feat: OpenRouter OneShot, Quick Chat, dimension/rerank/transcription fixes

- OpenRouter OneShot dialog with model classification (tool-calling, vision, embedding)
- OpenRouter transcription adapter (base64 JSON, no multipart)
- Quick Chat side panel with Direct RAG mode, session per collection
- Embedding dimension follows collection's actual Qdrant vector size
- Collection-aware embedding overrides in DirectQueryModule and recall
- Reranker fixes: base_url in OneShot, skip Qwen chat fallback for non-Qwen models
- Transcription preload: skip when external provider active, skip when already loading
- Collection creation: child_chunk_size default 256, ADVANCED collapsible section
- Recall filesMap: resolve source→display_name for search results and eval cases
- Cleanup stale selectedCollections IDs on fetch
- Various settings page improvements (auto-save, one-per-row layout, model folding)
- favicon.svg replaced

Author: superdd-coder

.env.template

@@ -0,0 +1,15 @@
+# SinkDuce environment variables
+# Copy this file to .env and modify as needed:
+#   cp .env.template .env
+
+# Backend API port (default: 18900)
+API_PORT=18900
+
+# Frontend dev server port (default: 5173)
+UI_PORT=5173
+
+# Qdrant vector database HTTP port (default: 6343)
+QDRANT_HTTP_PORT=6343
+
+# Qdrant vector database gRPC port (default: 6334)
+QDRANT_GRPC_PORT=6334

README.md

@@ -40,15 +40,25 @@ docker compose up -d --build
 
 Docker rebuilds the image with the latest code while preserving your `data/` directory (database, config, history).
 
-### Recommended Out-of-the-Box Setup (DashScope)
+### Recommended Out-of-the-Box Setup
 
-The project is highly optimized for [Alibaba Cloud DashScope (Bailian/阿里云百炼)](https://bailian.console.aliyun.com/). Go to **Settings** → **LLM Providers** → click **OneShot Setting (DashScope API)**, enter your API Key, and all services (LLM, Embedding, Reranker, STT) will automatically auto-configure with optimal defaults:
+Two OneShot configuration paths are available:
+
+**DashScope** (Alibaba Cloud): Go to **Settings** → **LLM Providers** → **OneShot Setting (DashScope API)**:
 
 * **LLM**: `deepseek-v4-flash`
-* **Embedding**: `text-embedding-v4`
+* **Embedding**: `text-embedding-v4` (1024d)
 * **Reranker**: `qwen3-rerank`
 * **Transcription**: `fun-asr` / `fun-asr-realtime`
 
+**OpenRouter**: Go to **Settings** → **LLM Providers** → **OneShot Setting (OpenRouter API)**. Enter your API Key; models are auto-fetched and classified:
+
+* **LLM**: `deepseek/deepseek-v4-flash`
+* **Chat** (function-calling, tools-filtered): `deepseek/deepseek-v4-pro`
+* **Visual** (vision models): `xiaomi/mimo-v2.5`
+* **Embedding**: `qwen/qwen3-embedding-4b` (default 1536d)
+* **Reranker**: `cohere/rerank-v3.5`
+
 ## 🏗️ Core Pillars
 
