smart_prompt 0.5.2 → 0.5.3

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 1477a83b116691863533a9b1726e40be07b03b4faa67b60ebc72fe6a290d60f1
-  data.tar.gz: c9e71e998318d186f296495679573ccb2ad9b539420b3b9c3ee02314db8e2d8b
+  metadata.gz: 28b52596217dded1ca7b39d43e098de3b01f9db8e0a6acb06201399ef6c6877d
+  data.tar.gz: f29cb242e7900d1c1321451f05b5aa1075155653c6b894707ed7b172893d6833
 SHA512:
-  metadata.gz: a433f3724915b38af6e3a1e66d3f52568ec305f85d3b038643ffeb1ae5522547a8e5c6769e6c53cf1a081ca353587ddd0156987a4d2c705f25f267e388a7f5b9
-  data.tar.gz: 90b3ba6033912705cc096b17765f0ad221d6f8a6e6b3b2a4350345d4b0777ebdee359d30604320918891ee7965cb55528a0aab4c12b9d9eebd67237ff52d7876
+  metadata.gz: ba8c550d079e37c5b5155393522568bef857821cb1440d810517f37d1067b9b7276ef46e52a5ef9e0605e066e65b53a6dc100c4a42a71585d9d99d0354322160
+  data.tar.gz: 4a2cb608d7b3aeb757a813e75ef5710ad4d106811e743d91e3dbdbb9c07f44c6d184e334eb1528a8999b551b32bbf2c4ab6037bfaabb405055fd4ab053ca4eec

