_for_model(_model)
+          # Default max tokens for Together.ai models
+          # This would ideally be model-specific
+          4096
+        end
+
+        def self.format_display_name(model_id)
+          model_id.split('/').last.tr('-', ' ').titleize
+        end
+
+        def self.model_family(model_id)
+          case model_id
+          when /llama/i then 'llama'
+          when /qwen/i then 'qwen'
+          when /mistral/i then 'mistral'
+          when /deepseek/i then 'deepseek'
+          when /gemma/i then 'gemma'
+          when /moonshot/i then 'kimi'
+          when /glm/i then 'glm'
+          when /cogito/i then 'cogito'
+          when /arcee/i then 'arcee'
+          when /marin/i then 'marin'
+          when /gryphe/i then 'mythomax'
+          when /openai/i then 'openai'
+          else 'other'
+          end
+        end
+
+        def self.context_window_for(model_id)
+          # Context windows based on Together.ai model specifications
+          # Using a hash lookup for better performance and maintainability
+          context_windows = {
+            # Ultra large context (1M+ tokens)
+            'meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8' => 524_288,
+            'meta-llama/Llama-4-Scout-17B-16E-Instruct' => 327_680,
+
+            # 256K+ context models
+            'moonshotai/Kimi-K2-Instruct-0905' => 262_144,
+            'moonshotai/Kimi-K2-Thinking' => 262_144,
+            'Qwen/Qwen3-235B-A22B-Thinking-2507' => 262_144,
+            'Qwen/Qwen3-235B-A22B-Instruct-2507-tput' => 262_144,
+            'Qwen/Qwen3-Next-80B-A3B-Instruct' => 262_144,
+            'Qwen/Qwen3-Next-80B-A3B-Thinking' => 262_144,
+            'Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8' => 256_000,
+
+            # ~200K context models
+            'zai-org/GLM-4.6' => 202_752,
+
+            # ~160K context models
+            'deepseek-ai/DeepSeek-R1' => 163_839,
+            'deepseek-ai/DeepSeek-R1-0528-tput' => 163_839,
+            'deepseek-ai/DeepSeek-V3' => 163_839,
+
+            # ~130K context models
+            'meta-llama/Llama-3.3-70B-Instruct-Turbo' => 131_072,
+            'meta-llama/Meta-Llama-3.1-8B-Instruct-Turbo' => 131_072,
+            'meta-llama/Meta-Llama-3.1-405B-Instruct-Turbo' => 130_815,
+            'meta-llama/Llama-3.2-3B-Instruct-Turbo' => 131_072,
+            'deepseek-ai/DeepSeek-R1-Distill-Llama-70B' => 131_072,
+            'deepseek-ai/DeepSeek-R1-Distill-Qwen-14B' => 131_072,
+            'zai-org/GLM-4.5-Air-FP8' => 131_072,
+
+            # ~128K context models
+            'moonshotai/Kimi-K2-Instruct' => 128_000,
+            'deepseek-ai/DeepSeek-V3.1' => 128_000,
+            'openai/gpt-oss-120b' => 128_000,
+            'openai/gpt-oss-20b' => 128_000,
+            'arcee-ai/virtuoso-medium-v2' => 128_000,
+            'arcee-ai/virtuoso-large' => 128_000,
+            'arcee-ai/maestro-reasoning' => 128_000,
+            'arcee_ai/arcee-spotlight' => 128_000,
+
+            # ~40K context models
+            'Qwen/Qwen3-235B-A22B-fp8-tput' => 40_960,
+            'mistralai/Magistral-Small-2506' => 40_960,
+
+            # ~32K context models (most common)
+            'Qwen/Qwen2.5-7B-Instruct-Turbo' => 32_768,
+            'Qwen/Qwen2.5-72B-Instruct-Turbo' => 32_768,
+            'Qwen/Qwen2.5-VL-72B-Instruct' => 32_768,
+            'Qwen/Qwen2.5-Coder-32B-Instruct' => 32_768,
+            'Qwen/QwQ-32B' => 32_768,
+            'mistralai/Mistral-Small-24B-Instruct-2501' => 32_768,
+            'mistralai/Mistral-7B-Instruct-v0.2' => 32_768,
+            'mistralai/Mistral-7B-Instruct-v0.3' => 32_768,
+            'google/gemma-3n-E4B-it' => 32_768,
+            'arcee-ai/coder-large' => 32_768,
+            'arcee-ai/caller' => 32_768,
+            'arcee-ai/arcee-blitz' => 32_768,
+
+            # ~8K context models
+            'meta-llama/Llama-3.3-70B-Instruct-Turbo-Free' => 8_193,
+            'meta-llama/Meta-Llama-3-8B-Instruct-Lite' => 8_192,
+            'meta-llama/Llama-3-70b-chat-hf' => 8_192,
+            'mistralai/Mistral-7B-Instruct-v0.1' => 8_192,
+            'google/gemma-2b-it' => 8_192,
+
+            # ~4K context models
+            'marin-community/marin-8b-instruct' => 4_096,
+            'Gryphe/MythoMax-L2-13b' => 4_096
+          }
+
+          # Check for exact match first
+          return context_windows[model_id] if context_windows.key?(model_id)
+
+          # Pattern matching for model families
+          case model_id
