smart_prompt 0.5.0 → 0.5.2

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 6417bfd0f16178f4e5b2d26b9c14f3ce8a22e8d9ab77f3cacf1845fa6531d6d0
-  data.tar.gz: 9eae765158a312e1605aad80917d59180109bb0a8c2603a2e2b7b95dd7284b00
+  metadata.gz: 1477a83b116691863533a9b1726e40be07b03b4faa67b60ebc72fe6a290d60f1
+  data.tar.gz: c9e71e998318d186f296495679573ccb2ad9b539420b3b9c3ee02314db8e2d8b
 SHA512:
-  metadata.gz: 4109ce9c8131961870d2c9b83d3d2ed534ebdf2ba42b34674516b61058034bab9af6c53079354d39a154759ca02bb39751eeb2c055465f2dfb7db8f2a3a12682
-  data.tar.gz: e12c080c4d95a7196e91497b02ef493a11b45e3da7ba6cdb5e342c00c40abb686fa1aac3e6c0db9e3905f5f18adab0395559188fe0fae5837c1128dad70f236b
+  metadata.gz: a433f3724915b38af6e3a1e66d3f52568ec305f85d3b038643ffeb1ae5522547a8e5c6769e6c53cf1a081ca353587ddd0156987a4d2c705f25f267e388a7f5b9
+  data.tar.gz: 90b3ba6033912705cc096b17765f0ad221d6f8a6e6b3b2a4350345d4b0777ebdee359d30604320918891ee7965cb55528a0aab4c12b9d9eebd67237ff52d7876

data/CHANGELOG.md

@@ -5,7 +5,14 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
-## [0.4.0] - 2026-06-21
+## [Unreleased]
+### Added
+- **硅基流动 (SiliconFlow) support** — unified `SiliconFlowAdapter` covering all REST categories: chat, multimodal vision (image/video/audio), embeddings (BAAI/bge-m3), rerank (bge-reranker-v2-m3), text-to-image (Kolors), image-edit (Qwen-Image-Edit), text-to-video (Wan2.2 async submit/poll), TTS (CosyVoice2), ASR (SenseVoiceSmall), and custom-voice management — with SSE streaming, reasoning_content passthrough, and the provider-correct async video flow (`POST /video/submit` → `requestId`, `POST /video/status` → `results.videos[].url`)
+
+### Changed
+- **Refactored the Zhipu / SenseNova / SiliconFlow adapters** — extracted byte-identical cross-provider logic into four shared concerns under `lib/smart_prompt/concerns/` (`HTTPClient`, `MultimodalMessages`, `OpenAIChatShaping`, `ImagePersistence`), and split the Zhipu and SiliconFlow adapters into per-modality capability modules under `lib/smart_prompt/adapters/<provider>/` (`Text` / `Embed` / `Image` / `Video` / `Voice` / `Rerank`). Pure internal refactor — no public-API change (`send_request` stays 5-arg, all DSL-delegated method names preserved), behavior unchanged. ~286 lines removed and the previously triplicated HTTP / multimodal / chat-shaping / image-persistence code now has a single source.
+
+## [0.5.1] - 2026-06-21
 ### Added
 - **SenseNova (商汤日日新) support** — unified `SenseNovaAdapter` covering chat (商量), multimodal vision, Cupido embeddings, and 秒画 text-to-image, with SSE streaming and reasoning-field handling
 - **智谱 AI (BigModel / GLM) support** — unified `ZhipuAIAdapter` covering all REST categories: chat (GLM-4), vision (GLM-4V), embeddings (embedding-3), text-to-image (CogView), text-to-video (CogVideoX async), TTS (GLM-TTS), ASR (GLM-ASR), and rerank
@@ -13,9 +20,25 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Intelligent conversation history management (sliding-window, relevance-based, summary-based, hybrid strategies) with session isolation, compression, persistence, and LRU caching
 - Token counter, message/session models, and persistence layer
 - Example configs, workers, and self-contained examples for every provider
+- Integrated upstream gemma4 multimodal support (`use_model`, `thinking`, `image`/`audio`/`video`, `multimodal_prompt`) and `request_options` plumbing
 
 ### Fixed
 - Expose `engine` on `WorkerContext` so workers can reach a configured adapter directly (fixes the `engine.llms[...]` pattern used by media workers)
+- `Worker#execute` default session_id was hard-coded to `"default"`, leaving the per-worker session branch as dead code and collapsing all history-using workers onto one shared session; now generates `worker_<name>_<ts>`
+- `AnthropicAdapter`: add `extract_content_from_response` and stop double-wrapping multimodal (array) content
+- file_upload multimodal fix: base64-encode local image/audio/video files instead of passing raw paths
+
+## [0.4.1] - 2026-04-22
+### Fixed
+- Re-release package with `lib/smart_prompt/anthropic_adapter.rb`, which is required by the gem entrypoint.
+
+## [0.4.0] - 2026-04-22
+### Added
+- Anthropic adapter support.
+
+## [0.3.6] - 2026-04-08
+### Changed
+- Bumped `ruby-openai` dependency from `8.1.0` to `8.3.0`
 