data/config/siliconflow_config.yml

@@ -0,0 +1,95 @@
+# 硅基流动 (SiliconFlow / SiliconCloud) Configuration for SmartPrompt
+#
+# Get an API key from https://cloud.siliconflow.cn/ and export it as
+# SILICONFLOW_API_KEY. A single SiliconFlowAdapter covers every model category —
+# just point a different llm entry at each model. Defaults use the free-tier models
+# so it runs out-of-box once the key is set.
+#
+# Base URL: https://api.siliconflow.cn/v1
+# Auth:     Authorization: Bearer <API_KEY>
+#
+# All categories share one base URL + Bearer auth. Notable SiliconFlow quirks
+# (handled by the adapter):
+#   * image response is images[].url  (not OpenAI's data[])
+#   * video is async:  POST /video/submit -> {requestId}; POST /video/status
+#     (body {requestId}) -> status Succeeded/InQueue/InProgress/Failed,
+#     result at results.videos[].url
+#   * rerank score field is relevance_score (not score)
+#   * TTS /audio/speech returns raw binary audio
+
+adapters:
+  siliconflow: "SiliconFlowAdapter"
+
+llms:
+  # 1. 文本对话 (free-tier Qwen2.5-7B; reasoning: deepseek-ai/DeepSeek-R1)
+  sf_chat:
+    adapter: "siliconflow"
+    url: "https://api.siliconflow.cn/v1"
+    api_key: ENV["SILICONFLOW_API_KEY"]
+    model: "Qwen/Qwen2.5-7B-Instruct"
+    temperature: 0.7
+    # Optional sampling extras (forwarded to /chat/completions):
+    # top_p: 0.7
+    # max_tokens: 2048
+    # enable_thinking: true   # for Qwen3 / GLM thinking models
+    # thinking_budget: 4096
+
+  # 2. 多模态 (vision; Qwen2.5-VL. audio/video needs an omni model like Qwen3-Omni)
+  sf_vision:
+    adapter: "siliconflow"
+    url: "https://api.siliconflow.cn/v1"
+    api_key: ENV["SILICONFLOW_API_KEY"]
+    model: "Qwen/Qwen2.5-VL-72B-Instruct"
+    # Paid alternatives: Qwen/Qwen3-VL-32B-Instruct, Qwen/Qwen3-Omni-30B-A3B-Instruct
+
+  # 3. 向量模型 (BAAI/bge-m3 — fixed 1024 dims. For custom dimensions use
+  #    Qwen/Qwen3-Embedding-8B and set `dimensions`.)
+  sf_embed:
+    adapter: "siliconflow"
+    url: "https://api.siliconflow.cn/v1"
+    api_key: ENV["SILICONFLOW_API_KEY"]
+    model: "BAAI/bge-m3"
+    # dimensions: 1024   # only honored by Qwen3-Embedding series
+
+  # 4. 重排 (rerank). relevance_score field, not score.
+  sf_rerank:
+    adapter: "siliconflow"
+    url: "https://api.siliconflow.cn/v1"
+    api_key: ENV["SILICONFLOW_API_KEY"]
+    model: "BAAI/bge-reranker-v2-m3"
+
+  # 5. 文生图 (free-tier Kolors; edit: Qwen/Qwen-Image-Edit-2509)
+  sf_image:
+    adapter: "siliconflow"
+    url: "https://api.siliconflow.cn/v1"
+    api_key: ENV["SILICONFLOW_API_KEY"]
+    model: "Kwai-Kolors/Kolors"
+
+  # 6. 文生视频 / 图生视频 (async; only Wan2.2 series is available)
+  sf_video:
+    adapter: "siliconflow"
+    url: "https://api.siliconflow.cn/v1"
+    api_key: ENV["SILICONFLOW_API_KEY"]
+    model: "Wan-AI/Wan2.2-T2V-A14B"
+    # image-to-video: model "Wan-AI/Wan2.2-I2V-A14B" and pass an image
+
+  # 7. 语音合成 (CosyVoice2; preset voices need the model prefix, e.g.
+  #    "FunAudioLLM/CosyVoice2-0.5B:alex")
+  sf_tts:
+    adapter: "siliconflow"
+    url: "https://api.siliconflow.cn/v1"
+    api_key: ENV["SILICONFLOW_API_KEY"]
+    model: "FunAudioLLM/CosyVoice2-0.5B"
+
+  # 8. 语音识别 (SenseVoiceSmall, free, multilingual)
+  sf_asr:
+    adapter: "siliconflow"
+    url: "https://api.siliconflow.cn/v1"
+    api_key: ENV["SILICONFLOW_API_KEY"]
+    model: "FunAudioLLM/SenseVoiceSmall"
+
+default_llm: "sf_chat"
+
+template_path: "./templates"
+worker_path: "./workers"
+logger_file: "./logs/smart_prompt.log"