 ### ⚡ 01. Spark: Fluid Friction Capture & Synthesis
@@ -58,8 +68,10 @@ The project is highly optimized for [Alibaba Cloud DashScope (Bailian/阿里云
 * **Full-Featured Markdown Workspace (Collection Notes)**: Create structured personal notes explicitly bound to specific operational contexts, featuring full WYSIWYG editing with markdown support (headings, tables, task lists, code blocks, images, YouTube embeds).
 * **Intelligent Note Distillation (Drag-to-Distill)**: Drag an existing old note or document into your current workspace, and the system automatically extracts the core insights into a dense citation block, seamlessly aggregating scattered ideas without manual rewriting.
 * **AI Image Ingestion**: One click prompts the AI to generate a precise contextual text description, weaving visual data into your markdown memory map for vector indexing. Supports image pasting, drag-and-drop, and resizing with inline captions.
-* **Audio Scribing & Tri-Fold Pipeline**: Upload recordings or record live meetings/lectures via WebSockets. Transcription (STT) can be handled via the embedded local engine or flexible external API endpoints. The system then generates a structured tri-fold artifact: *Summary* (semantic synthesis), *To-Do List/Action Items*, and *Detail* (deep-dive information extraction that filters conversational noise while preserving core intent).
+* **Audio Transcription & Multi-Phase Pipeline**: Upload recordings or record live meetings/lectures via WebSockets. Transcription (STT) can be handled by the embedded local FunASR engine or external APIs (DashScope, OpenRouter, etc.). After transcription, a multi-phase analysis runs: **Pass 1 — General Summary + Blueprint** (SSE streaming, builds section structure); **Pass 2 — Section Summary** (one-click custom sections, one keyword generates description and summary); **Pass 3 — Final Merge** producing Summary, To-Do List, and Detail artifacts.
 * **Hot Words**: Manage custom vocabulary libraries for domain-specific terminology to improve transcription accuracy.
+* **Quick Chat Panel**: A floating side panel in the Collection view for rapid Q&A via Direct RAG. Diamond button with animation, per-collection sessions, SSE streaming with thinking step display.
+* **Multi-turn Agentic Chat**: The main Chat supports multi-turn conversation with tool-calling Agent that autonomously decides whether, where, and how to search — exposing the full Agentic RAG capability through conversational interaction.
 
 ### 📥 02. Sink: Anti-Hoarding Ingestion Pipeline
 
@@ -70,13 +82,14 @@ The project is highly optimized for [Alibaba Cloud DashScope (Bailian/阿里云
 * **Granular Document Parsing & Chunking**: Utilizes the natively embedded parsing engine (supporting 12 format parsers) or links to powerful cloud parsing APIs (e.g., *MinerU*). Supports advanced **Parent-Child chunking** by headings, paragraphs, and max token configurations while keeping word boundaries intact.
 * **Context Enrichment Engine**: When enabled, an LLM evaluates each split chunk and injects its missing global context, mitigating the "chunk isolation" effect during retrieval.
 * **Auto-Summarization & Consolidation**: Every document gets a structured summary. Collections get consolidated overviews with automatic conflict detection when documents contradict each other.
+* **Auto-Updating Collection Coverage**: Collection-level coverage maps update automatically after each ingestion, comparing new and existing documents to keep retrieval context aligned with the latest knowledge boundaries.
 
 ### 🧠 03. Educe: High-Dimensional Contextual Reasoning
 
 **Educe** implements a cutting-edge retrieval architecture to turn cold data into active intelligence.
 
 * **Advanced Hybrid Retrieval**: Supports dense vector similarity search, keyword-semantic hybrid querying (BM25 + Dense via Qdrant), and advanced **Reranking** algorithms to surface top-tier context.
-* **Iterative Agentic RAG Pipeline**: Moves far beyond naïve single-shot semantic search. Sinkduce orchestrates a multi-step Agentic loop: *Analyze → Route → Retrieve → Grade → Decompose → Rerank → Synthesize* until it locks onto the exact underlying truth.
+* **Agentic RAG Pipeline**: Complex queries are decomposed into atomic sub-queries, routed to multiple Collections in parallel, each executing a Variant loop (Rewrite + Grade), then aggregated and synthesized. Full event stream observability (`decompose → variant_generation → scoring → synthesize_merge`).
 * **Multi-Collection Federated Search**: Context isolation does not limit high-dimensional synthesis. Users can choreograph inquiries spanning multiple explicit Collections simultaneously, harmonizing cross-domain shards via top-level reasoning.
 * **Absolute Source Traceability (3-Layer Traceability)**: Build bulletproof trust in AI answers by drilling straight to the raw text. Instantly inspect the source lineage across three deep-dive levels: the specific *Vector Chunk*, the *Full-Text Context*, or the *Original Source File*.
 * **Recall Evaluation**: Built-in benchmarking with adjustable parameters to evaluate retrieval recall and precision.
@@ -145,7 +158,7 @@ The MCP server uses **stdio** transport — Claude Code launches it as a subproc
 
 ### Backend & Frontend Stack
 
-Python 3.11+, FastAPI, React 19, Vite, TypeScript, Tailwind CSS, Qdrant, FunASR, Zustand, Shadcn UI.
+Python 3.11+, FastAPI, React 19, Vite, TypeScript, Tailwind CSS, Qdrant, FunASR, Zustand, Shadcn UI, SSE Streaming, Session-based Chat, Tool-calling LLM Agent.
 
