smart_prompt 0.5.1 → 0.5.3

data/lib/smart_prompt/concerns/multimodal_messages.rb

@@ -0,0 +1,108 @@
+require "base64"
+require "net/http"
+require "uri"
+
+module SmartPrompt
+  # Shared multimodal-message normalization for Net::HTTP adapters (ZhipuAI, SenseNova,
+  # SiliconFlow). Turns an OpenAI-style content array into the shape the provider expects,
+  # inlining local image/audio/video files as base64 data URLs and passing http(s)/data
+  # URLs through. Each adapter previously carried a near-identical copy of this logic.
+  #
+  # SiliconFlow's variant is the superset (image_url + video_url + audio_url, preserving
+  # detail/max_frames/fps); Zhipu/SenseNova only ever send image_url, which is a subset.
+  module MultimodalMessages
+    SUPPORTED_IMAGE_FORMATS = %w[jpg jpeg png gif bmp webp].freeze
+
+    def process_multimodal_messages(messages)
+      messages.map do |msg|
+        role = msg[:role] || msg["role"]
+        content = msg[:content] || msg["content"]
+        content = content.map { |item| normalize_content_item(item) } if content.is_a?(Array)
+        { "role" => role, "content" => content }
+      end
+    end
+
+    def normalize_content_item(item)
+      return { "type" => "text", "text" => item.to_s } unless item.is_a?(Hash)
+
+      case item[:type] || item["type"]
+      when "image_url"
+        normalize_media_part(item, "image_url", :image)
+      when "video_url"
+        normalize_media_part(item, "video_url", :video)
+      when "audio_url"
+        normalize_media_part(item, "audio_url", :audio)
+      else
+        stringify_hash(item)
+      end
+    end
+
+    # Build an image_url/video_url/audio_url part, inlining local files as data URLs and
+    # preserving any extra keys (detail, max_frames, fps) on the media hash.
+    def normalize_media_part(item, type, media_kind)
+      iu = item[type.to_sym] || item[type]
+      if iu.is_a?(Hash)
+        url = iu[:url] || iu["url"]
+        part = { "type" => type, type => { "url" => normalize_media_url(url, media_kind) } }
+        iu.each { |k, v| part[type][k.to_s] = stringify_hash(v) unless k.to_s == "url" }
+        part
+      else
+        { "type" => type, type => { "url" => normalize_media_url(iu, media_kind) } }
+      end
+    end
+
+    # Resolve a media URL embedded in a message: http(s)/data pass through; a local path
+    # is base64-encoded as a data URL.
+    def normalize_media_url(url, kind = :image)
+      return url if url.nil?
+      return url if url.start_with?("http://", "https://", "data:")
+
+      label = kind == :image ? "Image" : kind.to_s.capitalize
+      raise Error, "#{label} file not found: #{url}" unless File.exist?(url)
+      ext = File.extname(url).downcase.delete(".")
+      case kind
+      when :image
+        raise Error, "Unsupported image format: #{ext}" unless SUPPORTED_IMAGE_FORMATS.include?(ext)
+        mime = ext == "jpg" ? "jpeg" : ext
+        "data:image/#{mime};base64,#{Base64.strict_encode64(File.binread(url))}"
+      when :audio
+        "data:audio/#{ext.empty? ? 'wav' : ext};base64,#{Base64.strict_encode64(File.binread(url))}"
+      when :video
+        "data:video/#{ext.empty? ? 'mp4' : ext};base64,#{Base64.strict_encode64(File.binread(url))}"
+      end
+    end
+
+    # Single-arg image-only shim (call sites like generate_video pass a plain image URL).
+    def normalize_image_url(url)
+      normalize_media_url(url, :image)
+    end
+
+    # Accept a local path, a base64 data URL, or an http(s) URL for image-edit /
+    # image-to-video `image` fields.
+    def normalize_input_image(image)
+      return image if image.nil?
+
+      if image.is_a?(String)
+        return image if image.start_with?("data:")
+        return image if image.start_with?("http://", "https://")
+      end
+
+      raise Error, "Image file not found: #{image}" unless File.exist?(image)
+      ext = File.extname(image).downcase.delete(".")
+      raise Error, "Unsupported image format: #{ext}" unless SUPPORTED_IMAGE_FORMATS.include?(ext)
+      mime = ext == "jpg" ? "jpeg" : ext
+      "data:image/#{mime};base64,#{Base64.strict_encode64(File.binread(image))}"
+    end
+
+    def stringify_hash(hash)
+      case hash
+      when Hash
+        hash.each_with_object({}) { |(k, v), memo| memo[k.to_s] = stringify_hash(v) }
+      when Array
+        hash.map { |v| stringify_hash(v) }
+      else
+        hash
+      end
+    end
+  end
+end