data/examples/siliconflow_example.rb

@@ -0,0 +1,175 @@
+# 硅基流动 (SiliconFlow / SiliconCloud) Example for SmartPrompt
+#
+# Demonstrates every SiliconFlow model category through one SiliconFlowAdapter:
+#   1. 文本对话 (chat) — sync + streaming
+#   2. 多模态 (vision)
+#   3. 向量模型 (embeddings)
+#   4. 重排 (rerank)
+#   5. 文生图 (image)
+#   6. 文生视频 (async submit -> poll -> download)
+#   7. 语音合成 (TTS)
+#   8. 语音识别 (ASR)
+#
+# Requires a valid SiliconFlow API key in SILICONFLOW_API_KEY
+# (https://cloud.siliconflow.cn/). Defaults use free-tier models so it works
+# out-of-box once the key is set.
+
+require_relative "../lib/smart_prompt"
+
+api_key = ENV["SILICONFLOW_API_KEY"]
+base = "https://api.siliconflow.cn/v1"
+
+config = {
+  "adapters" => { "siliconflow" => "SiliconFlowAdapter" },
+  "llms" => {
+    "sf_chat"   => { "adapter" => "siliconflow", "url" => base, "api_key" => api_key, "model" => "Qwen/Qwen2.5-7B-Instruct" },
+    "sf_vision" => { "adapter" => "siliconflow", "url" => base, "api_key" => api_key, "model" => "Qwen/Qwen2.5-VL-72B-Instruct" },
+    "sf_embed"  => { "adapter" => "siliconflow", "url" => base, "api_key" => api_key, "model" => "BAAI/bge-m3" },
+    "sf_rerank" => { "adapter" => "siliconflow", "url" => base, "api_key" => api_key, "model" => "BAAI/bge-reranker-v2-m3" },
+    "sf_image"  => { "adapter" => "siliconflow", "url" => base, "api_key" => api_key, "model" => "Kwai-Kolors/Kolors" },
+    "sf_video"  => { "adapter" => "siliconflow", "url" => base, "api_key" => api_key, "model" => "Wan-AI/Wan2.2-T2V-A14B" },
+    "sf_tts"    => { "adapter" => "siliconflow", "url" => base, "api_key" => api_key, "model" => "FunAudioLLM/CosyVoice2-0.5B" },
+    "sf_asr"    => { "adapter" => "siliconflow", "url" => base, "api_key" => api_key, "model" => "FunAudioLLM/SenseVoiceSmall" },
+  },
+  "default_llm" => "sf_chat",
+  "template_path" => "./templates",
+  "worker_path" => "./workers",
+  "logger_file" => "./logs/smart_prompt.log",
+}
+
+File.write("siliconflow_config.yml", config.to_yaml)
+engine = SmartPrompt::Engine.new("siliconflow_config.yml")
+
+puts "=== SmartPrompt 硅基流动 SiliconFlow Demo ==="
+unless api_key
+  puts "Note: SILICONFLOW_API_KEY is not set — the API calls below will fail at the network layer."
+end
+
+# 1. Chat (sync)
+puts "\n=== Example 1: 文本对话 (sync) ==="
+begin
+  result = engine.call_worker(:siliconflow_chat, { prompt: "用一句话介绍硅基流动 SiliconFlow。" })
+  puts "Reply: #{result}"
+rescue => e
+  puts "Error: #{e.message}"
+end
+
+# 2. Chat (streaming)
+puts "\n=== Example 2: 文本对话 (streaming) ==="
+begin
+  engine.call_worker_by_stream(:siliconflow_chat, { prompt: "写两句关于春天的诗。" }) do |chunk, _|
+    print chunk.dig("choices", 0, "delta", "content").to_s
+  end
+  puts
+rescue => e
+  puts "Error: #{e.message}"
+end
+
+# 3. Multimodal vision
+puts "\n=== Example 3: 多模态 ==="
+begin
+  result = engine.call_worker(:siliconflow_vision, {
+    image_url: "https://img1.baidu.com/it/u=1966616150,2146512490&fm=253&fmt=auto&app=138&f=JPEG?w=500&h=282",
+    question: "图片里有什么？",
+  })
+  puts "Vision result: #{result}"
+rescue => e
+  puts "Error: #{e.message}"
+end
+
+# 4. Embeddings (BAAI/bge-m3)
+puts "\n=== Example 4: 向量模型 ==="
+begin
+  vector = engine.call_worker(:siliconflow_embed, { text: "硅基流动大模型", length: 1024 })
+  puts "Embedding dim: #{vector.is_a?(Array) ? vector.size : vector} (first 5: #{vector.first(5) rescue vector})"
+rescue => e
+  puts "Error: #{e.message}"
+end
+
+# 5. Rerank — reorder documents by relevance to a query.
+puts "\n=== Example 5: 重排 (rerank) ==="
+begin
+  result = engine.call_worker(:siliconflow_rerank, {
+    query: "如何用 Python 读取文件？",
+    documents: [
+      "在 Python 中可以用 open() 函数打开文件并读取内容。",
+      "JavaScript 是一种运行在浏览器中的脚本语言。",
+      "使用 with open(path) as f 可以安全地读取文件。",
+    ],
+    top_n: 2,
+  })
+  result.each { |r| puts "  idx=#{r[:index]} score=#{r[:relevance_score]}" }
+rescue => e
+  puts "Error: #{e.message}"
+end
+
+# 6. Text-to-image (Kolors)
+puts "\n=== Example 6: 文生图 ==="
+begin
+  result = engine.call_worker(:siliconflow_image, {
+    prompt: "一只在书房里读书的猫，水墨画风格",
+    image_size: "1024x1024",
+    save_to_file: true,
+    output_dir: "./generated_images",
+    filename_prefix: "siliconflow_cat",
+  })
+  if result.is_a?(Hash) && result[:images]
+    puts "Generated #{result[:images].size} image(s); first URL: #{result[:images].first[:url]}"
+    puts "Saved files: #{result[:saved_files]}"
+  else
+    puts "Result: #{result}"
+  end
+rescue => e
+  puts "Error: #{e.message}"
+end
+
+# 7. Text-to-video (Wan2.2, async) — may take a couple of minutes.
+puts "\n=== Example 7: 文生视频 (async) ==="
+begin
+  result = engine.call_worker(:siliconflow_video, {
+    prompt: "一只猫在阳光下打盹",
+    wait_for_completion: true,
+    download_to_file: true,
+    output_dir: "./generated_videos",
+    timeout: 600,
+  })
+  if result[:video]
+    puts "Video ready: #{result[:video][:video_url]}"
+    puts "Downloaded: #{result[:downloaded_file]}" if result[:downloaded_file]
+  else
+    puts "Submitted request: #{result[:submitted]}"
+  end
+rescue => e
+  puts "Error: #{e.message}"
+end
+
+# 8. TTS (CosyVoice2)
+puts "\n=== Example 8: 语音合成 (TTS) ==="
+begin
+  info = engine.call_worker(:siliconflow_tts, {
+    text: "你好，这是硅基流动语音合成的测试。",
+    voice: "FunAudioLLM/CosyVoice2-0.5B:alex",
+    output_path: "./generated_audio/siliconflow_tts.mp3",
+  })
+  puts "Audio saved: #{info[:file_path]}"
+rescue => e
+  puts "Error: #{e.message}"
+end
+
+# 9. ASR (SenseVoiceSmall) — needs a real audio file path.
+puts "\n=== Example 9: 语音识别 (ASR) ==="
+audio = ENV["SILICONFLOW_ASR_SAMPLE"] || "./generated_audio/siliconflow_tts.mp3"
+if File.exist?(audio)
+  begin
+    result = engine.call_worker(:siliconflow_asr, { audio_file: audio })
+    puts "Transcription: #{result[:text]}"
+  rescue => e
+    puts "Error: #{e.message}"
+  end
+else
+  puts "Skipped: set SILICONFLOW_ASR_SAMPLE to an audio file path (or run TTS first) to test ASR."
+end
+
+puts "\n=== All examples completed ==="
+
+File.delete("siliconflow_config.yml") if File.exist?("siliconflow_config.yml")

