feat(lifelog/crm+chat): 多服务商LLM + tool calling + 语义搜图embed落地

## 多服务商 & 任务槽（ModelConfig + SettingsScreen） - 新增 ModelConfig：LlmProvider 注册表（本地/Hermes + 用户自定义远程）+ ModelSlot（VL/EMBED 槽绑定），SettingsScreen 可视化管理 - ChatConfig 加 ChatTargetKind.REMOTE；targets() 动态合并本地+Hermes+用户服务商 - VisionClient / EmbedClient / PhotoVlWorker 统一走 ModelConfig.slotEndpoint 取端点 ## Tool Calling & Agentic Loop（LifelogTools + ChatViewModel） - LifelogTools：6 个本地工具（search_transcripts/persons/photos/timeline/summaries/followups），OpenAI function schema + 端上 Room 执行 - ChatViewModel：agenticLoop 检测 tool_calls → 调 LifelogTools.execute() → tool role 追回 → 多轮直到无 tool call ## 语义搜图 embed 落地（Pixel 设备侧 + 补算） - 下载 bge-small-zh-v1.5-q8_0.gguf（25MB，512维中文）到 Pixel ~/models/ - llama-swap config.yaml 新增 "embed" 条目（--embedding --pooling mean，ttl 600） - 补算向量：PhotoDao.withCaptionNoEmbedding/updateCaptionEmbedding + PhotologViewModel.backfillCaptionEmbeddings() - AlbumTab ProgressHeader 加「🔢 补算语义向量」按钮（有 caption 无 embedding 时显示，带进度） ## 文档 - HANDOFF.md：P0/P1/P2 全部打 ✅，补 v0.9 已完成列表 - CLAUDE.md：v0.8→v0.9，⭐待做项更新为已完成 Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

feat(lifelog/crm+chat): 多服务商LLM + tool calling + 语义搜图embed落地
## 多服务商 & 任务槽（ModelConfig + SettingsScreen） - 新增 ModelConfig：LlmProvider 注册表（本地/Hermes + 用户自定义远程）+ ModelSlot（VL/EMBED 槽绑定），SettingsScreen 可视化管理 - ChatConfig 加 ChatTargetKind.REMOTE；targets() 动态合并本地+Hermes+用户服务商 - VisionClient / EmbedClient / PhotoVlWorker 统一走 ModelConfig.slotEndpoint 取端点 ## Tool Calling & Agentic Loop（LifelogTools + ChatViewModel） - LifelogTools：6 个本地工具（search_transcripts/persons/photos/timeline/summaries/followups），OpenAI function schema + 端上 Room 执行 - ChatViewModel：agenticLoop 检测 tool_calls → 调 LifelogTools.execute() → tool role 追回 → 多轮直到无 tool call ## 语义搜图 embed 落地（Pixel 设备侧 + 补算） - 下载 bge-small-zh-v1.5-q8_0.gguf（25MB，512维中文）到 Pixel ~/models/ - llama-swap config.yaml 新增 "embed" 条目（--embedding --pooling mean，ttl 600） - 补算向量：PhotoDao.withCaptionNoEmbedding/updateCaptionEmbedding + PhotologViewModel.backfillCaptionEmbeddings() - AlbumTab ProgressHeader 加「🔢 补算语义向量」按钮（有 caption 无 embedding 时显示，带进度） ## 文档 - HANDOFF.md：P0/P1/P2 全部打 ✅，补 v0.9 已完成列表 - CLAUDE.md：v0.8→v0.9，⭐待做项更新为已完成 Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
yibo
1 parent d79aff6c
Showing 16 changed files with 892 additions and 83 deletions
CLAUDE.md
lifelog-recorder/HANDOFF.md
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/ChatScreen.kt
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/MainActivity.kt
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/PhotologScreen.kt
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/PhotologViewModel.kt
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/SettingsScreen.kt
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/chat/ChatClient.kt
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/chat/ChatConfig.kt
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/chat/ChatViewModel.kt
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/chat/LifelogTools.kt
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/chat/ModelConfig.kt
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/data/Daos.kt
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/vision/EmbedClient.kt
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/vision/PhotoVlWorker.kt
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/vision/VisionClient.kt
@@ -140,12 +140,12 @@ Foreground Service (foregroundServiceType=microphone, WAKE_LOCK partial)
 完整待办、估时、技术路线见 `lifelog-recorder/HANDOFF.md` §五。
-✅ **v0.2–v0.8 已完成**：Opus 编码 / BOOT 自启 + 电池白名单 / FTS4 中文搜索 / 段播放 / **声纹在线聚类** / **会话级纪要（按时间间隔切会话）** / **底部 4-tab + 子 tab UI** / **App 内对话（本地 2B/4B + Hermes）** / **本地 LLM 栈（llama-swap）** / **Room 迁移（去 destructive）** / **block-based diarization-first 录音管线（v0.7，攒块→整块分轮→逐 turn 转写）** / **声纹纠偏 L1 冻结 anchor + L2 多 anchor 重建** / **设置页（`SettingsScreen`：纪要模型后端 / 聚类阈值 / llama-swap 端点 / 电池豁免）** / **photolog 人脉 CRM 层（人物档案 + 相册人脸 + 微信通讯录/朋友圈/AI 印象）** / **存储 TTL（超期删 .opus，文字永留）** / **全相册自动索引（v0.8：MediaStore 全量扫描 + Tier-0 人脸/Tier-1 VL 分层管线 + 关键词/语义搜图 + VL 暂停/停止控制，Room v14）**。
+✅ **v0.2–v0.9 已完成**：Opus 编码 / BOOT 自启 + 电池白名单 / FTS4 中文搜索 / 段播放 / **声纹在线聚类** / **会话级纪要（按时间间隔切会话）** / **底部 4-tab + 子 tab UI** / **App 内对话（本地 2B/4B + Hermes）** / **本地 LLM 栈（llama-swap）** / **Room 迁移（去 destructive）** / **block-based diarization-first 录音管线（v0.7，攒块→整块分轮→逐 turn 转写）** / **声纹纠偏 L1 冻结 anchor + L2 多 anchor 重建** / **设置页（`SettingsScreen`：纪要模型后端 / 聚类阈值 / llama-swap 端点 / 电池豁免）** / **photolog 人脉 CRM 层（人物档案 + 相册人脸 + 微信通讯录/朋友圈/AI 印象）** / **存储 TTL（超期删 .opus，文字永留）** / **全相册自动索引（v0.8：MediaStore 全量扫描 + Tier-0 人脸/Tier-1 VL 分层管线 + 关键词/语义搜图 + VL 暂停/停止控制，Room v14）** / **微信抽取完全端上化（v0.8 P2 #6：SQLCipher key 端上推导 + 联系人/头像/聊天/朋友圈全不依赖 Mac）** / **Transcript→Person 提及关系（v0.9 P0 #5：词典反查 + MentionScanner 实时扫 + mentions 表 + 「被提到」tab）** / **人↔人共现边（v0.9 P1 #1：RelationBuilder + relations 表 + 「相关」tab + CommunitiesScreen）** / **待跟进 & 话题实体（FollowupExtractor + TopicExtractor，4B LLM 抽取）**。
 剩余摘要：
-- ⭐ **语义搜图落地**（v0.8 已接 `EmbedClient`，但需在 Pixel llama-swap 配 `embed` 模型才生效；中文 embedding GGUF）
-- ⭐ **内部 Transcript 实体 → Person 表**（端上 NER + 共现，取代旧「回链 lemonPCRM」；见下文 §下一步）
-- ⭐ **微信抽取端上化**（消除 photolog 仅剩的 Mac 依赖；见下文差距分析 #6）
+- ⭐ **语义搜图落地**（代码已完成：`EmbedClient` + `SettingsScreen` embed 字段；**需在 Pixel llama-swap config.yaml 加 embed 模型**才生效，中文 GGUF 如 bge-small-zh，参见 HANDOFF.md §七 调试陷阱）
+- ✅ **内部 Transcript 实体 → Person 表**（已完成：`crm/MentionExtractor.kt` + `MentionScanner` 词典反查 + `BlockProcessor` 实时扫 + `PersonProfileScreen` 「被提到」tab + `RelationBuilder` 共现边）
+- ✅ **微信抽取端上化**（已完成 2026-06-10：SQLCipher 端上解密 + 联系人/头像/朋友圈全端上，不依赖 Mac）
 ### App 即 Personal CRM：与旧 lemonPCRM 的能力对比
@@ -289,33 +289,29 @@ key = MD5(imei + uin).hex()[:7].lower()
 - v0.5 (2026-06)：hermes api_server 端点 + SummaryWorker + 纪要 UI section
 - v0.6 (2026-06-04)：**声纹在线聚类**（SpeakerClusterer，自动「说话人N」+ reclusterAll）/ **5-tab UI**（录音/对话/时间轴/声纹/纪要，时间轴全 CRUD + 批量删导出 + 按说话人/时间/未匹配筛选 + 排序）/ **真暂停**（停采集留模型）/ **会话级纪要**（按时间间隔切会话、可调阈值、要点带 [HH:mm:ss] 点击跳回时间轴、折叠头显示说话人名单）/ **App 内对话**（本地 2B/4B + Hermes 多端点流式）/ **本地 LLM 栈**（llama.cpp HEAD + llama-swap 8080，纪要+对话全本地，见 §六）/ **Room v3 迁移**（去 destructive 改 fail-loud）
 - v0.7 (2026-06-07)：**block-based diarization-first 录音管线**（攒块→整块 pyannote 分轮→逐 turn 转写+嵌入+认人，turn=时间轴一行；去实时字幕；BlockBuffer/BlockProcessor；崩溃 listPending 续跑；commit 07f3d22）/ **声纹纠偏 L1+L2**（L1=reclusterAll & assign 冻结命名 anchor 不漂移止串人；L2=重命名/合并时 rebuildAnchor 用该人全部段均值重建多 anchor）/ 阈值默认 0.45、设置页 0.35-0.65 可调 / **photolog 人脉 CRM 层**（人物档案 + 相册人脸聚类 + 微信通讯录/朋友圈/AI 印象）
-- v0.8 (2026-06-09)：**全相册自动索引 + 分层管线**（GalleryScanner 扫 MediaStore 全量、content uri 原地登记不拷文件；Tier-0 人脸/时间秒级全量铺、Tier-1 VL 充电时排队补；Room v14；详见 §四 photolog 段）/ **搜图**（关键词 caption+人名 LIKE / 语义 caption 向量余弦，embed 走 llama-swap）/ **VL 控制**（PhotoVlState：暂停/继续/停止切换 + 单张转圈 + 格子状态角标）
+- v0.8 (2026-06-09)：**全相册自动索引 + 分层管线**（GalleryScanner 扫 MediaStore 全量、content uri 原地登记不拷文件；Tier-0 人脸/时间秒级全量铺、Tier-1 VL 充电时排队补；Room v14；详见 §四 photolog 段）/ **搜图**（关键词 caption+人名 LIKE / 语义 caption 向量余弦，embed 走 llama-swap）/ **VL 控制**（PhotoVlState：暂停/继续/停止切换 + 单张转圈 + 格子状态角标）/ **微信抽取完全端上化（P2 #6）**（SQLCipher key 端上推导 + WeChatChatImporter + 联系人/头像端上解密；好友/群友区分）
+- v0.9 (2026-06-10)：**Transcript→Person 提及关系（P0 #5）**（`crm/MentionExtractor.kt` 词典反查；`BlockProcessor` 实时扫；`MentionScanner.rescanAll()` 全量重建；Room v15 `mentions` 表；`PersonProfileScreen` 「被提到」tab）/ **人↔人共现边（P1 #1）**（`RelationBuilder.rebuild()` 按会话在场+被提到配对；Room v16 `relations` 表；「相关」tab + `CommunitiesScreen` 圈子可视化）/ **待跟进（P3 #3 部分）**（`FollowupExtractor` 4B LLM 抽待办；Room v17 `followups` 表；`FollowupsScreen`）/ **话题实体（P3 #2 部分）**（`TopicExtractor` 4B 抽 org/project/event；Room v18/v19 `topics`/`topic_mentions`；`TopicsScreen`）
 剩下：
 1. ✅ **存储 TTL**（2026-06-08 完成）：`StorageManager.prune` 按天目录删超期 .opus、文字永留；每天 periodic worker；设置页「存储」段（用量 + 保留时长 + 立即清理），默认永久保留。
 2. ✅ **设置页**（基本完成）：`SettingsScreen` 已含 纪要模型后端 / 聚类·合并·最短识别阈值 / llama-swap 端点 / 分段间隔 / **定时不录（quiet hours）** / 存储用量+TTL / 电池豁免。
 3. ✅ **多 anchor 平均**（2026-06-07 完成，L2）：重命名/合并某说话人时 `rebuildAnchor` 用该人名下全部段的均值重建 anchor，取代单段注册。