data/lib/smart_prompt/concerns/openai_chat_shaping.rb

@@ -0,0 +1,87 @@
+module SmartPrompt
+  # Shared shaping of Net::HTTP chat responses into the OpenAI completion / stream
+  # shape that the rest of SmartPrompt (Engine#@stream_proc, Conversation) expects.
+  #
+  # Reasoning models expose a thinking trace under a provider-specific field —
+  # surfaced here uniformly as `reasoning_content`. Adapters override one hook:
+  #
+  #   reasoning_field_name  — the source field on message/delta (default
+  #     "reasoning_content"; SenseNova uses "reasoning"). Its value is remapped to
+  #     reasoning_content so Engine#@stream_proc needs no per-provider logic.
+  #
+  #   extra_top_level_fields(raw) — extra top-level keys to copy onto the shaped
+  #     response/chunk (default {}; SenseNova adds system_fingerprint).
+  module OpenAIChatShaping
+    def build_completion_response(raw)
+      msg = raw.dig("choices", 0, "message") || {}
+      message = { "role" => msg["role"] || "assistant" }
+      message["content"] = msg["content"]
+      reasoning = msg[reasoning_field_name]
+      message["reasoning_content"] = reasoning if reasoning
+      message["tool_calls"] = msg["tool_calls"] if msg["tool_calls"]
+
+      response = {
+        "id"      => raw["id"],
+        "object"  => raw["object"] || "chat.completion",
+        "created" => raw["created"],
+        "model"   => raw["model"],
+        "choices" => [{
+          "index"         => 0,
+          "message"       => message,
+          "finish_reason" => raw.dig("choices", 0, "finish_reason"),
+        }],
+      }
+      response["usage"] = raw["usage"] if raw["usage"]
+      merge_extra_top_level(response, raw)
+      response
+    end
+
+    def build_stream_chunk(data)
+      chunk = {
+        "id"      => data["id"],
+        "object"  => data["object"],
+        "created" => data["created"],
+        "model"   => data["model"],
+      }
+      chunk["usage"] = data["usage"] if data["usage"]
+      merge_extra_top_level(chunk, data)
+
+      choices = data["choices"] || []
+      if choices.any?
+        delta = choices[0]["delta"] || {}
+        new_delta = {}
+        new_delta["role"]              = delta["role"]        if delta["role"]
+        new_delta["content"]           = delta["content"]     if delta["content"]
+        reasoning = delta[reasoning_field_name]
+        new_delta["reasoning_content"] = reasoning if reasoning
+        new_delta["tool_calls"]        = delta["tool_calls"]  if delta["tool_calls"]
+        chunk["choices"] = [{
+          "index"         => choices[0]["index"] || 0,
+          "delta"         => new_delta,
+          "finish_reason" => choices[0]["finish_reason"],
+        }]
+      else
+        chunk["choices"] = []
+      end
+      chunk
+    end
+
+    # ---- hooks (override in adapter) -----------------------------------------
+
+    def reasoning_field_name
+      "reasoning_content"
+    end
+
+    def extra_top_level_fields(_raw)
+      {}
+    end
+
+    private
+
+    def merge_extra_top_level(target, raw)
+      extra_top_level_fields(raw).each do |k, v|
+        target[k] = v unless v.nil? || target.key?(k)
+      end
+    end
+  end
+end