data/lib/smart_prompt/adapters/siliconflow/embed.rb

@@ -0,0 +1,33 @@
+module SmartPrompt
+  module SiliconFlow
+    # Embeddings (BAAI/bge-m3 default; Qwen3-Embedding supports custom dimensions).
+    module Embed
+      # BAAI/bge-m3 (default, fixed 1024 dims) or Qwen3-Embedding (custom dimensions).
+      # `dimensions` is only honored when set in config (Qwen3-Embedding series only).
+      # Returns the first embedding vector.
+      def embeddings(text, model)
+        model_name = model || @config["embedding_model"] || @config["model"]
+        SmartPrompt.logger.info "SiliconFlowAdapter: embeddings model=#{model_name}"
+
+        body = { "model" => model_name, "input" => text.is_a?(Array) ? text : [text.to_s] }
+        body["dimensions"] = @config["dimensions"] if @config["dimensions"]
+        body["encoding_format"] = @config["encoding_format"] if @config["encoding_format"]
+
+        response =
+          begin
+            http_post_json("#{@base_url}/embeddings", body)
+          rescue LLMAPIError, Error
+            raise
+          rescue => e
+            raise LLMAPIError, "Failed to call SiliconFlow embeddings: #{e.message}"
+          end
+
+        items = response["data"]
+        unless items.is_a?(Array) && items.any? && items[0]["embedding"]
+          raise LLMAPIError, "No embedding vector in SiliconFlow response: #{response.inspect}"
+        end
+        items[0]["embedding"]
+      end
+    end
+  end
+end