-4. **内部 Transcript 实体 → Person 表**（取代旧「回链 lemonPCRM」）：在 **app 内部**从 `Transcript.text`/`Summary` 抽实体（人名/项目，端上 LLM 或规则 NER）→ 匹配本地 `Person` → 建「提及」关系（新表 `mentions(segmentId, personId)`），让人物档案显示「TA 被哪些录音提到」（现仅显示「TA 自己说的」=声纹段）。同段共现的人 → 人↔人边。**不再回链 Mac 端 lemonPCRM**（已废弃）。
+4. ✅ **内部 Transcript 实体 → Person 表**（已完成）：`crm/MentionExtractor.kt`（词典反查：managed 人物的 name/realName/nickname/remark/alias）+ `MentionScanner`（单段实时 & 全量重扫）；`BlockProcessor` 每块处理时实时扫 → `mentions(segmentId, personId, matchedText)`；`PersonProfileScreen` 加「被提到」tab；`rescanMentions()` 结束后自动调 `RelationBuilder.rebuild()`。
 5. ✅ **hermes gateway 自启**（2026-06-04 完成）：用 Magisk `service.d` 脚本开机拉起（Termux:Boot 未装，root 方案更稳），见下方「§ 自启（Magisk service.d）」。
-6. **微信抽取端上化**：消除 photolog 仅剩的 Mac 依赖（微信联系人/头像/聊天/朋友圈现由 Mac 一次性抽取推到 `filesDir/wx_enroll/`）。
+6. ✅ **微信抽取端上化**（2026-06-10 完成，P2 #6）：`WeChatChatImporter`（SQLCipher key 端上推导 + 聊天读取）+ `WeChatSnsImporter`（朋友圈明文库）+ 联系人/头像端上解密（`rcontact`/`img_flag`）；好友 vs 群友区分（`wechatFriend` 字段）；通讯录「🔄更新」一次性整库解密。不再依赖 Mac 导出器。
 ### 端上 CRM 补齐 — TODO（对齐并超越旧 lemonPCRM）
 > 旧 Mac 端 lemonPCRM 已废弃，能力在 app 内重建。下表是「app 当前缺、旧 lemonPCRM 有」的 6 项，
 > 按优先级排（P0 先做）。能力对比详见 CLAUDE.md §「App 即 Personal CRM」。
-- [ ] **P0 · #5 Transcript 内容 → Person「提及」关系**（最高价值，最贴 app 形态）
-  - 现状：录音只按**声纹**连到人（谁说的）；不从转写文本抽人名/项目。
-  - 做法：块处理后（或事后 worker）对 `Transcript.text` 抽实体（端上 LLM 4B 或规则/词典 NER）→ 匹配本地 `Person`（name/realName/nickname/alias，拼音兜底）→ 写新表 `mentions(segmentId, personId, name, confidence)`（Room 迁移 +1）。
-  - 出口：`PersonProfileScreen` 加「被提到」tab（现只有「TA 自己说的」=声纹段）。
-  - 依赖：端上 4B 已就绪（llama-swap 8080）。
-- [ ] **P1 · #1 人↔人关系（共现边）**
-  - 做法：基于 #5，同一会话/时间段被同时提到 + 同时在场（声纹）的人 → `relations(personA, personB, kind, weight)` 共现/认识边，weight 随共现次数累加。
-  - 出口：人物档案「相关的人」；后续可做关系图可视化。
-- [ ] **P2 · #6 微信抽取端上化**（消除最后的 Mac 依赖，工程债）
-  - 现状：微信联系人/头像/聊天/朋友圈由 Mac 解密 `EnMicroMsg.db`(SQLCipher)+抠头像，推到 `filesDir/wx_enroll/`。
-  - 做法：把解密+抽取搬到端上（root + SQLCipher key 端上取），`WeChatEnrollWorker` 直接读微信库；配方见 memory `wechat-android-decrypt.md` / `wechat-avatar-face-enrollment.md`。
+- [x] **P0 · #5 Transcript 内容 → Person「提及」关系**（已完成）
+  - 词典反查（managed 人物 name/realName/nickname/remark/alias 多别名）；`BlockProcessor` 每块实时扫；`MentionScanner.rescanAll()` 全量重扫；`mentions` 表 + `PersonProfileScreen` 「被提到」tab。
+- [x] **P1 · #1 人↔人关系（共现边）**（已完成）
+  - `RelationBuilder.rebuild()`（rescanAll 后自动调）按 5min 静音切会话，在场（声纹）+ 被提到（mentions）的人两两配对 → `relations` 表；`PersonProfileScreen` 「相关」tab + `CommunitiesScreen` 圈子。
+- [x] **P2 · #6 微信抽取端上化**（已完成，2026-06-10）
+  - SQLCipher key 端上推导（MD5(imei+uin)[:7]）+ `WeChatChatImporter` + 联系人/头像/朋友圈全端上。
 - [ ] **P3 · #2 Organization / Project / Event 实体**（产品决策：要不要从「联系人表」长成「图谱」）
   - 新增节点类型 + 与 Person 的边（属于/参与）；可由 #5 的实体抽取顺带产出（项目名/公司名/事件）。
 - [ ] **P3 · #3 关系分析**：关系路径（两人间最短链）/ 社群发现（聚类）/ 待跟进（followups：从录音承诺「下周给他发X」抽待办）。依赖 #1 的关系边。
@@ -192,7 +192,7 @@ internal fun ChatTab(repo: Repository) {
                 }
             }
-            // lifelog 模式开关。
+            // 模式开关行
             Row(
                 modifier = Modifier.fillMaxWidth().padding(horizontal = 12.dp),
                 verticalAlignment = Alignment.CenterVertically,
@@ -201,12 +201,20 @@ internal fun ChatTab(repo: Repository) {
                 FilterChip(
                     selected = vm.lifelogMode,
                     onClick = { vm.setLifelog(!vm.lifelogMode) },
-                    label = { Text("🧠 lifelog 模式") },
+                    label = { Text("🧠 lifelog") },
                 )
-                Text(
-                    if (vm.lifelogMode) "回答会带上你的纪要/转写检索" else "纯聊天，不读 lifelog",
-                    style = MaterialTheme.typography.labelSmall,
+                FilterChip(
+                    selected = vm.toolMode,
+                    onClick = { vm.setTool(!vm.toolMode) },
+                    label = { Text("🔧 工具") },
                 )
+                val statusHint = when {
+                    vm.toolMode && vm.toolStatus.isNotBlank() -> vm.toolStatus
+                    vm.toolMode -> "AI 可自动查录音/联系人/相册"
+                    vm.lifelogMode -> "回答带纪要/转写检索"
+                    else -> "纯聊天"
+                }
+                Text(statusHint, style = MaterialTheme.typography.labelSmall, modifier = Modifier.weight(1f))
             }
             HorizontalDivider()
@@ -122,9 +122,9 @@ class MainActivity : ComponentActivity() {
 }
 private enum class Tab(val label: String, val icon: String) {
-    People("人物", "👤"),       // 通讯录 + 联系人档案（CRM 核心，放第一位）
-    Recording("录音", "⏺"),   // 顶部子页：录音 / 时间轴 / 声纹 / 纪要
-    Chat("对话", "💬"),
+    Chat("对话", "💬"),         // 默认首页：AI 对话 + 工具模式
+    People("人物", "👤"),
+    Recording("录音", "⏺"),
     Album("相册", "🖼"),
     Settings("设置", "⚙"),
 }
@@ -144,7 +144,7 @@ enum class TimeFilter(val label: String) {
 @Composable
 private fun MainScreen() {
-    var tab by remember { mutableStateOf(Tab.People) }
+    var tab by remember { mutableStateOf(Tab.Chat) }
     val activity = androidx.compose.ui.platform.LocalContext.current
     val repo = remember(activity) { Repository(activity) }
@@ -161,6 +161,7 @@ internal fun AlbumTab(repo: Repository) {
     val vlRunning by PhotoVlState.running.collectAsState()
     val vlPaused by PhotoVlState.paused.collectAsState()
     val vlCurrentId by PhotoVlState.currentPhotoId.collectAsState()
+    val embBackfill by vm.embBackfill.collectAsState()
     // 读相册权限（13+ READ_MEDIA_IMAGES；12 及以下 READ_EXTERNAL_STORAGE）。授予即扫全相册。
     val mediaPerm = if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.TIRAMISU)
@@ -221,6 +222,7 @@ internal fun AlbumTab(repo: Repository) {
         } else {
             ProgressHeader(
                 p = progress, running = vlRunning, paused = vlPaused,
+                embBackfill = embBackfill,
                 onRun = {
                     vm.scanGallery(); vm.runVlNow()
                     android.widget.Toast.makeText(
@@ -229,6 +231,7 @@ internal fun AlbumTab(repo: Repository) {
                 },
                 onPause = { vm.pauseVl() },
                 onStop = { vm.stopVl() },
+                onBackfillEmbed = { vm.backfillCaptionEmbeddings() },
             )
         }
         // 搜索框 + 语义按钮。
@@ -312,9 +315,11 @@ private fun ProgressHeader(
     p: PhotoProgress,
     running: Boolean,
     paused: Boolean,
+    embBackfill: Pair<Int, Int>?,
     onRun: () -> Unit,
     onPause: () -> Unit,
     onStop: () -> Unit,
+    onBackfillEmbed: () -> Unit,
 ) {
     if (p.total == 0) {
         Row(
@@ -359,6 +364,23 @@ private fun ProgressHeader(
         } else {
             Text("✓ 全部已理解", style = MaterialTheme.typography.labelSmall)
         }
+        // 补算语义向量（有 caption 但无 embedding 时显示）。
+        if (embBackfill != null) {
+            val (done, total) = embBackfill
+            Row(
+                verticalAlignment = Alignment.CenterVertically,
+                horizontalArrangement = Arrangement.spacedBy(8.dp),
+                modifier = Modifier.padding(top = 4.dp),
+            ) {
+                CircularProgressIndicator(modifier = Modifier.size(14.dp), strokeWidth = 2.dp)
+                Text("语义向量 $done/$total", style = MaterialTheme.typography.labelSmall)
+            }
+        } else if (p.captionDone > 0) {
+            TextButton(
+                onClick = onBackfillEmbed,
+                contentPadding = androidx.compose.foundation.layout.PaddingValues(horizontal = 0.dp, vertical = 0.dp),
+            ) { Text("🔢 补算语义向量", style = MaterialTheme.typography.labelSmall) }
+        }
     }
 }
@@ -116,7 +116,8 @@ class PhotologViewModel(app: Application) : AndroidViewModel(app) {
             val app = getApplication<Application>()
             try {
                 val rows = withContext(Dispatchers.IO) {
-                    val embed = EmbedClient(ChatConfig.localBase(app), ChatConfig.modelEmbed(app))
+                    val (embedBase, embedKey, embedModel) = com.lemon.lifelog.chat.ModelConfig.slotEndpoint(app, com.lemon.lifelog.chat.ModelSlot.EMBED)
+                    val embed = EmbedClient(embedBase, embedModel, embedKey)
                     val q = embed.embed(query.trim())
                     val vecs = repo.photos.captionVecs()
                     if (vecs.isEmpty()) return@withContext emptyList<PhotoRow>()