data/lib/smart_prompt/sensenova_adapter.rb

@@ -3,6 +3,10 @@ require "json"
 require "net/http"
 require "uri"
 require "fileutils"
+require_relative "concerns/image_persistence"
+require_relative "concerns/openai_chat_shaping"
+require_relative "concerns/multimodal_messages"
+require_relative "concerns/http_client"
 
 module SmartPrompt
   # Adapter for SenseNova (商汤 日日新) — the SenseCore large-model platform.
@@ -38,7 +42,6 @@ module SmartPrompt
       2048x2048 2752x1536 1536x2752 3072x1376 1344x3136 2560x720 3072x864
     ].freeze
     DEFAULT_IMAGE_SIZE = "2048x2048".freeze
-    SUPPORTED_IMAGE_FORMATS = %w[jpg jpeg png gif bmp webp].freeze
 
     # SenseNova sampling parameters forwarded from config to the chat request when present.
     CHAT_OPTIONAL_KEYS = %w[
@@ -46,6 +49,31 @@ module SmartPrompt
       reasoning_effort max_completion_tokens max_tokens
     ].freeze
 
+    include ImagePersistence
+    include OpenAIChatShaping
+    include MultimodalMessages
+    include HTTPClient
+
+    # ---- hooks for shared concerns -------------------------------------------
+    def provider_label
+      "SenseNova"
+    end
+
+    def default_image_prefix
+      "sensenova_image"
+    end
+
+    # SenseNova exposes the reasoning trace under `reasoning` (not reasoning_content)
+    # and also returns system_fingerprint — override the OpenAIChatShaping hooks so the
+    # shared shaper still produces the uniform reasoning_content / fingerprint output.
+    def reasoning_field_name
+      "reasoning"
+    end
+
+    def extra_top_level_fields(raw)
+      { "system_fingerprint" => raw["system_fingerprint"] }
+    end
+
     def initialize(config)
       super
       SmartPrompt.logger.info "Start create the SmartPrompt SenseNovaAdapter."
@@ -166,16 +194,7 @@ module SmartPrompt
       images
     end
 
-    # Save one or many generated images to disk (Array from #generate_image or a single hash).
-    def save_image(image_data, output_dir = "./output", filename_prefix = "sensenova_image")
-      FileUtils.mkdir_p(output_dir)
-      images = image_data.is_a?(Array) ? image_data : [image_data]
-      saved = images.each_with_index.map do |img, index|
-        save_single_image(img, output_dir, "#{filename_prefix}_#{index + 1}")
-      end
-      SmartPrompt.logger.info "Saved #{saved.size} SenseNova image(s) to #{output_dir}"
-      saved
-    end
+    # (save_image / save_single_image provided by ImagePersistence concern.)
 
     private
 
@@ -192,171 +211,10 @@ module SmartPrompt
       body
     end
 
-    # Pass messages through, normalizing any multimodal content. Local image paths inside
-    # image_url.url are converted to data: URLs; http(s)/data URLs and plain text pass through.
-    def process_multimodal_messages(messages)
-      messages.map do |msg|
-        role = msg[:role] || msg["role"]
-        content = msg[:content] || msg["content"]
-        content = content.map { |item| normalize_content_item(item) } if content.is_a?(Array)
-        { "role" => role, "content" => content }
-      end
-    end
-
-    def normalize_content_item(item)
-      return { "type" => "text", "text" => item.to_s } unless item.is_a?(Hash)
-
-      type = item[:type] || item["type"]
-      if type == "image_url"
-        iu = item[:image_url] || item["image_url"]
-        url = iu.is_a?(Hash) ? (iu[:url] || iu["url"]) : iu
-        { "type" => "image_url", "image_url" => { "url" => normalize_image_url(url) } }
-      else
-        stringify_hash(item)
-      end
-    end
-
-    def normalize_image_url(url)
-      return url if url.nil?
-      return url if url.start_with?("http://", "https://", "data:")
+    # (process_multimodal_messages / normalize_* / stringify_hash provided by MultimodalMessages concern.)
+    # (build_completion_response / build_stream_chunk provided by OpenAIChatShaping concern.)
 