 ## [0.3.2] - 2025-05-18
 ### Added
@@ -46,4 +69,4 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Initial gem release
 - Llama.cpp adapter
 - Basic configuration parameters
-- Environment bug fixes
+- Environment bug fixes

data/README.cn.md

@@ -14,6 +14,7 @@ SmartPrompt 是一个强大的 Ruby gem，提供了优雅的领域特定语言
 - **Anthropic Claude**: 原生支持 Claude 模型及多模态能力
 - **商汤 SenseNova（日日新）**: 单一适配器覆盖商量文本对话、图文多模态、Cupido 向量、秒画文生图四类 API，详见 `examples/sensenova_example.rb`
 - **智谱 AI（BigModel / GLM）**: 单一适配器覆盖全部模型类别——文本对话（GLM-4）、图文多模态（GLM-4V）、向量（embedding-3）、文生图（CogView）、文生视频（CogVideoX）、语音合成（GLM-TTS）、语音识别（GLM-ASR），详见 `examples/zhipu_example.rb`
+- **硅基流动（SiliconFlow）**: 单一适配器覆盖全部模型类别——文本对话、图文多模态、向量（bge-m3）、重排（rerank）、文生图（Kolors）、图像编辑（Qwen-Image-Edit）、文生视频（Wan2.2 异步）、语音合成（CosyVoice2）、语音识别（SenseVoiceSmall）及自定义音色管理，详见 `examples/siliconflow_example.rb`
 - **Llama.cpp 集成**: 直接集成本地 Llama.cpp 服务器
 - **可扩展适配器**: 易于扩展的适配器系统，支持新的 LLM 提供商
 - **统一接口**: 无论底层 LLM 提供商如何，都使用相同的 API
@@ -95,12 +96,33 @@ llms:
     adapter: openai
     url: http://localhost:11434/
     default_model: deepseek-r1
+  gemma4_local:
+    adapter: openai
+    url: http://localhost:8000/v1
+    api_key: dummy
+    default_model: gemma-4-12B-it
+    temperature: 1.0
+    top_p: 0.95
+    top_k: 64
   deepseek:
     adapter: openai
     url: https://api.deepseek.com
     api_key: ENV["DSKEY"]
     default_model: deepseek-reasoner
 
+# 模型别名配置
+models:
+  local/qwen3.5:
+    use: local
+    model: qwen3.5
+  deepseekv3.2:
+    use: SiliconFlow
+    model: Pro/deepseek-ai/DeepSeek-V3.2
+  gemma4/12b:
+    use: gemma4_local
+    model: gemma-4-12B-it
+    max_tokens: 1024
+
 # 默认设置
 default_llm: SiliconFlow
 template_path: "./templates"
@@ -128,9 +150,8 @@ logger_file: "./logs/smart_prompt.log"
 **workers/chat_worker.rb**:
 ```ruby
 SmartPrompt.define_worker :chat_assistant do
-  # 使用特定的 LLM
-  use "SiliconFlow"
-  model "deepseek-ai/DeepSeek-V3"
+  # 使用配置好的模型别名
+  use_model "deepseekv3.2"
   # 设置系统消息
   sys_msg("你是一个有用的 AI 助手。", params)
   # 使用模板和参数  
@@ -182,6 +203,26 @@ engine.call_worker_by_stream(:streaming_chat, {
 end
 ```
 
+### Gemma 4 12B 多模态
+
+Gemma 4 12B 可以通过 LiteRT-LM、LM Studio、Ollama、llama.cpp 等 OpenAI 兼容本地服务接入。SmartPrompt 会把图片放在文本前、音频放在文本后，以匹配 Gemma 4 的多模态最佳实践。
+
+```ruby
+SmartPrompt.define_worker :gemma_multimodal_assistant do
+  use_model "gemma4/12b"
+  thinking params.fetch(:thinking, true)
+  sys_msg("你是一个严谨的本地多模态助手。", params)
+
+  image(params[:image], token_budget: params[:token_budget] || 280) if params[:image]
+  video(params[:video], fps: 1, max_seconds: 60) if params[:video]
+  audio(params[:audio]) if params[:audio]
+  prompt(params[:message])
+
+  request_options(response_format: { type: "json_object" }) if params[:json]
+  send_msg
+end
+```
+
 ### 工具集成
 
 ```ruby
@@ -561,6 +602,17 @@ llms:
     model: "FunAudioLLM/CosyVoice2-0.5B"
 ```
 