@@ -138,6 +139,35 @@ class PhotologViewModel(app: Application) : AndroidViewModel(app) {
         }
     }
+    /**
+     * 补算 captionEmbedding：对已有 caption 但缺 embedding 的照片（embed 模型后配时回填），
+     * 只调 embed 端点，不重跑 VL。返回成功补算数量。
+     */
+    private val _embBackfill = MutableStateFlow<Pair<Int, Int>?>(null) // (done, total) | null=idle
+    val embBackfill: StateFlow<Pair<Int, Int>?> = _embBackfill
+
+    fun backfillCaptionEmbeddings() {
+        if (_embBackfill.value != null) return
+        viewModelScope.launch(Dispatchers.IO) {
+            val app = getApplication<Application>()
+            val pending = repo.photos.withCaptionNoEmbedding()
+            if (pending.isEmpty()) { _embBackfill.value = null; return@launch }
+            val (embedBase, embedKey, embedModel) = com.lemon.lifelog.chat.ModelConfig.slotEndpoint(app, com.lemon.lifelog.chat.ModelSlot.EMBED)
+            val client = EmbedClient(embedBase, embedModel, embedKey)
+            var done = 0
+            _embBackfill.value = 0 to pending.size
+            for (p in pending) {
+                runCatching {
+                    val emb = client.embed(p.caption)
+                    repo.photos.updateCaptionEmbedding(p.id, com.lemon.lifelog.vision.ImageUtil.packEmbedding(emb))
+                    done++
+                }
+                _embBackfill.value = done to pending.size
+            }
+            _embBackfill.value = null
+        }
+    }
+
     fun createPersonAndTag(faceId: Long, name: String) {
         viewModelScope.launch {
             // 手动建的人 = 我关心的 → 直接 managed。
@@ -17,6 +17,7 @@ import androidx.compose.foundation.layout.height
 import androidx.compose.foundation.layout.padding
 import androidx.compose.foundation.rememberScrollState
 import androidx.compose.foundation.verticalScroll
+import androidx.compose.material3.AlertDialog
 import androidx.compose.material3.Button
 import androidx.compose.material3.DropdownMenu
 import androidx.compose.material3.DropdownMenuItem
@@ -26,6 +27,7 @@ import androidx.compose.material3.HorizontalDivider
 import androidx.compose.material3.MaterialTheme
 import androidx.compose.material3.OutlinedTextField
 import androidx.compose.material3.Text
+import androidx.compose.material3.TextButton
 import androidx.compose.runtime.Composable
 import androidx.compose.runtime.LaunchedEffect
 import androidx.compose.runtime.getValue
@@ -40,6 +42,9 @@ import com.lemon.lifelog.asr.SpeakerClusterer
 import com.lemon.lifelog.service.BlockBuffer
 import com.lemon.lifelog.service.QuietHours
 import com.lemon.lifelog.chat.ChatConfig
+import com.lemon.lifelog.chat.LlmProvider
+import com.lemon.lifelog.chat.ModelConfig
+import com.lemon.lifelog.chat.ModelSlot
 import com.lemon.lifelog.storage.StorageManager
 import com.lemon.lifelog.summary.SummaryScheduler
 import com.lemon.lifelog.summary.SummaryWorker
@@ -71,14 +76,29 @@ fun SettingsScreen(activity: Context) {
     var pruneMsg by remember { mutableStateOf("") }
     val scope = rememberCoroutineScope()
     LaunchedEffect(Unit) { audioUsage = StorageManager.audioUsageBytes(activity.filesDir) }
-    val backendTargets = remember { ChatConfig.targets(activity) }
+    var configRev by remember { mutableStateOf(0) }
+    val backendTargets = remember(configRev) { ChatConfig.targets(activity) }
     var localBase by remember { mutableStateOf(ChatConfig.localBase(activity)) }
     var m2b by remember { mutableStateOf(ChatConfig.model2b(activity)) }
     var m4b by remember { mutableStateOf(ChatConfig.model4b(activity)) }
-    var mEmbed by remember { mutableStateOf(ChatConfig.modelEmbed(activity)) }
     var endpointMsg by remember { mutableStateOf("") }
+    var hermesUrl by remember(configRev) { mutableStateOf(ModelConfig.hermesProvider(activity).baseUrl) }
+    var hermesKey by remember(configRev) { mutableStateOf(ModelConfig.hermesProvider(activity).apiKey) }
+    var hermesModel by remember(configRev) { mutableStateOf(ModelConfig.hermesProvider(activity).chatModel) }
+    var hermesSaveMsg by remember { mutableStateOf("") }
+
+    var editingProvider by remember { mutableStateOf<LlmProvider?>(null) }
+    var isNewProvider by remember { mutableStateOf(false) }
+
+    var vlPid by remember(configRev) { mutableStateOf(ModelConfig.slotProviderId(activity, ModelSlot.VL)) }
+    var vlModel by remember(configRev) { mutableStateOf(ModelConfig.slotModel(activity, ModelSlot.VL)) }
+    var vlSaveMsg by remember { mutableStateOf("") }
+    var embedPid by remember(configRev) { mutableStateOf(ModelConfig.slotProviderId(activity, ModelSlot.EMBED)) }
+    var embedModel by remember(configRev) { mutableStateOf(ModelConfig.slotModel(activity, ModelSlot.EMBED)) }
+    var embedSaveMsg by remember { mutableStateOf("") }
+
     var batteryExempt by remember { mutableStateOf(isBatteryExempt(activity)) }
     Column(
@@ -125,7 +145,7 @@ fun SettingsScreen(activity: Context) {
             style = MaterialTheme.typography.labelSmall)
         HorizontalDivider()
-        Text("本地模型端点（对话 + 纪要共用 llama-swap）", style = MaterialTheme.typography.titleMedium)
+        Text("本地 llama-swap", style = MaterialTheme.typography.titleMedium)
         OutlinedTextField(value = localBase, onValueChange = { localBase = it },
             label = { Text("baseUrl") }, singleLine = true, modifier = Modifier.fillMaxWidth())
         Row(horizontalArrangement = Arrangement.spacedBy(8.dp)) {
@@ -134,18 +154,101 @@ fun SettingsScreen(activity: Context) {
             OutlinedTextField(value = m4b, onValueChange = { m4b = it },
                 label = { Text("4B model 名") }, singleLine = true, modifier = Modifier.weight(1f))
         }
-        OutlinedTextField(value = mEmbed, onValueChange = { mEmbed = it },
-            label = { Text("embed model 名（语义搜图，llama-swap 里配 --embedding）") },
-            singleLine = true, modifier = Modifier.fillMaxWidth())
         Row(verticalAlignment = Alignment.CenterVertically, horizontalArrangement = Arrangement.spacedBy(8.dp)) {
             Button(onClick = {
-                ChatConfig.save(activity, localBase, m2b, m4b, mEmbed)
+                ChatConfig.save(activity, localBase, m2b, m4b)
                 endpointMsg = "已保存"
-            }) { Text("保存端点") }
+                configRev++
+            }) { Text("保存") }
             if (endpointMsg.isNotEmpty()) Text(endpointMsg, style = MaterialTheme.typography.labelSmall)
         }
         HorizontalDivider()
+        Text("Hermes 网关", style = MaterialTheme.typography.titleMedium)
+        OutlinedTextField(value = hermesUrl, onValueChange = { hermesUrl = it },
+            label = { Text("接口地址") }, singleLine = true, modifier = Modifier.fillMaxWidth())
+        OutlinedTextField(value = hermesKey, onValueChange = { hermesKey = it },
+            label = { Text("API Key") }, singleLine = true, modifier = Modifier.fillMaxWidth())
+        OutlinedTextField(value = hermesModel, onValueChange = { hermesModel = it },
+            label = { Text("路由模型名（如 kimi-for-coding / qwen-plus）") },
+            singleLine = true, modifier = Modifier.fillMaxWidth())
+        Row(verticalAlignment = Alignment.CenterVertically, horizontalArrangement = Arrangement.spacedBy(8.dp)) {
+            Button(onClick = {
+                ModelConfig.saveHermes(activity, hermesUrl, hermesKey, hermesModel)
+                hermesSaveMsg = "已保存"
+                configRev++
+            }) { Text("保存 Hermes") }
+            if (hermesSaveMsg.isNotEmpty()) Text(hermesSaveMsg, style = MaterialTheme.typography.labelSmall)
+        }
+
+        HorizontalDivider()
+        Text("远程服务商", style = MaterialTheme.typography.titleMedium)
+        Text("添加任意 OpenAI 兼容 API（通义千问 / Kimi / MiniMax 等），对话模型名不为空则出现在聊天列表。",
+            style = MaterialTheme.typography.labelSmall)
+        val userProviders = remember(configRev) { ModelConfig.userProviders(activity) }
+        userProviders.forEach { p ->
+            Row(
+                modifier = Modifier.fillMaxWidth(),
+                verticalAlignment = Alignment.CenterVertically,