-      raise Error, "Image file not found: #{url}" unless File.exist?(url)
-      ext = File.extname(url).downcase.delete(".")
-      raise Error, "Unsupported image format: #{ext}" unless SUPPORTED_IMAGE_FORMATS.include?(ext)
-      mime = ext == "jpg" ? "jpeg" : ext
-      "data:image/#{mime};base64,#{Base64.strict_encode64(File.binread(url))}"
-    end
-
-    # ---- response shaping -----------------------------------------------------
-
-    # Convert a non-streaming SenseNova response into the OpenAI completion shape the
-    # rest of SmartPrompt expects, surfacing the reasoning model's `reasoning` field.
-    def build_completion_response(raw)
-      msg = raw.dig("choices", 0, "message") || {}
-      message = { "role" => msg["role"] || "assistant" }
-      message["content"] = msg["content"]
-      message["reasoning_content"] = msg["reasoning"] if msg["reasoning"]
-      message["tool_calls"] = msg["tool_calls"] if msg["tool_calls"]
-
-      response = {
-        "id"      => raw["id"],
-        "object"  => raw["object"] || "chat.completion",
-        "created" => raw["created"],
-        "model"   => raw["model"],
-        "choices" => [{
-          "index"         => 0,
-          "message"       => message,
-          "finish_reason" => raw.dig("choices", 0, "finish_reason"),
-        }],
-      }
-      response["usage"] = raw["usage"] if raw["usage"]
-      response["system_fingerprint"] = raw["system_fingerprint"] if raw["system_fingerprint"]
-      response
-    end
-
-    # Convert one SSE event from SenseNova's stream into an OpenAI-style streaming chunk.
-    # The key remap is delta.reasoning -> delta.reasoning_content, which is what
-    # Engine#@stream_proc reads for reasoning models.
-    def build_stream_chunk(data)
-      chunk = {
-        "id"      => data["id"],
-        "object"  => data["object"],
-        "created" => data["created"],
-        "model"   => data["model"],
-      }
-      chunk["usage"]             = data["usage"]             if data["usage"]
-      chunk["system_fingerprint"] = data["system_fingerprint"] if data["system_fingerprint"]
-
-      choices = data["choices"] || []
-      if choices.any?
-        delta = choices[0]["delta"] || {}
-        new_delta = {}
-        new_delta["role"]              = delta["role"]      if delta["role"]
-        new_delta["content"]           = delta["content"]   if delta["content"]
-        new_delta["reasoning_content"] = delta["reasoning"] if delta["reasoning"]
-        new_delta["tool_calls"]        = delta["tool_calls"] if delta["tool_calls"]
-        chunk["choices"] = [{
-          "index"         => choices[0]["index"] || 0,
-          "delta"         => new_delta,
-          "finish_reason" => choices[0]["finish_reason"],
-        }]
-      else
-        # Usage-only final event (choices is an empty array).
-        chunk["choices"] = []
-      end
-      chunk
-    end
-
-    # ---- HTTP -----------------------------------------------------------------
-
-    def http_post_json(url, body)
-      uri = URI.parse(url)
-      http = Net::HTTP.new(uri.host, uri.port)
-      http.use_ssl     = (uri.scheme == "https")
-      http.open_timeout = 30
-      http.read_timeout = 240
-
-      request = Net::HTTP::Post.new(uri.request_uri)
-      request["Content-Type"]  = "application/json"
-      request["Authorization"] = "Bearer #{@api_key}"
-      request.body = body.to_json
-
-      SmartPrompt.logger.debug "SenseNova POST #{uri} body=#{body.to_json}"
-      response = http.request(request)
-
-      if response.is_a?(Net::HTTPSuccess)
-        JSON.parse(response.body)
-      else
-        SmartPrompt.logger.error "SenseNova API error: #{response.code} - #{response.body}"
-        raise LLMAPIError, "SenseNova API error: #{response.code} - #{response.body}"
-      end
-    end
-
-    # POST with stream:true and yield each parsed SSE `data:` payload to the block.
-    def stream_chat(url, body)
-      uri = URI.parse(url)
-      http = Net::HTTP.new(uri.host, uri.port)
-      http.use_ssl      = (uri.scheme == "https")
-      http.open_timeout  = 30
-      http.read_timeout  = 300
-
-      request = Net::HTTP::Post.new(uri.request_uri)
-      request["Content-Type"]  = "application/json"
-      request["Authorization"] = "Bearer #{@api_key}"
-      request["Accept"]        = "text/event-stream"
-      request.body = body.to_json
-
-      buffer = ""
-      done = false
-
-      http.request(request) do |response|
-        unless response.is_a?(Net::HTTPSuccess)
-          raise LLMAPIError, "SenseNova stream error: #{response.code} - #{response.body}"
-        end
-
-        response.read_body do |segment|
-          break if done
-          buffer << segment
-          while (idx = buffer.index("\n"))
-            line = buffer.slice!(0, idx + 1).strip
-            next if line.empty? || !line.start_with?("data:")
-
-            payload = line.sub(/\Adata:\s*/, "")
-            if payload == "[DONE]"
-              done = true
-              break
-            end
-
-            begin
-              data = JSON.parse(payload)
-            rescue JSON::ParserError
-              next
-            end
-            yield data
-          end
-        end
-      end
-    end
+    # (http_post_json / stream_chat provided by HTTPClient concern.)
 