 ### Directory Layout
 

README_CN.md

@@ -44,15 +44,25 @@ docker compose up -d --build
 
 Docker 将使用最新代码重新构建镜像，同时完整保留您的 `data/` 目录（包含数据库、配置及历史记录）。
 
-### 推荐开箱即用配置 (阿里云百炼 DashScope)
+### 推荐开箱即用配置
 
-本项目对 [阿里云百炼 (DashScope)](https://bailian.console.aliyun.com/) 进行了深度优化。前往 **Settings** → **LLM Providers** → 点击 **OneShot Setting (DashScope API)** 并输入您的 API Key，系统将自动一键配置所有服务，并匹配最佳默认模型：
+项目提供两套 OneShot 一键配置方案：
+
+**阿里云百炼 (DashScope)**：前往 **Settings** → **LLM Providers** → **OneShot Setting (DashScope API)**，输入 API Key 自动配置全部服务：
 
 * **LLM**: `deepseek-v4-flash`
-* **Embedding**: `text-embedding-v4`
+* **Embedding**: `text-embedding-v4`（1024 维）
 * **Reranker**: `qwen3-rerank`
 * **Transcription**: `fun-asr` / `fun-asr-realtime`
 
+**OpenRouter**：前往 **Settings** → **LLM Providers** → **OneShot Setting (OpenRouter API)**，输入 API Key 后自动拉取模型列表并分类：
+
+* **LLM**: `deepseek/deepseek-v4-flash`
+* **Chat**（自动筛选支持工具调用的模型）: `deepseek/deepseek-v4-pro`
+* **Visual**（自动筛选视觉模型）: `xiaomi/mimo-v2.5`
+* **Embedding**: `qwen/qwen3-embedding-4b`（默认 1536 维）
+* **Reranker**: `cohere/rerank-v3.5`
+
 ## 🏗️ 三大核心支柱
 
 ### ⚡ 01. Spark: Fluid Friction Capture & Synthesis
@@ -62,8 +72,10 @@ Docker 将使用最新代码重新构建镜像，同时完整保留您的 `data/
 * **Full-Featured Markdown Workspace (Collection Notes)**: 创建明确绑定到特定业务上下文的结构化个人笔记。支持全功能所见即所得（WYSIWYG）编辑及 Markdown 语法（标题、表格、任务列表、代码块、图片和 YouTube 视频内嵌）。
 * **Intelligent Note Distillation (Drag-to-Distill)**: 直接将现有的旧笔记或文档拖入当前工作区，系统会自动将其核心洞察提取为高密度的引用块，无需手动重写即可无缝聚合零散灵感。
 * **AI Image Ingestion**: One click 触发 AI 生成精准的图像上下文文本描述，将视觉数据织入您的 Markdown 记忆网络中以供向量索引。支持图片粘贴、拖拽和带行内字幕的尺寸调整。
-* **音频转录与“三折页”流水线**: 支持上传录音或通过 WebSockets 录制实时会议/讲座。语音转文字（STT）可由内置的本地引擎或灵活的外部 API 端点处理。系统随后会生成结构化的三折页伪影：*Summary（语义合成摘要）*、*To-Do List/Action Items（清晰的行动项）* 以及 *Detail（深度信息提取，过滤口语杂音并保留核心意图）*。
+* **音频转录与多阶段智能管道**: 支持上传录音或通过 WebSockets 录制实时会议/讲座。语音转文字（STT）可由内置本地 FunASR 引擎或外部 API（DashScope、OpenRouter 等）处理。转写完成后自动运行多阶段分析：**Pass 1 — General Summary + Blueprint（章节结构发现）**，SSE 流式输出，动态构建章节骨架；**Pass 2 — Section Summary（逐章节深度总结）**，支持一键添加自定义 Section，仅需一个关键词即可智能生成该章节的描述与摘要；**Pass 3 — 合并产出**，最终生成 Summary、To-Do List 及 Detail 三折页产物。
 * **Hot Words**: 管理特定专业领域的自定义词汇库，以显著提高特定术语的语音转写准确率。
+* **Quick Chat Panel**: Collection 视图中通过悬浮菱形按钮呼出的侧边栏，使用 Direct RAG 快速问答。每个 Collection 独立会话，支持 SSE 流式输出与 thinking 过程展示。
+* **Multi-turn Agentic Chat**: 主 Chat 支持多轮对话，Agent 通过工具调用（Tool-calling）自主决策是否检索、检索哪些 Collection、以何种模式检索，将 Agentic RAG 的完整能力暴露在对话交互中。
 