+                horizontalArrangement = Arrangement.spacedBy(8.dp),
+            ) {
+                Column(modifier = Modifier.weight(1f)) {
+                    Text(p.label, style = MaterialTheme.typography.bodyMedium)
+                    Text(
+                        buildString {
+                            append(p.baseUrl.take(40))
+                            if (p.chatModel.isNotBlank()) append("  ${p.chatModel}")
+                        },
+                        style = MaterialTheme.typography.labelSmall,
+                    )
+                }
+                TextButton(onClick = { editingProvider = p; isNewProvider = false }) { Text("编辑") }
+                TextButton(onClick = { ModelConfig.deleteProvider(activity, p.id); configRev++ }) { Text("删除") }
+            }
+        }
+        Button(onClick = {
+            editingProvider = LlmProvider(ModelConfig.generateId(), "", "", "", "")
+            isNewProvider = true
+        }) { Text("+ 添加服务商") }
+
+        HorizontalDivider()
+        Text("模型分配", style = MaterialTheme.typography.titleMedium)
+        Text("指定 VL（图像理解）和语义向量各用哪个服务商 + 模型。",
+            style = MaterialTheme.typography.labelSmall)
+        val allProviders = remember(configRev) { ModelConfig.allProviders(activity) }
+        // VL slot
+        Text("图像理解(VL)", style = MaterialTheme.typography.labelLarge)
+        DropdownSetting(
+            "服务商", allProviders.firstOrNull { it.id == vlPid }?.label ?: vlPid,
+            allProviders.map { it.label to it.id },
+        ) { vlPid = it }
+        Row(verticalAlignment = Alignment.CenterVertically, horizontalArrangement = Arrangement.spacedBy(8.dp)) {
+            OutlinedTextField(value = vlModel, onValueChange = { vlModel = it },
+                label = { Text("模型名") }, singleLine = true, modifier = Modifier.weight(1f))
+            Button(onClick = {
+                ModelConfig.setSlot(activity, ModelSlot.VL, vlPid, vlModel)
+                vlSaveMsg = "已保存"
+            }) { Text("保存") }
+        }
+        if (vlSaveMsg.isNotEmpty()) Text(vlSaveMsg, style = MaterialTheme.typography.labelSmall)
+        // EMBED slot
+        Text("语义向量", style = MaterialTheme.typography.labelLarge)
+        DropdownSetting(
+            "服务商", allProviders.firstOrNull { it.id == embedPid }?.label ?: embedPid,
+            allProviders.map { it.label to it.id },
+        ) { embedPid = it }
+        Row(verticalAlignment = Alignment.CenterVertically, horizontalArrangement = Arrangement.spacedBy(8.dp)) {
+            OutlinedTextField(value = embedModel, onValueChange = { embedModel = it },
+                label = { Text("模型名（llama-swap 里配 --embedding 的 GGUF 名）") },
+                singleLine = true, modifier = Modifier.weight(1f))
+            Button(onClick = {
+                ModelConfig.setSlot(activity, ModelSlot.EMBED, embedPid, embedModel)
+                embedSaveMsg = "已保存"
+            }) { Text("保存") }
+        }
+        if (embedSaveMsg.isNotEmpty()) Text(embedSaveMsg, style = MaterialTheme.typography.labelSmall)
+
+        HorizontalDivider()
         Text("录音 / 电池", style = MaterialTheme.typography.titleMedium)
         DropdownSetting("分段间隔（静音多久收尾一段）", "$blockSilenceMin 分钟",
             listOf(1, 2, 3, 5, 10, 15).map { "$it 分钟" to it }) {
@@ -209,6 +312,65 @@ fun SettingsScreen(activity: Context) {
         Spacer(Modifier.height(24.dp))
     }
+
+    editingProvider?.let { p ->
+        ProviderEditDialog(
+            initial = p,
+            isNew = isNewProvider,
+            onConfirm = { updated ->
+                if (isNewProvider) ModelConfig.addProvider(activity, updated)
+                else ModelConfig.updateProvider(activity, updated)
+                editingProvider = null
+                configRev++
+            },
+            onDismiss = { editingProvider = null },
+        )
+    }
+}
+
+@Composable
+private fun ProviderEditDialog(
+    initial: LlmProvider,
+    isNew: Boolean,
+    onConfirm: (LlmProvider) -> Unit,
+    onDismiss: () -> Unit,
+) {
+    var label by remember { mutableStateOf(initial.label) }
+    var baseUrl by remember { mutableStateOf(initial.baseUrl) }
+    var apiKey by remember { mutableStateOf(initial.apiKey) }
+    var chatModel by remember { mutableStateOf(initial.chatModel) }
+    AlertDialog(
+        onDismissRequest = onDismiss,
+        title = { Text(if (isNew) "添加服务商" else "编辑服务商") },
+        text = {
+            Column(verticalArrangement = Arrangement.spacedBy(8.dp)) {
+                OutlinedTextField(value = label, onValueChange = { label = it },
+                    label = { Text("名称") }, singleLine = true, modifier = Modifier.fillMaxWidth())
+                OutlinedTextField(value = baseUrl, onValueChange = { baseUrl = it },
+                    label = { Text("接口地址 (baseUrl)") }, singleLine = true, modifier = Modifier.fillMaxWidth())
+                OutlinedTextField(value = apiKey, onValueChange = { apiKey = it },
+                    label = { Text("API Key") }, singleLine = true, modifier = Modifier.fillMaxWidth())
+                OutlinedTextField(value = chatModel, onValueChange = { chatModel = it },
+                    label = { Text("对话模型名（留空则不出现在聊天列表）") },
+                    singleLine = true, modifier = Modifier.fillMaxWidth())
+                Text("接口地址需兼容 OpenAI /v1/chat/completions 协议。",
+                    style = MaterialTheme.typography.labelSmall)
+            }
+        },
+        confirmButton = {
+            TextButton(onClick = {
+                if (label.isNotBlank() && baseUrl.isNotBlank()) {
+                    onConfirm(initial.copy(
+                        label = label.trim(),
+                        baseUrl = baseUrl.trimEnd('/'),
+                        apiKey = apiKey.trim(),
+                        chatModel = chatModel.trim(),
+                    ))
+                }
+            }) { Text("保存") }
+        },
+        dismissButton = { TextButton(onClick = onDismiss) { Text("取消") } },
+    )
 }
 /** 字节数 → 人类可读（KB/MB/GB）。 */
@@ -12,6 +12,18 @@ import org.json.JSONObject
 import java.io.IOException
 import java.util.concurrent.TimeUnit
+data class ToolCall(
+    val id: String,
+    val name: String,
+    val arguments: String,
+)
+
+data class ToolChatResult(
+    val text: String,
+    val toolCalls: List<ToolCall>?,
+    val rawToolCallsJson: JSONArray?,
+)
+
 /**
  * OpenAI 兼容的聊天客户端，支持 SSE 流式（stream:true）增量回复。
  *
@@ -124,4 +136,65 @@ class ChatClient {
         if (choices.length() == 0) return ""
         return choices.getJSONObject(0).optJSONObject("message")?.optString("content", "").orEmpty()
     }
+
+    /**
+     * 非流式 function calling 请求。
+     * 传入 [messages]（JSONObject 列表，支持 tool/assistant with tool_calls 等 role）和
+     * [tools]（LifelogTools.SCHEMAS），返回 [ToolChatResult]：
+     *   - 若模型给出 tool_calls → toolCalls 非 null，调用方执行工具后再次调用本方法
+     *   - 若模型给出最终文字 → toolCalls == null，text 是答案
+     */
+    @Throws(IOException::class)
+    suspend fun chatWithTools(
+        target: ChatTarget,
+        messages: List<JSONObject>,
+        tools: JSONArray,
+        maxTokens: Int = 1024,
+    ): ToolChatResult = withContext(Dispatchers.IO) {
+        val payload = JSONObject().apply {
+            put("model", target.model)
+            put("stream", false)
+            put("max_tokens", maxTokens)
+            put("tools", tools)
+            put("tool_choice", "auto")
+            if (target.kind == ChatTargetKind.LOCAL) {
+                put("chat_template_kwargs", JSONObject().put("enable_thinking", false))
+            }
+            put("messages", JSONArray().apply { messages.forEach { put(it) } })
+        }
+        val builder = Request.Builder()
+            .url("${target.baseUrl}/v1/chat/completions")
+            .header("Content-Type", "application/json")
+            .post(payload.toString().toRequestBody(JSON))
+        if (target.apiKey.isNotBlank()) builder.header("Authorization", "Bearer ${target.apiKey}")
+
+        http.newCall(builder.build()).execute().use { resp ->
+            val body = resp.body?.string().orEmpty()
