add image uploader and update README

Author: Oli51467

README.md

@@ -20,6 +20,8 @@ LoFS 聚焦“本地优先”的知识管理场景，将文件管理与语义检
 - 🔍 **多模态处理**：自动解析 `.md`、`.txt`、`.docx`、`.pdf`、`.pptx`、`.json` 等文本与图像内容。
 - 📁 **文件树交互**：支持一键挂载/卸载、PDF 专属解析、进度实时可视化。
 - 🧠 **主题驱动检索**：在传统 Faiss/BM25s/reranker 之上引入“文档主题摘要 + 语义/词法加权”的两阶段检索，问题可先与摘要匹配再回退全文。
+- 🤖 **多平台模型接入**：内置 SiliconFlow，并新增对 ModelScope、通义千问（DashScope）的兼容，支持在线连通性检测、流式问答与 `<think>` 思考返回。
+- 💬 **智能聊天体验**：聊天页记住上次选用的模型，重启后即刻恢复；模型输出的思考内容会以灰色卡片独立展示，方便分析与引用。
 - 🔒 **本地优先**：所有文件、向量与索引均保存在本地 SQLite 与 Faiss 中，核心数据不出域。
 - 🛠️ **开箱即用**：Electron 桌面端 + FastAPI 后端，兼容多系统打包。
 
@@ -31,6 +33,7 @@ LoFS 聚焦“本地优先”的知识管理场景，将文件管理与语义检
 - **桌面端（Electron）**：负责文件树、解析状态、检索结果的交互体验。
 - **后端（FastAPI）**：承载挂载、解析、索引构建等核心任务，提供 REST API。
 - **检索引擎**：通过 Faiss、BM25s 与 FlagEmbedding（BGE 系列）构成的混合检索流水线，支持文本与图像向量，并内置主题摘要检索通路。
+- **模型管理**：前端提供模型库视图，可添加自定义模型并测试连通性；支持 SiliconFlow、ModelScope 与通义千问（DashScope）三大平台。
 - **数据存储**：SQLite 保存结构化元数据，Faiss 保存向量索引，本地文件系统管理模型与缓存。
 
 ```text
@@ -51,6 +54,7 @@ LoFS 聚焦“本地优先”的知识管理场景，将文件管理与语义检
 3. 🧮 **索引构建**：将文本向量写入 Faiss，将关键词倒排写入 BM25s，元数据落盘至 SQLite。
 4. 🧾 **生成主题摘要**（可选）：启用“文档总结”后，调用指定模型为整篇文档生成主题概述，并写入独立向量与 SQLite。
 5. 🔎 **分层检索**：输入查询后先与主题摘要做语义/词法混合匹配（分数 ≥ 0.7 即入选），命中文档的完整内容 + 摘要会一并注入提示词；若无命中则自动回退至传统混合检索。
+6. 💬 **LLM 问答**：选择任一已配置的模型即可发起流式对话；若模型支持思考输出（ModelScope、DashScope），LoFS 会将 `<think>` 内的内容转成灰色卡片单独展示。
 
 模型目录按需懒加载，首次使用对应能力时自动下载；也可以提前拉取以免首轮等待：
 
@@ -64,6 +68,14 @@ LoFS 聚焦“本地优先”的知识管理场景，将文件管理与语义检
 python -c "from service.model_manager import get_model_manager; manager = get_model_manager(); [manager.get_model_path(key) for key in ('bge_m3', 'bge_reranker_v2_m3', 'clip_vit_b_32', 'clip_vit_b_32_multilingual', 'pdf_extract_kit')]"
 ```
 
+### 3.2 多平台模型接入
+- 在 **设置 → API Key** 中填写对应凭证：
+  - SiliconFlow：`siliconflwApiKey`
+  - ModelScope：`modelscopeApiKey`
+  - 通义千问（DashScope）：`qwenApiKey`
+- 进入 **模型库 → 添加模型**，选择平台即可测试连通性；失败时会返回接口错误信息辅助排查。
+- 聊天页面的模型下拉会记住上次选择，重启后自动恢复；支持流式输出，若模型返回 `<think>` 思考段落会以灰色卡片单独展示（默认适用于 ModelScope、DashScope）。
+
 ## 4. 部署使用
 ### 4.1 环境要求
 | 组件 | 最低版本 | 推荐 |

README_EN.md

@@ -20,6 +20,8 @@ LoFS fuses local file organization with semantic retrieval to deliver an “alwa
 - 🔍 **Multimodal ingestion**: parses `.md`, `.txt`, `.docx`, `.pdf`, `.pptx`, `.json`, and extracts both text and images.
 - 📁 **Explorer-style UX**: mount/remount flows, PDF-only parsing, and live progress indicators keep operations transparent.
 - 🧠 **Theme-driven retrieval**: augments the Faiss/BM25s/reranker stack with document-level summaries plus semantic/lexical blending so questions land on the right topic before falling back to full-text recall.
+- 🤖 **Multi-provider LLM support**: built-in adapters for SiliconFlow, ModelScope, and Alibaba DashScope (Qwen) let you plug in your favorite service, validate connectivity, and use streaming responses with `<think>` reasoning blocks.
+- 💬 **Chat experience upgrades**: the chat view remembers your last-selected model across restarts, and thinking traces render as dedicated gray cards for easier auditing.
 - 🔒 **Local-first by design**: SQLite and Faiss stay on disk, ensuring data never leaves your machine.
 - 🛠️ **Shipping-ready app**: Electron desktop shell plus FastAPI backend, complete with cross-platform packaging scripts.
 