data/lib/smart_prompt/adapters/siliconflow/image.rb

@@ -0,0 +1,103 @@
+module SmartPrompt
+  module SiliconFlow
+    # Text-to-image (Kolors) + image editing (Qwen-Image-Edit). save_image comes from
+    # the ImagePersistence concern.
+    module Image
+      # Default resolution for text-to-image (Kolors accepts these "WxH" values).
+      DEFAULT_IMAGE_SIZE = "1024x1024".freeze
+
+      # Text-to-image. SiliconFlow response is images[].url (not OpenAI's data[]),
+      # and uses its own param names (image_size, batch_size, guidance_scale, ...).
+      # Returns an Array of {url:, b64_json:, seed:}.
+      def generate_image(prompt, params = {})
+        SmartPrompt.logger.info "SiliconFlowAdapter: generating image"
+        raise Error, "Prompt cannot be empty" if prompt.nil? || prompt.to_s.strip.empty?
+
+        model_name = params[:model] || @config["image_model"] || @config["model"]
+        raise Error, "No model configured for image generation" if model_name.nil? || model_name.to_s.strip.empty?
+
+        body = { "model" => model_name, "prompt" => prompt.to_s }
+        body["image_size"]          = resolve_image_size(params[:image_size] || params[:size])
+        body["negative_prompt"]     = params[:negative_prompt]     if params[:negative_prompt]
+        body["seed"]                = params[:seed]                if params[:seed]
+        body["num_inference_steps"] = params[:num_inference_steps] if params[:num_inference_steps]
+        body["guidance_scale"]      = params[:guidance_scale]      if params[:guidance_scale]
+        # batch_size only applies to a subset of models (e.g. Kolors); send it only
+        # when the caller explicitly asks for it.
+        batch = params[:batch_size] || params[:n]
+        body["batch_size"] = batch if batch
+
+        SmartPrompt.logger.info "SiliconFlow image params: #{body.except('prompt').inspect}"
+        response =
+          begin
+            http_post_json(@image_url, body)
+          rescue LLMAPIError, Error
+            raise
+          rescue => e
+            raise Error, "Failed to call SiliconFlow image generation: #{e.message}"
+          end
+
+        images = parse_image_response(response)
+        SmartPrompt.logger.info "SiliconFlowAdapter: generated #{images.size} image(s)"
+        images
+      end
+
+      # Image editing / image-to-image (Qwen/Qwen-Image-Edit-2509 and Kolors composable).
+      # +image+ (and optionally +image2+/+image3+) may be a local file path, a base64
+      # data URL, or a public http(s) URL. Edit models reject image_size, so we omit it.
+      def edit_image(prompt, params = {})
+        SmartPrompt.logger.info "SiliconFlowAdapter: editing image"
+        raise Error, "Prompt cannot be empty" if prompt.nil? || prompt.to_s.strip.empty?
+        raise Error, "An input image is required for image editing" if params[:image].nil? && params[:image_file].nil?
+
+        model_name = params[:model] || @config["image_model"] || @config["model"]
+        raise Error, "No model configured for image generation" if model_name.nil? || model_name.to_s.strip.empty?
+
+        body = { "model" => model_name, "prompt" => prompt.to_s }
+        body["image"]          = normalize_input_image(params[:image] || params[:image_file])
+        body["image2"]         = normalize_input_image(params[:image2]) if params[:image2]
+        body["image3"]         = normalize_input_image(params[:image3]) if params[:image3]
+        body["negative_prompt"] = params[:negative_prompt] if params[:negative_prompt]
+        body["seed"]           = params[:seed]            if params[:seed]
+        body["guidance_scale"] = params[:guidance_scale]  if params[:guidance_scale]
+
+        SmartPrompt.logger.info "SiliconFlow image edit params: #{body.except('prompt', 'image', 'image2', 'image3').inspect}"
+        response =
+          begin
+            http_post_json(@image_url, body)
+          rescue LLMAPIError, Error
+            raise
+          rescue => e
+            raise Error, "Failed to call SiliconFlow image edit: #{e.message}"
+          end
+
+        images = parse_image_response(response)
+        SmartPrompt.logger.info "SiliconFlowAdapter: edited into #{images.size} image(s)"
+        images
+      end
+
+      private
+
+      # SiliconFlow image response: images[].url (FLAT). Fall back to OpenAI's
+      # data[] or a bare-url array for defensive compatibility.
+      def parse_image_response(response)
+        items = response["images"] || response["data"]
+        items = [] unless items.is_a?(Array)
+        if items.empty?
+          SmartPrompt.logger.error "No image data in SiliconFlow response: #{response.inspect}"
+          raise LLMAPIError, "No image data in SiliconFlow response"
+        end
+        items.map do |d|
+          d = { "url" => d } if d.is_a?(String)
+          { url: d["url"], b64_json: d["b64_json"], seed: d["seed"] }
+        end
+      end
+
+      # Resolve the image size: default 1024x1024 when none given.
+      def resolve_image_size(size)
+        return DEFAULT_IMAGE_SIZE if size.nil? || size.to_s.strip.empty?
+        size.to_s
+      end
+    end
+  end
+end