+            if (!resp.isSuccessful) throw IOException("${target.label} ${resp.code}: ${body.take(300)}")
+            parseToolChatResponse(body)
+        }
+    }
+
+    private fun parseToolChatResponse(body: String): ToolChatResult {
+        val json = JSONObject(body)
+        val choices = json.optJSONArray("choices") ?: throw IOException("no choices")
+        if (choices.length() == 0) throw IOException("empty choices")
+        val message = choices.getJSONObject(0).optJSONObject("message")
+            ?: throw IOException("no message")
+        val text = if (message.isNull("content")) "" else message.optString("content", "")
+        val rawCalls = message.optJSONArray("tool_calls")
+        if (rawCalls != null && rawCalls.length() > 0) {
+            val calls = (0 until rawCalls.length()).map { i ->
+                val tc = rawCalls.getJSONObject(i)
+                val fn = tc.getJSONObject("function")
+                ToolCall(
+                    id = tc.getString("id"),
+                    name = fn.getString("name"),
+                    arguments = fn.optString("arguments", "{}"),
+                )
+            }
+            return ToolChatResult(text, calls, rawCalls)
+        }
+        return ToolChatResult(text, null, null)
+    }
 }
@@ -3,7 +3,7 @@ package com.lemon.lifelog.chat
 import android.content.Context
 import com.lemon.lifelog.BuildConfig
-enum class ChatTargetKind { LOCAL, HERMES }
+enum class ChatTargetKind { LOCAL, HERMES, REMOTE }
 /**
  * 一个可聊天的对象 = 一个 OpenAI 兼容端点 + 一个 model 名。
@@ -34,6 +34,7 @@ object ChatConfig {
     /** 语义搜索的文本 embedding model 名（llama-swap 里配 --embedding 的 GGUF，如 bge-small-zh）。 */
     const val DEFAULT_MODEL_EMBED = "embed"
+    const val ID_LOCAL = "local-swap"
     const val ID_2B = "local-2b"
     const val ID_4B = "local-4b"
     const val ID_HERMES = "hermes"
@@ -61,21 +62,20 @@ object ChatConfig {
             .apply()
     }
-    /** 当前可选的聊天对象列表。 */
+    /** 当前可选的聊天对象列表（本地 2B/4B + Hermes + 用户添加的远程服务商）。 */
     fun targets(ctx: Context): List<ChatTarget> {
         val local = localBase(ctx)
-        return listOf(
-            ChatTarget(ID_2B, "本地 2B", local, "", model2b(ctx), ChatTargetKind.LOCAL),
-            ChatTarget(ID_4B, "本地 4B", local, "", model4b(ctx), ChatTargetKind.LOCAL),
-            ChatTarget(
-                ID_HERMES,
-                "Hermes Agent",
-                BuildConfig.HERMES_API_URL.trimEnd('/'),
-                BuildConfig.HERMES_API_KEY,
-                BuildConfig.HERMES_MODEL,
-                ChatTargetKind.HERMES,
-            ),
-        )
+        val hermes = ModelConfig.hermesProvider(ctx)
+        return buildList {
+            add(ChatTarget(ID_2B, "本地 2B", local, "", model2b(ctx), ChatTargetKind.LOCAL))
+            add(ChatTarget(ID_4B, "本地 4B", local, "", model4b(ctx), ChatTargetKind.LOCAL))
+            add(ChatTarget(ID_HERMES, hermes.label, hermes.baseUrl, hermes.apiKey, hermes.chatModel, ChatTargetKind.HERMES))
+            ModelConfig.userProviders(ctx)
+                .filter { it.chatModel.isNotBlank() }
+                .forEach { p ->
+                    add(ChatTarget(p.id, p.label, p.baseUrl, p.apiKey, p.chatModel, ChatTargetKind.REMOTE))
+                }
+        }
     }
     fun targetById(ctx: Context, id: String?): ChatTarget {
@@ -10,6 +10,8 @@ import androidx.lifecycle.viewModelScope
 import com.lemon.lifelog.data.ChatConversation
 import com.lemon.lifelog.data.ChatMessage
 import com.lemon.lifelog.data.Repository
+import org.json.JSONArray
+import org.json.JSONObject
 import kotlinx.coroutines.Dispatchers
 import kotlinx.coroutines.launch
 import kotlinx.coroutines.withContext
@@ -33,6 +35,10 @@ class ChatViewModel(app: Application) : AndroidViewModel(app) {
         private set
     var lifelogMode by mutableStateOf(prefs.getBoolean(K_LIFELOG, false))
         private set
+    var toolMode by mutableStateOf(prefs.getBoolean(K_TOOL, false))
+        private set
+    var toolStatus by mutableStateOf("")
+        private set
     var input by mutableStateOf("")
     var streaming by mutableStateOf(false)
@@ -47,6 +53,13 @@ class ChatViewModel(app: Application) : AndroidViewModel(app) {
         prefs.edit().putBoolean(K_LIFELOG, on).apply()
     }
+    fun setTool(on: Boolean) {
+        toolMode = on
+        prefs.edit().putBoolean(K_TOOL, on).apply()
+        // 工具模式和 lifelog 模式互斥
+        if (on && lifelogMode) setLifelog(false)
+    }
+
     fun selectTarget(id: String) {
         selectedTargetId = id
         prefs.edit().putString(K_TARGET, id).apply()
@@ -90,8 +103,8 @@ class ChatViewModel(app: Application) : AndroidViewModel(app) {
         input = ""
         val app = getApplication<Application>()
         val target = ChatConfig.targetById(app, selectedTargetId)
-        val useLifelog = lifelogMode
         viewModelScope.launch {
+            // ── 公共准备：建会话 + 存 user 消息 + 取历史 ──────────────────
             val now = System.currentTimeMillis()
             var convId = currentConvId
             if (convId == null) {
@@ -108,48 +121,128 @@ class ChatViewModel(app: Application) : AndroidViewModel(app) {
                 repo.chats.insertMessage(ChatMessage(conversationId = cid, role = "user", content = text, createdAt = now))
                 repo.chats.touch(cid, now)
             }
-            val history = withContext(Dispatchers.IO) { repo.chats.messagesOf(cid) }.map { it.role to it.content }
+            val history = withContext(Dispatchers.IO) { repo.chats.messagesOf(cid) }
             streaming = true
             streamingText = ""
             try {
-                val messagesToSend = if (useLifelog) {
-                    val c = LifelogRetriever.buildContext(repo, text)
-                    if (c != null) listOf("system" to c) + history else history
+                if (toolMode) {
+                    agenticLoop(target, cid, history)
                 } else {
-                    history
-                }
-                val full = client.stream(target, messagesToSend) { delta -> streamingText += delta }
-                withContext(Dispatchers.IO) {
-                    repo.chats.insertMessage(
-                        ChatMessage(
-                            conversationId = cid, role = "assistant",
-                            content = full.ifBlank { "(空回复)" }, model = target.model,
-                            createdAt = System.currentTimeMillis(),
-                        ),
-                    )
-                    repo.chats.touch(cid, System.currentTimeMillis())
+                    normalChat(text, target, cid, history.map { it.role to it.content })
                 }
             } catch (t: Throwable) {
                 withContext(Dispatchers.IO) {
-                    repo.chats.insertMessage(
-                        ChatMessage(
-                            conversationId = cid, role = "assistant",
-                            content = "⚠ 出错：${t.message ?: t.toString()}", model = target.model,
-                            createdAt = System.currentTimeMillis(),
-                        ),
-                    )
+                    repo.chats.insertMessage(ChatMessage(
+                        conversationId = cid, role = "assistant",
+                        content = "⚠ 出错：${t.message ?: t.toString()}", model = target.model,
+                        createdAt = System.currentTimeMillis(),
+                    ))
                 }
             } finally {
                 streaming = false
                 streamingText = ""
+                toolStatus = ""
+            }
+        }
+    }
+
+    private suspend fun normalChat(
+        text: String,
+        target: ChatTarget,
+        cid: Long,
+        history: List<Pair<String, String>>,
+    ) {
+        val messagesToSend = if (lifelogMode) {
+            val c = LifelogRetriever.buildContext(repo, text)
+            if (c != null) listOf("system" to c) + history else history
+        } else {
+            history
+        }
+        val full = client.stream(target, messagesToSend) { delta -> streamingText += delta }