     # Resolve the image size: default to 2048x2048 when none given, and warn (but still
     # send) when the caller asks for a size sensenova-u1-fast does not accept.
@@ -370,41 +228,6 @@ module SmartPrompt
       size
     end
 
-    def save_single_image(image_data, output_dir, filename)
-      if image_data[:b64_json]
-        file_path = File.join(output_dir, "#{filename}.png")
-        File.binwrite(file_path, Base64.decode64(image_data[:b64_json]))
-      elsif image_data[:url]
-        uri = URI.parse(image_data[:url])
-        response = Net::HTTP.get_response(uri)
-        raise Error, "Failed to download image from URL: #{response.code}" unless response.is_a?(Net::HTTPSuccess)
-
-        ext = case response["content-type"]
-              when "image/jpeg", "image/jpg" then "jpg"
-              when "image/png"               then "png"
-              when "image/gif"               then "gif"
-              when "image/webp"              then "webp"
-              else "png"
-              end
-        file_path = File.join(output_dir, "#{filename}.#{ext}")
-        File.binwrite(file_path, response.body)
-      else
-        raise Error, "No image data available to save"
-      end
-      file_path
-    end
-
-    def stringify_hash(hash)
-      case hash
-      when Hash
-        hash.each_with_object({}) do |(k, v), memo|
-          memo[k.to_s] = stringify_hash(v)
-        end
-      when Array
-        hash.map { |v| stringify_hash(v) }
-      else
-        hash
-      end
-    end
+    # (stringify_hash provided by MultimodalMessages concern.)
   end
 end