+### 模型别名配置
+
+```yaml
+models:
+  model_alias:
+    use: "llm_name"
+    model: "model_identifier"
+```
+
+在 worker 中，`use_model "model_alias"` 等价于调用 `use "llm_name"` 和 `model "model_identifier"`。
+
 ### 路径配置
 
 ```yaml
@@ -685,4 +737,4 @@ end
 
 ---
 
-**SmartPrompt** - 让 Ruby 应用中的 LLM 集成变得简单、强大且优雅。
+**SmartPrompt** - 让 Ruby 应用中的 LLM 集成变得简单、强大且优雅。

data/README.md

@@ -14,6 +14,7 @@ SmartPrompt is a powerful Ruby gem that provides an elegant domain-specific lang
 - **Anthropic Claude**: Native support for Claude models with multimodal capabilities
 - **SenseNova (商汤日日新)**: One adapter covers chat (商量), multimodal vision (图文多模态), Cupido embeddings (向量), and 秒画 text-to-image — see `examples/sensenova_example.rb`
 - **智谱 AI (BigModel / GLM)**: One adapter covers all categories — chat (GLM-4), vision (GLM-4V), embeddings (embedding-3), text-to-image (CogView), text-to-video (CogVideoX), TTS (GLM-TTS), ASR (GLM-ASR) — see `examples/zhipu_example.rb`
+- **硅基流动 (SiliconFlow)**: One adapter covers all categories — chat, multimodal vision, embeddings, rerank, text-to-image (Kolors), image-edit (Qwen-Image-Edit), text-to-video (Wan2.2 async), TTS (CosyVoice2), ASR (SenseVoiceSmall) — see `examples/siliconflow_example.rb`
 - **Llama.cpp Integration**: Direct integration with local Llama.cpp servers
 - **Extensible Adapters**: Easy-to-extend adapter system for new LLM providers
 - **Unified Interface**: Same API regardless of the underlying LLM provider
@@ -95,12 +96,33 @@ llms:
     adapter: openai
     url: http://localhost:11434/
     default_model: deepseek-r1
+  gemma4_local:
+    adapter: openai
+    url: http://localhost:8000/v1
+    api_key: dummy
+    default_model: gemma-4-12B-it
+    temperature: 1.0
+    top_p: 0.95
+    top_k: 64
   deepseek:
     adapter: openai
     url: https://api.deepseek.com
     api_key: ENV["DSKEY"]
     default_model: deepseek-reasoner
 
+# Model aliases
+models:
+  local/qwen3.5:
+    use: local
+    model: qwen3.5
+  deepseekv3.2:
+    use: SiliconFlow
+    model: Pro/deepseek-ai/DeepSeek-V3.2
+  gemma4/12b:
+    use: gemma4_local
+    model: gemma-4-12B-it
+    max_tokens: 1024
+
 # Default settings
 default_llm: SiliconFlow
 template_path: "./templates"
@@ -128,9 +150,8 @@ Create worker files in your `workers/` directory:
 **workers/chat_worker.rb**:
 ```ruby
 SmartPrompt.define_worker :chat_assistant do
-  # Use a specific LLM
-  use "SiliconFlow"
-  model "deepseek-ai/DeepSeek-V3"
+  # Use a configured model alias
+  use_model "deepseekv3.2"
   # Set system message
   sys_msg("You are a helpful AI assistant.", params)
   # Use template with parameters  
@@ -182,6 +203,26 @@ engine.call_worker_by_stream(:streaming_chat, {
 end
 ```
 
+### Gemma 4 12B Multimodal
+
+Gemma 4 12B can be connected through OpenAI-compatible local servers such as LiteRT-LM, LM Studio, Ollama, or llama.cpp. SmartPrompt places images before text and audio after text to match Gemma 4 multimodal best practices.
+
+```ruby
+SmartPrompt.define_worker :gemma_multimodal_assistant do
+  use_model "gemma4/12b"
+  thinking params.fetch(:thinking, true)
+  sys_msg("You are a precise local multimodal assistant.", params)
+
+  image(params[:image], token_budget: params[:token_budget] || 280) if params[:image]
+  video(params[:video], fps: 1, max_seconds: 60) if params[:video]
+  audio(params[:audio]) if params[:audio]
+  prompt(params[:message])
+
+  request_options(response_format: { type: "json_object" }) if params[:json]
+  send_msg
+end
+```
+
 ### Tool Integration
 
 ```ruby
@@ -565,6 +606,17 @@ llms:
     model: "FunAudioLLM/CosyVoice2-0.5B"
 ```
 
+### Model Alias Configuration
+
+```yaml
+models:
+  model_alias:
+    use: "llm_name"
+    model: "model_identifier"
+```
+
+In a worker, `use_model "model_alias"` is equivalent to calling `use "llm_name"` and `model "model_identifier"`.
+
 ### Path Configuration
 
 ```yaml
@@ -689,4 +741,4 @@ This project is licensed under the MIT License - see the [LICENSE.txt](LICENSE.t
 
 ---
 
-**SmartPrompt** - Making LLM integration in Ruby applications simple, powerful, and elegant.
+**SmartPrompt** - Making LLM integration in Ruby applications simple, powerful, and elegant.