+        withContext(Dispatchers.IO) {
+            repo.chats.insertMessage(ChatMessage(
+                conversationId = cid, role = "assistant",
+                content = full.ifBlank { "(空回复)" }, model = target.model,
+                createdAt = System.currentTimeMillis(),
+            ))
+            repo.chats.touch(cid, System.currentTimeMillis())
+        }
+    }
+
+    /**
+     * 工具模式 agentic loop：
+     *  1. 发带 tools 的请求 → 若 LLM 返回 tool_calls，本地执行 → 追加结果 → 再请求
+     *  2. 直到 LLM 不再调工具（给出最终文字）或达到最大轮数
+     *  3. 只把 user 消息和最终 assistant 文字存 Room（中间工具消息不持久化）
+     */
+    private suspend fun agenticLoop(
+        target: ChatTarget,
+        cid: Long,
+        history: List<ChatMessage>,
+    ) {
+        // 构建初始 messages（system prompt + 全部 Room 历史，已含刚存的 user 消息）
+        val messages: MutableList<JSONObject> = mutableListOf()
+        messages.add(JSONObject().put("role", "system").put("content", TOOL_SYSTEM_PROMPT))
+        history.forEach { m ->
+            messages.add(JSONObject().put("role", m.role).put("content", m.content))
+        }
+
+        var finalText = ""
+        var rounds = 0
+        toolStatus = "思考中…"
+
+        while (rounds < MAX_TOOL_ROUNDS) {
+            val result = withContext(Dispatchers.IO) {
+                client.chatWithTools(target, messages, LifelogTools.SCHEMAS)
+            }
+            if (result.toolCalls.isNullOrEmpty()) {
+                finalText = result.text
+                break
+            }
+            // assistant 消息（含 tool_calls，API 要求原样回传）
+            val assistantMsg = JSONObject().put("role", "assistant").put("content", result.text.ifBlank { "" })
+            result.rawToolCallsJson?.let { assistantMsg.put("tool_calls", it) }
+            messages.add(assistantMsg)
+
+            // 逐个执行工具
+            for (call in result.toolCalls) {
+                toolStatus = "⚙ ${call.name}"
+                streamingText = "调用工具：${call.name}\n"
+                val args = runCatching { JSONObject(call.arguments) }.getOrElse { JSONObject() }
+                val toolResult = LifelogTools.execute(repo, call.name, args)
+                streamingText += toolResult.take(200)
+                messages.add(JSONObject()
+                    .put("role", "tool")
+                    .put("tool_call_id", call.id)
+                    .put("content", toolResult))
             }
+            rounds++
+            toolStatus = "整理回复中（第 $rounds 轮）…"
+            streamingText = toolStatus
+        }
+
+        if (finalText.isBlank()) finalText = "（未获得有效回复，请重试或检查服务商是否支持 function calling）"
+        withContext(Dispatchers.IO) {
+            repo.chats.insertMessage(ChatMessage(
+                conversationId = cid, role = "assistant",
+                content = finalText, model = target.model,
+                createdAt = System.currentTimeMillis(),
+            ))
+            repo.chats.touch(cid, System.currentTimeMillis())
         }
     }
     companion object {
         private const val K_LIFELOG = "lifelog_mode"
+        private const val K_TOOL   = "tool_mode"
         private const val K_TARGET = "selected_target"
-        private const val K_CONV = "current_conv"
+        private const val K_CONV   = "current_conv"
+        private const val MAX_TOOL_ROUNDS = 8
+        private val TOOL_SYSTEM_PROMPT = """
+            你是用户的 lifelog AI 助手，可以访问用户 24/7 录音转写、会话纪要、联系人档案和相册。
+            当用户询问记忆、联系人或照片时，请主动调用相应工具获取真实数据后再回答。
+            多步骤问题可多次调用工具。最终回复用中文，简洁清晰。
+        """.trimIndent()
     }
 }
+package com.lemon.lifelog.chat
+
+import com.lemon.lifelog.data.Repository
+import kotlinx.coroutines.Dispatchers
+import kotlinx.coroutines.flow.first
+import kotlinx.coroutines.withContext
+import org.json.JSONArray
+import org.json.JSONObject
+import java.text.SimpleDateFormat
+import java.util.Date
+import java.util.Locale
+
+/**
+ * lifelog 工具集：OpenAI function calling schema + 本地执行。
+ *
+ * 所有工具直接查 Room DB，在 app 进程内执行，无需网络/额外服务。
+ * agenticLoop（ChatViewModel）检测到 tool_calls → 调 execute() → 把结果以 tool role 追回给 LLM。
+ *
+ * 当前 6 个工具：
+ *   search_transcripts  搜录音转写片段
+ *   get_person          查联系人档案
+ *   list_persons        列 CRM 人物
+ *   get_summaries       取会话纪要
+ *   search_photos       关键词搜相册
+ *   get_followups       待跟进事项
+ */
+object LifelogTools {
+
+    private val TS = SimpleDateFormat("MM-dd HH:mm", Locale.US)
+    private val TSL = SimpleDateFormat("MM-dd HH:mm:ss", Locale.US)
+
+    // ── Schema（OpenAI function calling 格式）────────────────────────────────
+
+    val SCHEMAS: JSONArray by lazy {
+        JSONArray().apply {
+            put(schema(
+                name = "search_transcripts",
+                desc = "在录音转写片段中搜索关键词，返回匹配段落（时间、说话人、内容）",
+                props = mapOf(
+                    "query"   to prop("string",  "搜索关键词"),
+                    "speaker" to prop("string",  "限定说话人姓名，可选"),
+                    "limit"   to prop("integer", "返回条数，默认 12"),
+                ),
+                required = listOf("query"),
+            ))
+            put(schema(
+                name = "get_person",
+                desc = "按姓名或备注查找联系人档案（基本资料、AI 印象、简历、关联声纹/人脸）",
+                props = mapOf("name" to prop("string", "人物姓名或备注名")),
+                required = listOf("name"),
+            ))
+            put(schema(
+                name = "list_persons",
+                desc = "列出 CRM 中的联系人（默认只列我管理的人）",
+                props = mapOf(
+                    "managed_only" to prop("boolean", "true=只看管理的人（默认），false=全部通讯录"),
+                ),
+                required = emptyList(),
+            ))
+            put(schema(
+                name = "get_summaries",
+                desc = "获取最近的会话纪要（LLM 自动生成的录音摘要，时间倒序）",
+                props = mapOf("limit" to prop("integer", "返回篇数，默认 5，最多 20")),
+                required = emptyList(),
+            ))
+            put(schema(
+                name = "search_photos",
+                desc = "在相册中按关键词或人名搜索照片（需要 VL 索引完成后才有描述内容）",
+                props = mapOf(
+                    "query" to prop("string",  "关键词或人名"),
+                    "limit" to prop("integer", "返回张数，默认 8"),
+                ),
+                required = listOf("query"),
+            ))
+            put(schema(
+                name = "get_followups",
+                desc = "获取待跟进事项列表（从录音中 AI 抽取的 TODO/行动项，未完成在前）",
+                props = emptyMap(),
+                required = emptyList(),
+            ))
+        }
+    }
+
+    private fun schema(name: String, desc: String, props: Map<String, JSONObject>, required: List<String>) =
+        JSONObject().apply {
+            put("type", "function")
+            put("function", JSONObject().apply {
+                put("name", name)
+                put("description", desc)
+                put("parameters", JSONObject().apply {
+                    put("type", "object")
+                    put("properties", JSONObject().apply { props.forEach { (k, v) -> put(k, v) } })
+                    put("required", JSONArray().apply { required.forEach { put(it) } })
+                })
+            })
+        }
+
+    private fun prop(type: String, description: String) =
+        JSONObject().put("type", type).put("description", description)
+
+    // ── 执行（在 IO 线程调）──────────────────────────────────────────────────
+
+    suspend fun execute(repo: Repository, name: String, args: JSONObject): String =
+        withContext(Dispatchers.IO) {
+            try {
+                when (name) {
+                    "search_transcripts" -> searchTranscripts(repo, args)
+                    "get_person"         -> getPerson(repo, args)
+                    "list_persons"       -> listPersons(repo, args)
+                    "get_summaries"      -> getSummaries(repo, args)
+                    "search_photos"      -> searchPhotos(repo, args)
+                    "get_followups"      -> getFollowups(repo)
+                    else                 -> "未知工具：$name"
+                }
+            } catch (t: Throwable) {
+                "工具执行出错（$name）：${t.message?.take(120)}"
+            }
+        }
+