+          when %r{^deepcogito/cogito-v2.*} then 32_768
+          when %r{^Qwen/Qwen3.*235B.*} then 262_144
+          when %r{^meta-llama/Llama-4.*} then 1_048_576
+          else 16_384 # Default context window for unknown models
+          end
+        end
+
+        def self.max_tokens_for(model_id)
+          max_tokens_for_model(model_id)
+        end
+
+        def self.modalities_for(model_id)
+          input_modalities = ['text']
+          output_modalities = ['text']
+
+          input_modalities << 'image' if supports_vision_for?(model_id)
+          input_modalities << 'audio' if supports_audio_for?(model_id) && !model_id.match?(/sonic/i)
+
+          output_modalities = ['image'] if supports_images_for?(model_id)
+          output_modalities << 'audio' if model_id.match?(/sonic|voxtral/i)
+
+          { input: input_modalities, output: output_modalities }
+        end
+
+        def self.capabilities_for(model_id)
+          capabilities = primary_capabilities(model_id)
+          capabilities.concat(chat_capabilities(model_id)) if supports_chat_for?(model_id)
+          capabilities.concat(specialized_capabilities(model_id))
+          capabilities
+        end
+
+        def self.primary_capabilities(model_id)
+          [].tap do |caps|
+            caps << 'chat' if supports_chat_for?(model_id)
+            caps << 'embeddings' if supports_embeddings_for?(model_id)
+            caps << 'images' if supports_images_for?(model_id)
+          end
+        end
+
+        def self.chat_capabilities(model_id)
+          [].tap do |caps|
+            caps << 'streaming' if supports_streaming?(model_id)
+            caps << 'tools' if supports_tools_for?(model_id)
+            caps << 'json_mode' if supports_json_mode?(model_id)
+            caps << 'vision' if supports_vision_for?(model_id)
+          end
+        end
+
+        def self.specialized_capabilities(model_id)
+          [].tap do |caps|
+            caps << 'transcription' if supports_audio_for?(model_id)
+            caps << 'moderation' if supports_moderation_for?(model_id)
+          end
+        end
+
+        def self.supports_tools_for?(model_id)
+          # Most chat models support function calling, exclude non-chat models
+          return false if supports_embeddings_for?(model_id)
+          return false if supports_images_for?(model_id)
+          return false if supports_audio_for?(model_id)
+          return false if supports_moderation_for?(model_id)
+
+          true
+        end
+
+        def self.supports_chat_for?(model_id)
+          # Chat models are the main category, exclude non-chat models
+          return false if supports_embeddings_for?(model_id)
+          return false if supports_images_for?(model_id)
+          return false if supports_audio_for?(model_id) && !supports_vision_for?(model_id)
+          return false if supports_moderation_for?(model_id) && !model_id.match?(/Llama-Guard/i)
+
+          true
+        end
+
+        def self.supports_embeddings_for?(model_id)
+          # Embedding models
+          model_id.match?(/bge-|m2-bert|gte-|multilingual-e5/i)
+        end
+
+        # Methods for detecting different model capabilities
+        def self.supports_images_for?(model_id)
+          # Image generation models (FLUX, Stable Diffusion, Imagen)
+          model_id.match?(/FLUX|stable-diffusion|imagen/i)
+        end
+
+        def self.supports_vision_for?(model_id)
+          # Vision models (multimodal models that can process images)
+          model_id.match?(/Scout|VL|spotlight/i)
+        end
+
+        def self.supports_video_for?(_model_id)
+          false # Video generation support will be added in future PR
+        end
+
+        def self.supports_audio_for?(model_id)
+          # Audio models (TTS and transcription)
+          model_id.match?(/sonic|whisper|voxtral|orpheus/i)
+        end
+
+        def self.supports_transcription_for?(model_id)
+          # Transcription-specific models
+          model_id.match?(/whisper/i)
+        end
+
+        def self.supports_moderation_for?(model_id)
+          # Moderation models
+          model_id.match?(/Guard|VirtueGuard/i)
+        end
+
+        def self.supports_rerank_for?(_model_id)
+          false # Rerank support will be added in future PR
+        end
+
+        def self.pricing_for(_model_id)
+          # Placeholder pricing - should be model-specific
+          {
+            input_tokens: 0.001,
+            output_tokens: 0.002
+          }
+        end
+      end
+    end
+  end
+end