@@ -31,6 +33,7 @@ LoFS fuses local file organization with semantic retrieval to deliver an “alwa
 - **Electron desktop**: renders the file tree, orchestration panels, and search UI.
 - **FastAPI backend**: exposes REST endpoints for mounting, parsing, indexing, and retrieval orchestration.
 - **Retrieval pipeline**: Faiss + BM25s + FlagEmbedding (BGE family) power fast semantic/keyword blending, now enhanced with summary vectors for topic-first search, plus CLIP embeddings for images.
+- **Model management**: the desktop UI exposes a model library so you can add custom endpoints, run health checks, and manage keys for SiliconFlow, ModelScope, and DashScope in one place.
 - **Storage layer**: SQLite for metadata, Faiss for vector indices, and the local filesystem for model caches.
 
 ```text
@@ -51,6 +54,7 @@ LoFS fuses local file organization with semantic retrieval to deliver an “alwa
 3. 🧮 **Index build**: persist embeddings to Faiss, keyword metadata to BM25s, and structured info to SQLite.
 4. 🧾 **Topic summaries (optional)**: when the “Document Theme Summary” setting is on, a chosen LLM produces per-document abstracts that are stored in SQLite and embedded separately.
 5. 🔎 **Layered retrieval**: each query first matches against the summary vectors with a semantic (0.6) + lexical (0.4) score fusion; summaries scoring ≥ 0.7 bring their full document chunks and the summary into the LLM prompt. If nothing passes the threshold, the pipeline transparently falls back to classic hybrid retrieval.
+6. 💬 **Conversational hand-off**: when a model is selected in the chat panel, LoFS streams replies (including `<think>` reasoning for ModelScope/DashScope) and caches the selection so it’s restored on the next launch.
 
 Model assets download lazily the first time a capability is invoked. Prefetch them to warm the cache:
 
@@ -64,6 +68,11 @@ python -c "from service.model_manager import get_model_manager; manager = get_mo
 - 🔁 **Graceful fallback**: if no summary clears the threshold, LoFS automatically reverts to the standard hybrid pipeline; matched responses surface a “Reference Theme” card for transparency.
 - 🧩 **Prompt enrichment**: selected documents contribute both the full chunk set and their summaries to the LLM context, helping the assistant stay on-topic and cite accurately.
 
+### 3.2 Multi-provider LLM integration
+- Configure API keys under **Settings → API Key** (`siliconflwApiKey`, `modelscopeApiKey`, `qwenApiKey`).
+- Use the **Model Library** to add endpoints and run connectivity tests; failures return verbose error payloads for troubleshooting.
+- During chat, the dropdown retains your last pick even after restarts. Models that emit `<think>` blocks (ModelScope/DashScope) render their reasoning in a separate, gray “Thinking” panel ahead of the final answer.
+
 ## 4. Deployment & Usage
 ### 4.1 Requirements
 | Component | Minimum | Recommended |

electron/index.html

@@ -277,7 +277,19 @@ <h1 class="splash-title">LoFS</h1>
       </div>
       <div id="chat-input-area">
         <div id="chat-status-text"></div>
+        <div id="chat-image-preview" class="chat-image-preview" hidden>
+          <div id="chat-image-preview-list" class="chat-image-preview-list"></div>
+        </div>
         <div class="chat-input-wrapper">
+          <button id="chat-upload-btn" class="chat-upload-btn" type="button" aria-label="上传图片">
+            <svg width="18" height="18" viewBox="0 0 24 24" aria-hidden="true" focusable="false">
+              <path d="M4 7h16v12H4z" stroke="currentColor" stroke-width="1.5" stroke-linejoin="round" fill="none" />
+              <path d="M4 16l3.5-3.5a2 2 0 0 1 2.9 0L16 18" stroke="currentColor" stroke-width="1.5"
+                stroke-linecap="round" stroke-linejoin="round" fill="none" />
+              <circle cx="15.5" cy="11" r="1.75" fill="currentColor" />
+            </svg>
+          </button>
+          <input id="chat-upload-input" type="file" accept="image/*" multiple aria-hidden="true">
           <textarea id="chat-input" rows="1" placeholder="询问任何问题" autocomplete="off" spellcheck="false"></textarea>
           <button id="chat-send-btn" type="button" aria-label="发送消息">
             <svg class="icon-arrow" width="18" height="18" viewBox="0 0 24 24" aria-hidden="true" focusable="false">

electron/package-lock.json

@@ -1,6 +1,6 @@
 {
   "name": "LoFS",
-  "version": "1.1.8",
+  "version": "1.1.9",
   "description": "LoFS (Load once Fast Search) is a desktop-first knowledge base manager for local files. Mount a directory once and enjoy millisecond hybrid search afterwards.",
   "main": "main.js",
   "scripts": {