+    // ── 各工具实现 ────────────────────────────────────────────────────────────
+
+    private suspend fun searchTranscripts(repo: Repository, args: JSONObject): String {
+        val query = args.optString("query", "").ifBlank { return "请提供搜索关键词" }
+        val speaker = args.optString("speaker", "").ifBlank { null }
+        val limit = args.optInt("limit", 12).coerceIn(1, 40)
+        val rows = repo.segments.searchTextLike(query, limit * 3)
+            .filter { it.text != null }
+            .let { list -> if (speaker != null) list.filter { it.speakerName?.contains(speaker) == true } else list }
+            .take(limit)
+        if (rows.isEmpty()) return "未找到包含「$query」的录音片段。"
+        return buildString {
+            appendLine("找到 ${rows.size} 条片段（关键词：「$query」）：")
+            rows.forEach { r ->
+                val who = r.speakerName ?: r.personName ?: "未知"
+                appendLine("[${TSL.format(Date(r.startTs))}] $who：${r.text?.trim()}")
+            }
+        }
+    }
+
+    private suspend fun getPerson(repo: Repository, args: JSONObject): String {
+        val q = args.optString("name", "").ifBlank { return "请提供人物姓名" }
+        val persons = repo.persons.all().filter { p ->
+            p.name.contains(q) || p.realName?.contains(q) == true ||
+                p.nickname?.contains(q) == true || p.remark?.contains(q) == true
+        }
+        if (persons.isEmpty()) return "未找到与「$q」匹配的联系人。"
+        return buildString {
+            persons.take(3).forEach { p ->
+                appendLine("=== ${p.name} ===")
+                p.realName?.takeIf { it.isNotBlank() }?.let { appendLine("真实姓名：$it") }
+                p.nickname?.takeIf { it.isNotBlank() }?.let { appendLine("微信昵称：$it") }
+                p.remark?.takeIf { it.isNotBlank() }?.let { appendLine("备注：$it") }
+                p.phone?.takeIf { it.isNotBlank() }?.let { appendLine("手机：$it") }
+                if (p.wechatFriend == true) appendLine("微信好友：是")
+                p.impression?.takeIf { it.isNotBlank() }?.let { appendLine("AI 印象：${it.take(400)}") }
+                p.resume?.takeIf { it.isNotBlank() }?.let { appendLine("简历：${it.take(300)}") }
+                val tags = buildList {
+                    if (p.managed) add("CRM 管理中")
+                    if (p.voiceSpeakerId != null) add("有声纹")
+                    if (p.faceEmbedding != null) add("有人脸")
+                }.joinToString(" / ")
+                if (tags.isNotBlank()) appendLine("标签：$tags")
+                appendLine()
+            }
+        }
+    }
+
+    private suspend fun listPersons(repo: Repository, args: JSONObject): String {
+        val managedOnly = args.optBoolean("managed_only", true)
+        val persons = repo.persons.all().let { if (managedOnly) it.filter { p -> p.managed } else it }
+        if (persons.isEmpty()) return if (managedOnly) "CRM 中暂无管理的人物。" else "通讯录为空。"
+        return buildString {
+            appendLine("共 ${persons.size} 位${if (managedOnly) "（管理中）" else ""}：")
+            persons.forEach { p ->
+                val tags = buildList {
+                    if (p.wechatFriend == true) add("微信")
+                    if (p.voiceSpeakerId != null) add("声纹")
+                    if (p.faceEmbedding != null) add("人脸")
+                    if (!p.phone.isNullOrBlank()) add("手机")
+                }.joinToString("/")
+                appendLine("- ${p.name}${if (tags.isNotEmpty()) "（$tags）" else ""}")
+            }
+        }
+    }
+
+    private suspend fun getSummaries(repo: Repository, args: JSONObject): String {
+        val limit = args.optInt("limit", 5).coerceIn(1, 20)
+        val list = repo.summaries.recentList(limit)
+        if (list.isEmpty()) return "暂无会话纪要（需先生成）。"
+        return buildString {
+            appendLine("最近 ${list.size} 篇纪要：")
+            list.forEach { s ->
+                val range = "${TS.format(Date(s.windowStartTs))}–${TS.format(Date(s.windowEndTs))}"
+                appendLine("\n【$range】")
+                appendLine(s.markdown.trim().take(700))
+            }
+        }
+    }
+
+    private suspend fun searchPhotos(repo: Repository, args: JSONObject): String {
+        val query = args.optString("query", "").ifBlank { return "请提供搜索关键词" }
+        val limit = args.optInt("limit", 8).coerceIn(1, 20)
+        val photos = repo.photos.searchByKeyword(query, limit)
+        if (photos.isEmpty()) return "未找到与「$query」相关的照片（VL 描述索引可能尚未完成）。"
+        return buildString {
+            appendLine("找到 ${photos.size} 张相关照片（关键词：「$query」）：")
+            photos.forEach { p ->
+                val date = TS.format(Date(p.takenTs))
+                val caption = p.caption?.take(100) ?: "（暂无场景描述）"
+                val faces = if (p.faceCount > 0) "${p.faceCount}人" else "无人脸"
+                appendLine("[$date $faces] $caption")
+            }
+        }
+    }
+
+    private suspend fun getFollowups(repo: Repository): String {
+        val items = repo.followups.observeAll().first()
+        if (items.isEmpty()) return "暂无待跟进事项。"
+        val open = items.filter { !it.done }
+        val done = items.filter { it.done }
+        return buildString {
+            if (open.isNotEmpty()) {
+                appendLine("未完成（${open.size} 条）：")
+                open.forEach { f ->
+                    val who = f.personName?.let { "（$it）" } ?: ""
+                    appendLine("- ${f.text}$who  [${TS.format(Date(f.sourceTs))}]")
+                }
+            }
+            if (done.isNotEmpty()) {
+                appendLine("\n已完成（${done.size} 条，最近 5 条）：")
+                done.take(5).forEach { f -> appendLine("- ✓ ${f.text}") }
+            }
+        }
+    }
+}
+package com.lemon.lifelog.chat
+
+import android.content.Context
+import com.lemon.lifelog.BuildConfig
+import org.json.JSONArray
+import org.json.JSONObject
+import java.util.UUID
+
+data class LlmProvider(
+    val id: String,
+    val label: String,
+    val baseUrl: String,
+    val apiKey: String,
+    val chatModel: String,
+    val builtIn: Boolean = false,
+)
+
+enum class ModelSlot(val key: String, val label: String, val defaultModel: String) {
+    VL("slot_vl", "图像理解(VL)", "vl"),
+    EMBED("slot_embed", "语义向量", "embed"),
+}
+
+/**
+ * 统一服务商注册表 + 任务槽绑定。
+ *
+ * 服务商（LlmProvider）= 一个 OpenAI 兼容端点（baseUrl + apiKey）+ 对话默认模型名。
+ * 内置两个：本地 llama-swap (ID_LOCAL) 和 本地 Hermes (ID_HERMES)，不可删除。
+ * 用户可添加任意远程服务商（通义千问 / Kimi / MiniMax 等）。
+ *
+ * 任务槽（ModelSlot）= 某项能力用哪个 (provider, model)：
+ *   - VL：图像理解，PhotoVlWorker 读取
+ *   - EMBED：语义向量，PhotologViewModel / PhotoVlWorker 读取
+ *   - CHAT：用哪个 target 聊天 → 由 ChatConfig.targets() + ChatViewModel.selectedTargetId 管理，不在此
+ *   - SUMMARY：纪要用哪个 target → 同上，存 SummaryWorker.KEY_BACKEND（ChatTarget.id）
+ *
+ * 持久化：SharedPreferences "model_config"，JSON 数组 + 独立 key，无 Room 依赖。
+ */
+object ModelConfig {
+    const val PREFS = "model_config"
+    private const val K_PROVIDERS = "user_providers"
+    private const val K_HERMES_URL = "hermes_url"
+    private const val K_HERMES_KEY = "hermes_key"
+    private const val K_HERMES_MODEL = "hermes_model"
+
+    private fun prefs(ctx: Context) = ctx.getSharedPreferences(PREFS, Context.MODE_PRIVATE)
+
+    // ── 内置：Hermes 网关 ──────────────────────────────────────────────────────
+
+    fun hermesProvider(ctx: Context): LlmProvider {
+        val p = prefs(ctx)
+        return LlmProvider(
+            id = ChatConfig.ID_HERMES,
+            label = "本地 Hermes",
+            baseUrl = p.getString(K_HERMES_URL, null)?.ifBlank { null } ?: BuildConfig.HERMES_API_URL,
+            apiKey = p.getString(K_HERMES_KEY, null) ?: BuildConfig.HERMES_API_KEY,
+            chatModel = p.getString(K_HERMES_MODEL, null)?.ifBlank { null } ?: BuildConfig.HERMES_MODEL,
+            builtIn = true,
+        )
+    }
+
+    fun saveHermes(ctx: Context, url: String, key: String, model: String) {
+        prefs(ctx).edit()
+            .putString(K_HERMES_URL, url.trim())
+            .putString(K_HERMES_KEY, key.trim())
+            .putString(K_HERMES_MODEL, model.trim())
+            .apply()
+    }
+
+    // ── 内置：本地 llama-swap ─────────────────────────────────────────────────