data/lib/smart_prompt/adapters/siliconflow/rerank.rb

@@ -0,0 +1,41 @@
+module SmartPrompt
+  module SiliconFlow
+    # Rerank (reorder documents by relevance to a query).
+    module Rerank
+      # Reorder documents by relevance to a query. SiliconFlow returns
+      # results[].relevance_score (NOT "score"). Returns an Array of
+      # {index:, relevance_score:} sorted by the provider.
+      def rerank(query, documents, model: nil, **opts)
+        model_name = model || @config["rerank_model"] || @config["model"]
+        SmartPrompt.logger.info "SiliconFlowAdapter: rerank model=#{model_name}"
+
+        body = { "model" => model_name, "query" => query.to_s, "documents" => documents }
+        body["top_n"]                = opts[:top_n]                if opts[:top_n]
+        body["return_documents"]     = opts[:return_documents]     unless opts[:return_documents].nil?
+        body["max_chunks_per_doc"]   = opts[:max_chunks_per_doc]   if opts[:max_chunks_per_doc]
+        body["chunk_overlap_tokens"] = opts[:chunk_overlap_tokens] if opts[:chunk_overlap_tokens]
+        body["instruction"]          = opts[:instruction]          if opts[:instruction]
+
+        response =
+          begin
+            http_post_json("#{@base_url}/rerank", body)
+          rescue LLMAPIError, Error
+            raise
+          rescue => e
+            raise LLMAPIError, "Failed to call SiliconFlow rerank: #{e.message}"
+          end
+
+        parse_rerank_response(response)
+      end
+
+      private
+
+      # SiliconFlow rerank response: results[].relevance_score (NOT "score").
+      def parse_rerank_response(response)
+        (response["results"] || []).map do |r|
+          { index: r["index"], relevance_score: r["relevance_score"] || r["score"] }
+        end
+      end
+    end
+  end
+end