data/lib/smart_prompt/siliconflow_adapter.rb

@@ -0,0 +1,91 @@
+require "base64"
+require "json"
+require "net/http"
+require "uri"
+require "fileutils"
+require_relative "concerns/image_persistence"
+require_relative "concerns/openai_chat_shaping"
+require_relative "concerns/multimodal_messages"
+require_relative "concerns/http_client"
+require_relative "adapters/siliconflow/text"
+require_relative "adapters/siliconflow/embed"
+require_relative "adapters/siliconflow/image"
+require_relative "adapters/siliconflow/video"
+require_relative "adapters/siliconflow/voice"
+require_relative "adapters/siliconflow/rerank"
+
+module SmartPrompt
+  # Adapter for 硅基流动 (SiliconFlow / SiliconCloud) — one adapter owns the whole
+  # provider: every category shares the base URL https://api.siliconflow.cn/v1 and
+  # Bearer auth.
+  #
+  # Per-modality behavior lives in capability modules under adapters/siliconflow/
+  # (Text / Embed / Image / Video / Voice / Rerank); cross-provider plumbing (HTTP,
+  # multimodal normalization, chat shaping, image saving) comes from the shared
+  # concerns. This class wires them together + holds config/credentials.
+  #
+  # Provider-specific quirks (all vs https://docs.siliconflow.cn/cn/api-reference):
+  #   chat/vision — POST {base}/chat/completions (reasoning_content, no remap)
+  #   embeddings  — POST {base}/embeddings        (dimensions only for Qwen3-Embedding)
+  #   rerank      — POST {base}/rerank            (results[].relevance_score)
+  #   image/edit  — POST {base}/images/generations (images[].url; image_size/batch_size/guidance_scale)
+  #   video       — POST {base}/video/submit -> POST {base}/video/status (async; results.videos[].url)
+  #   tts         — POST {base}/audio/speech       (binary audio response)
+  #   asr         — POST {base}/audio/transcriptions (multipart, field "file")
+  #   voice       — /uploads/audio/voice, /audio/voice/list, /audio/voice/deletions
+  class SiliconFlowAdapter < LLMAdapter
+    DEFAULT_BASE_URL = "https://api.siliconflow.cn/v1".freeze
+
+    # Cross-provider shared concerns
+    include ImagePersistence
+    include OpenAIChatShaping
+    include MultimodalMessages
+    include HTTPClient
+
+    # Per-capability modules
+    include SiliconFlow::Text
+    include SiliconFlow::Embed
+    include SiliconFlow::Image
+    include SiliconFlow::Video
+    include SiliconFlow::Voice
+    include SiliconFlow::Rerank
+
+    # ---- hooks for shared concerns -------------------------------------------
+    def provider_label
+      "SiliconFlow"
+    end
+
+    def default_image_prefix
+      "siliconflow_image"
+    end
+
+    def initialize(config)
+      super
+      SmartPrompt.logger.info "Start create the SmartPrompt SiliconFlowAdapter."
+
+      api_key = @config["api_key"]
+      if api_key.is_a?(String) && api_key.start_with?("ENV[") && api_key.end_with?("]")
+        api_key = eval(api_key)
+      end
+      # Tolerate a missing key at construction (e.g. when the ENV var isn't set yet)
+      # and let the first request fail with a clear auth error.
+      SmartPrompt.logger.warn "SiliconFlow api_key is empty — API calls will fail until it is set." if api_key.nil? || api_key.to_s.strip.empty?
+
+      @api_key  = api_key
+      @base_url = (@config["url"] || DEFAULT_BASE_URL).to_s.chomp("/")
+      # Optional per-method URL overrides (default to the standard paths off @base_url).
+      @image_url         = (@config["image_url"]         || "#{@base_url}/images/generations").to_s
+      @video_submit_url  = (@config["video_submit_url"]  || "#{@base_url}/video/submit").to_s
+      @video_status_url  = (@config["video_status_url"]  || "#{@base_url}/video/status").to_s
+      @speech_url        = (@config["speech_url"]        || "#{@base_url}/audio/speech").to_s
+      @transcription_url = (@config["transcription_url"] || "#{@base_url}/audio/transcriptions").to_s
+      @voice_upload_url  = (@config["voice_upload_url"]  || "#{@base_url}/uploads/audio/voice").to_s
+      @voice_list_url    = (@config["voice_list_url"]    || "#{@base_url}/audio/voice/list").to_s
+      @voice_delete_url  = (@config["voice_delete_url"]  || "#{@base_url}/audio/voice/deletions").to_s
+      SmartPrompt.logger.info "SiliconFlow base_url=#{@base_url}"
+    rescue => e
+      SmartPrompt.logger.error "Failed to initialize SiliconFlow client: #{e.message}"
+      raise e.is_a?(SmartPrompt::Error) ? e : LLMAPIError, "Invalid SiliconFlow configuration: #{e.message}"
+    end
+  end
+end

data/lib/smart_prompt/version.rb

@@ -1,3 +1,3 @@
 module SmartPrompt
-  VERSION = "0.5.1"
+  VERSION = "0.5.3"
 end