+
+    fun localSwapProvider(ctx: Context) = LlmProvider(
+        id = ChatConfig.ID_LOCAL,
+        label = "本地 llama-swap",
+        baseUrl = ChatConfig.localBase(ctx),
+        apiKey = "",
+        chatModel = "",
+        builtIn = true,
+    )
+
+    // ── 用户添加的远程服务商 ───────────────────────────────────────────────────
+
+    fun userProviders(ctx: Context): List<LlmProvider> {
+        val json = prefs(ctx).getString(K_PROVIDERS, null) ?: return emptyList()
+        return try {
+            val arr = JSONArray(json)
+            (0 until arr.length()).map { arr.getJSONObject(it).toProvider() }
+        } catch (t: Throwable) {
+            emptyList()
+        }
+    }
+
+    fun allProviders(ctx: Context): List<LlmProvider> = buildList {
+        add(localSwapProvider(ctx))
+        add(hermesProvider(ctx))
+        addAll(userProviders(ctx))
+    }
+
+    fun addProvider(ctx: Context, p: LlmProvider) {
+        saveUserProviders(ctx, userProviders(ctx) + p)
+    }
+
+    fun updateProvider(ctx: Context, p: LlmProvider) {
+        if (p.id == ChatConfig.ID_HERMES) {
+            saveHermes(ctx, p.baseUrl, p.apiKey, p.chatModel)
+            return
+        }
+        if (p.builtIn) return
+        val list = userProviders(ctx).toMutableList()
+        val idx = list.indexOfFirst { it.id == p.id }
+        if (idx >= 0) list[idx] = p else list.add(p)
+        saveUserProviders(ctx, list)
+    }
+
+    fun deleteProvider(ctx: Context, id: String) {
+        if (id == ChatConfig.ID_LOCAL || id == ChatConfig.ID_HERMES) return
+        saveUserProviders(ctx, userProviders(ctx).filter { it.id != id })
+        // 如果某个 slot 引用了被删的 provider，重置回默认
+        for (slot in ModelSlot.entries) {
+            if (slotProviderId(ctx, slot) == id) {
+                setSlot(ctx, slot, ChatConfig.ID_LOCAL, slot.defaultModel)
+            }
+        }
+    }
+
+    private fun saveUserProviders(ctx: Context, list: List<LlmProvider>) {
+        val arr = JSONArray()
+        list.forEach { arr.put(it.toJson()) }
+        prefs(ctx).edit().putString(K_PROVIDERS, arr.toString()).apply()
+    }
+
+    // ── 任务槽 ────────────────────────────────────────────────────────────────
+
+    fun slotProviderId(ctx: Context, slot: ModelSlot): String =
+        prefs(ctx).getString("${slot.key}_pid", null)?.ifBlank { null } ?: ChatConfig.ID_LOCAL
+
+    fun slotModel(ctx: Context, slot: ModelSlot): String =
+        prefs(ctx).getString("${slot.key}_model", null)?.ifBlank { null } ?: slot.defaultModel
+
+    fun setSlot(ctx: Context, slot: ModelSlot, providerId: String, model: String) {
+        prefs(ctx).edit()
+            .putString("${slot.key}_pid", providerId)
+            .putString("${slot.key}_model", model.trim())
+            .apply()
+    }
+
+    /** (baseUrl, apiKey, modelName) 三元组，供 VisionClient / EmbedClient 直接使用。 */
+    fun slotEndpoint(ctx: Context, slot: ModelSlot): Triple<String, String, String> {
+        val pid = slotProviderId(ctx, slot)
+        val model = slotModel(ctx, slot)
+        val provider = allProviders(ctx).firstOrNull { it.id == pid } ?: localSwapProvider(ctx)
+        return Triple(provider.baseUrl, provider.apiKey, model)
+    }
+
+    fun generateId(): String = "p_${UUID.randomUUID().toString().take(8)}"
+
+    // ── JSON 序列化 ────────────────────────────────────────────────────────────
+
+    private fun JSONObject.toProvider() = LlmProvider(
+        id = getString("id"),
+        label = getString("label"),
+        baseUrl = getString("baseUrl"),
+        apiKey = optString("apiKey", ""),
+        chatModel = optString("chatModel", ""),
+    )
+
+    private fun LlmProvider.toJson() = JSONObject().apply {
+        put("id", id)
+        put("label", label)
+        put("baseUrl", baseUrl)
+        put("apiKey", apiKey)
+        put("chatModel", chatModel)
+    }
+}
@@ -501,6 +501,9 @@ data class PhotoVec(
     val captionEmbedding: ByteArray,
 )
+/** 补算 embedding 用的轻投影：已有 caption 但缺 embedding 的照片。 */
+data class CaptionOnly(val id: Long, val caption: String)
+
 /** 某人物名下照片数（相册「人物文件夹」封面 + 计数）。 */
 data class PersonPhotoCount(
     val personId: Long,
@@ -749,6 +752,14 @@ interface PhotoDao {
     /** 按 path 模糊查 id（清理旧版误导入的朋友圈图用，pat 形如 "%/moment_<wxid>_%"）。 */
     @Query("SELECT id FROM photos WHERE path LIKE :pat")
     suspend fun idsByPathLike(pat: String): List<Long>
+
+    /** 补算向量：已有 caption 但还没算 embedding 的照片（embed 模型后配时回填用）。 */
+    @Query("SELECT id, caption FROM photos WHERE caption IS NOT NULL AND caption != '' AND captionEmbedding IS NULL")
+    suspend fun withCaptionNoEmbedding(): List<CaptionOnly>
+
+    /** 只写 captionEmbedding，不改其他字段（backfill 用，不重跑 VL）。 */
+    @Query("UPDATE photos SET captionEmbedding = :emb WHERE id = :id")
+    suspend fun updateCaptionEmbedding(id: Long, emb: ByteArray)
 }
 @Dao
@@ -22,6 +22,7 @@ import kotlin.math.sqrt
 class EmbedClient(
     private val baseUrl: String,
     private val model: String = "embed",
+    private val apiKey: String = "",
 ) {
     companion object {
         private val JSON = "application/json; charset=utf-8".toMediaType()
@@ -38,11 +39,12 @@ class EmbedClient(
             put("model", model)
             put("input", text)
         }
-        val req = Request.Builder()
+        val reqBuilder = Request.Builder()
             .url("$baseUrl/v1/embeddings")
             .header("Content-Type", "application/json")
             .post(payload.toString().toRequestBody(JSON))
-            .build()
+        if (apiKey.isNotBlank()) reqBuilder.header("Authorization", "Bearer $apiKey")
+        val req = reqBuilder.build()
         http.newCall(req).execute().use { resp ->
             val body = resp.body?.string().orEmpty()
             if (!resp.isSuccessful) throw IOException("embed ${resp.code}: ${body.take(200)}")
@@ -4,7 +4,8 @@ import android.content.Context
 import android.util.Log
 import androidx.work.CoroutineWorker
 import androidx.work.WorkerParameters
-import com.lemon.lifelog.chat.ChatConfig
+import com.lemon.lifelog.chat.ModelConfig
+import com.lemon.lifelog.chat.ModelSlot
 import com.lemon.lifelog.data.Repository
 import java.util.concurrent.atomic.AtomicBoolean
@@ -28,9 +29,10 @@ class PhotoVlWorker(ctx: Context, params: WorkerParameters) : CoroutineWorker(ct
             val pending = repo.photos.pendingCaption()
             if (pending.isEmpty()) return Result.success()
-            val base = ChatConfig.localBase(applicationContext)
-            val vision = VisionClient(base)
-            val embed = EmbedClient(base, ChatConfig.modelEmbed(applicationContext))
+            val (vlBase, vlKey, vlModel) = ModelConfig.slotEndpoint(applicationContext, ModelSlot.VL)
+            val (embedBase, embedKey, embedModel) = ModelConfig.slotEndpoint(applicationContext, ModelSlot.EMBED)
+            val vision = VisionClient(vlBase, vlModel, vlKey)
+            val embed = EmbedClient(embedBase, embedModel, embedKey)
             val proc = PhotoProcessor(applicationContext, repo)
             val deadline = System.currentTimeMillis() + RUN_BUDGET_MS
             var done = 0
@@ -17,6 +17,7 @@ import java.util.concurrent.TimeUnit
 class VisionClient(
     private val baseUrl: String,
     private val model: String = "vl",
+    private val apiKey: String = "",
 ) {
     companion object {
         private val JSON = "application/json; charset=utf-8".toMediaType()
@@ -48,11 +49,12 @@ class VisionClient(
                 })
             }))
         }
-        val req = Request.Builder()
+        val reqBuilder = Request.Builder()
             .url("$baseUrl/v1/chat/completions")
             .header("Content-Type", "application/json")
             .post(payload.toString().toRequestBody(JSON))
-            .build()
+        if (apiKey.isNotBlank()) reqBuilder.header("Authorization", "Bearer $apiKey")
+        val req = reqBuilder.build()
         http.newCall(req).execute().use { resp ->
             val body = resp.body?.string().orEmpty()
             if (!resp.isSuccessful) throw IOException("vl ${resp.code}: ${body.take(200)}")