data/lib/smart_prompt/adapters/siliconflow/text.rb

@@ -0,0 +1,54 @@
+module SmartPrompt
+  module SiliconFlow
+    # Text chat + multimodal vision (OpenAI-compatible /chat/completions, SSE streaming,
+    # reasoning_content passthrough).
+    module Text
+      CHAT_OPTIONAL_KEYS = %w[
+        top_p top_k frequency_penalty presence_penalty
+        max_tokens max_completion_tokens stop response_format
+        enable_thinking thinking_budget min_p reasoning_effort seed
+      ].freeze
+
+      # Chat / multimodal. Non-streaming returns a full OpenAI-format hash (so
+      # last_response carries usage + reasoning_content); streaming calls +proc+
+      # with each OpenAI-shaped chunk and returns nil.
+      def send_request(messages, model = nil, temperature = nil, tools = nil, proc = nil)
+        model_name = model || @config["model"]
+        body = build_chat_body(messages, model_name, temperature, tools)
+        SmartPrompt.logger.info "SiliconFlowAdapter: chat request model=#{model_name} stream=#{!proc.nil?}"
+
+        url = "#{@base_url}/chat/completions"
+        if proc
+          body["stream"] = true
+          stream_chat(url, body) { |data| proc.call(build_stream_chunk(data), 0) }
+          SmartPrompt.logger.info "SiliconFlowAdapter: streaming request finished"
+          nil
+        else
+          raw = http_post_json(url, body)
+          response = build_completion_response(raw)
+          @last_response = response
+          SmartPrompt.logger.info "SiliconFlowAdapter: received chat response"
+          response
+        end
+      rescue LLMAPIError, Error
+        raise
+      rescue => e
+        SmartPrompt.logger.error "SiliconFlow chat error: #{e.message}"
+        raise LLMAPIError, "Failed to call SiliconFlow chat: #{e.message}"
+      end
+
+      private
+
+      def build_chat_body(messages, model_name, temperature, tools)
+        body = {
+          "model"       => model_name,
+          "messages"    => process_multimodal_messages(messages),
+          "temperature" => @config["temperature"] || temperature || 0.7,
+        }
+        CHAT_OPTIONAL_KEYS.each { |k| body[k] = @config[k] if @config.key?(k) }
+        body["tools"] = tools if tools && !tools.empty?
+        body
+      end
+    end
+  end
+end