 ### 📥 02. Sink: Anti-Hoarding Ingestion Pipeline
 
@@ -74,13 +86,14 @@ Docker 将使用最新代码重新构建镜像，同时完整保留您的 `data/
 * **Granular Document Parsing & Chunking**: 利用内置的本地解析引擎（支持 12 种格式解析器）或链接到强大的云端解析 API（如 *MinerU*）。支持基于标题、段落和最大 Token 配置的高级 **Parent-Child chunking**，同时保持完整的词边界。
 * **Context Enrichment Engine**: 开启后，LLM 将评估每个切分出的数据块，并为其注入缺失的全局上下文，从而缓解检索过程中的“分块孤立”效应。
 * **Auto-Summarization & Consolidation**: 每个文档都会获得一个结构化摘要。同时，集合会生成固化综述，并在文档间出现事实冲突时自动进行冲突检测与提示。
+* **Auto-Updating Collection Coverage**: 集合在每次文档入库后自动触发摘要固化与变化检测，对比新旧文档并更新集合级别的主题覆盖图（Coverage Map），确保检索上下文始终反映最新的知识边界。
 
 ### 🧠 03. Educe: High-Dimensional Contextual Reasoning
 
 **Educe** 实现了前沿的检索架构，将静态的数据转化为活跃的智能。
 
 * **Advanced Hybrid Retrieval**: 支持标准稠密向量相似度检索、关键词-语义混合查询（基于 Qdrant 的 BM25 + Dense），并结合先进的 **Reranking算法** 来锁定顶级上下文。
-* **Iterative Agentic RAG Pipeline**: 远超传统朴素的单次语义检索。Sinkduce 编排了一个多步骤的 Agentic RAG 循环：*Analyze → Route → Retrieve → Grade → Decompose → Rerank → Synthesize*，直到精准锁定底层事实。
+* **Agentic RAG Pipeline**: 复杂查询 → 拆解为原子子查询 → 并行路由至多个 Collection → 各自执行 Variant（Rewrite + Grade loop）→ 聚合合成。全程事件流可观测（`decompose → variant_generation → scoring → synthesize_merge`），每阶段结果可追溯。
 * **Multi-Collection Federated Search**: Context-Isolated 架构并不会限制高维度的知识合成。用户可以设计同时跨越多个明确集合的联合查询，系统将触发相互隔离的多路检索流水线，并通过顶层推理协调跨领域碎片。
 * **Absolute Source Traceability (3-Layer Traceability)**: 允许直接穿透至原始文本，建立对 AI 回答的铁证信任。您可以立即向下钻取三层源头脉络：特定 *Vector Chunk*、*Full-Text Context* 或 *Original Source File*。
 * **Recall Evaluation**: 内置基准测试功能，可通过可调参数评估检索的召回率（Recall）和精确度（Precision）。
@@ -149,7 +162,7 @@ JSON
 
 ### Backend & Frontend Stack
 
-Python 3.11+, FastAPI, React 19, Vite, TypeScript, Tailwind CSS, Qdrant, FunASR, Zustand, Shadcn UI.
+Python 3.11+, FastAPI, React 19, Vite, TypeScript, Tailwind CSS, Qdrant, FunASR, Zustand, Shadcn UI, SSE Streaming, Session-based Chat, Tool-calling LLM Agent.
 
 ### Directory Layout
 

frontend/package.json

@@ -1,7 +1,7 @@
 {
   "name": "frontend",
   "private": true,
-  "version": "0.0.0",
+  "version": "1.0.0",
   "type": "module",
   "scripts": {
     "dev": "vite",