data/lib/smart_prompt/adapters/siliconflow/video.rb

@@ -0,0 +1,111 @@
+module SmartPrompt
+  module SiliconFlow
+    # Text-to-video / image-to-video (Wan2.2, async submit -> poll -> download).
+    module Video
+      # Text-to-video image_size enum (SiliconFlow rejects anything else).
+      VALID_VIDEO_SIZES = %w[1280x720 720x1280 960x960].freeze
+      DEFAULT_VIDEO_SIZE = "1280x720".freeze
+
+      # Submit a text-to-video (or image-to-video) job. Returns the requestId.
+      # SiliconFlow's submit endpoint returns {"requestId": "..."} (camelCase).
+      def generate_video(prompt, params = {})
+        SmartPrompt.logger.info "SiliconFlowAdapter: submitting video job"
+        model_name = params[:model] || @config["video_model"] || @config["model"]
+        raise Error, "No model configured for video generation" if model_name.nil? || model_name.to_s.strip.empty?
+
+        body = { "model" => model_name, "prompt" => prompt.to_s }
+        body["image_size"]      = resolve_video_size(params[:image_size] || params[:size])
+        body["negative_prompt"] = params[:negative_prompt] if params[:negative_prompt]
+        body["seed"]            = params[:seed]            if params[:seed]
+        body["image"]           = normalize_input_image(params[:image]) if params[:image]
+
+        SmartPrompt.logger.info "SiliconFlow video params: #{body.except('prompt').inspect}"
+        response =
+          begin
+            http_post_json(@video_submit_url, body)
+          rescue LLMAPIError, Error
+            raise
+          rescue => e
+            raise Error, "Failed to submit SiliconFlow video job: #{e.message}"
+          end
+
+        request_id = response["requestId"] || response["request_id"]
+        raise LLMAPIError, "No requestId in SiliconFlow video response: #{response.inspect}" unless request_id
+        SmartPrompt.logger.info "SiliconFlowAdapter: video request #{request_id} submitted"
+        { request_id: request_id, model: model_name, raw: response }
+      end
+
+      # Poll an async task. SiliconFlow's status endpoint is a POST (NOT GET) that
+      # takes {requestId} in the body. Returns the raw status hash.
+      def check_video_status(request_id)
+        SmartPrompt.logger.info "SiliconFlowAdapter: polling video request #{request_id}"
+        http_post_json(@video_status_url, { "requestId" => request_id })
+      rescue LLMAPIError, Error
+        raise
+      rescue => e
+        raise LLMAPIError, "Failed to query SiliconFlow video status: #{e.message}"
+      end
+
+      # Block until the task finishes (or times out), then return the video URL.
+      # SiliconFlow status values are exactly: Succeeded / InQueue / InProgress / Failed.
+      def wait_for_video_completion(request_id, check_interval: 10, timeout: 600)
+        start = Time.now
+        loop do
+          status = check_video_status(request_id)
+          case video_status_of(status)
+          when "Succeed"
+            url = video_url_of(status)
+            raise LLMAPIError, "Video succeeded but no url in: #{status.inspect}" unless url
+            SmartPrompt.logger.info "SiliconFlowAdapter: video ready #{url}"
+            return { request_id: request_id, status: "Succeeded", video_url: url, raw: status }
+          when "Failed"
+            raise LLMAPIError, "SiliconFlow video generation failed: #{status["reason"] || status.inspect}"
+          else
+            if Time.now - start > timeout
+              raise LLMAPIError, "SiliconFlow video generation timeout after #{timeout}s"
+            end
+            SmartPrompt.logger.info "SiliconFlow video request #{request_id} #{video_status_of(status)}..."
+            sleep(check_interval)
+          end
+        end
+      end
+
+      def download_video(video_url, output_path)
+        uri = URI.parse(video_url)
+        http = Net::HTTP.new(uri.host, uri.port); http.use_ssl = (uri.scheme == "https")
+        response = http.request(Net::HTTP::Get.new(uri.request_uri))
+        raise Error, "Failed to download video: #{response.code}" unless response.is_a?(Net::HTTPSuccess)
+        FileUtils.mkdir_p(File.dirname(output_path))
+        File.binwrite(output_path, response.body)
+        SmartPrompt.logger.info "SiliconFlow video saved to #{output_path}"
+        output_path
+      rescue => e
+        raise e.is_a?(SmartPrompt::Error) ? e : Error, "Error downloading SiliconFlow video: #{e.message}"
+      end
+
+      private
+
+      # SiliconFlow video status is under the top-level `status` field.
+      def video_status_of(status)
+        status["status"] || "InQueue"
+      end
+
+      # The video url lives at results.videos[].url (results is an OBJECT, not array).
+      def video_url_of(status)
+        videos = status.dig("results", "videos")
+        item = videos.is_a?(Array) ? videos[0] : videos
+        item.is_a?(Hash) ? (item["url"] || item["video_url"]) : nil
+      end
+
+      # Resolve the video image_size: default 1280x720; warn on unknown values.
+      def resolve_video_size(size)
+        size = size.nil? || size.to_s.strip.empty? ? DEFAULT_VIDEO_SIZE : size.to_s
+        unless VALID_VIDEO_SIZES.include?(size)
+          SmartPrompt.logger.warn "SiliconFlow video image_size '#{size}' is not in the known-valid list " \
+                                  "(#{VALID_VIDEO_SIZES.join(', ')}); the API may reject it."
+        end
+        size
+      end
+    end
+  end
+end