feat(chat): record and show what each answer was built from

DESIGN.md

@@ -216,6 +216,73 @@ Two rules for the composer and for leaving the workspace:
   through. The in-panel Back button is not special-cased — it uses the same
   state. A navigation path that skips this guard is a bug.
 
+### An answer shows what it was built from
+
+Implementation: `components/notebook/chat/AnswerSources.tsx`, fed by
+`chat_messages.metadata` (see [message metadata](#message-metadata)).
+
+Retrieval already knew which passages it used and threw the identity away: the
+prompt kept a title and some text, and `chunkId` / `documentId` were dropped. The
+answer therefore looked identical whether it came from the reader's documents or
+from nowhere, which is the one distinction this product cannot afford to lose.
+
+The region has three states, and they are three different statements — do not
+collapse them:
+
+| State            | Shown                                                         |
+| ---------------- | ------------------------------------------------------------- |
+| `used`           | A disclosure: "Based on N source passages" → the passages     |
+| `none`           | One T3 line: nothing from your sources was used               |
+| `failed`         | One T3 line: searching your sources failed                    |
+| no status at all | **Nothing** — an older message is "unknown", not "ungrounded" |
+
+Rules:
+
+- **The passage is quoted verbatim**, never summarised or truncated into a
+  paraphrase. It is the evidence; if it is too long to read, it is still too long
+  to invent.
+- Each passage sits in `bg-muted` — a recessed container inside the panel, per the
+  surface rules. The region is **not** a `Card`: cards inside panels nest, and this
+  belongs to the message, not beside it.
+- Density is the disclosure's job. The collapsed state is one line; the passages
+  are behind it.
+- Only after the turn ends. Rendering an empty evidence list mid-answer would read
+  as "nothing was used".
+- The ungrounded and failed lines are T3: they must be readable, and they must not
+  compete with the answer.
+
+### Message metadata
+
+The structured part of `chat_messages.metadata` is typed (`ChatMessageMetadata`
+in `shared/types/chat.ts`) and read **only** through
+`shared/utils/answerSources.ts`. Two rules:
+
+- **Parse defensively, drop per entry.** The column is an open JSON bag written by
+  whichever version of the app produced the message, so a reader never assumes a
+  shape. One malformed entry must not hide the passages that did survive: two
+  usable sources out of three still show two.
+- **"Not recorded" is not "none".** An answer from before this existed has no
+  status; saying it was ungrounded would accuse a grounded answer. Unknown
+  renders nothing.
+
+### Cross-panel requests
+
+The transcript (centre) and the library (left) are siblings, so a citation click
+cannot pass a prop. It writes to `uiStore` and `SourcePanel` derives from it **during
+render** — never in an effect:
+
+```tsx
+const openDocument = selectedDocument ?? focusedDocument
+```
+
+Two things this avoids, both of which were tried and rejected:
+
+- `setState` inside an effect to consume the request: the lint rule
+  `react-hooks/set-state-in-effect` fails the build, and it cascades renders.
+- Clearing the store request from inside render: writing to a store during render
+  notifies other components mid-render. The clear happens in the event handlers
+  instead (list click, Back), which is why the derivation needs no cleanup step.
+
 ### Geometry lives in a module, and it is tested
 
 The arithmetic above — `availableFrom`, `maxSideWidth`, the golden-ratio split, the
@@ -638,3 +705,20 @@ runs in the `Verify` workflow, before the build matrix.
 What it does **not** enforce, and therefore relies on review: the surface ladder
 (which surface a panel uses), accent discipline, spacing and density, and the
 four interaction states. Those are the rules a reviewer must hold the line on.
+
+## Locales
+
+`en-US` and `zh-CN` move together (CONTRIBUTING.md), and the check is **parity**:
+neither file may carry a key the other lacks.
+
+**One exception: plural forms.** English declares `key_one` / `key_other` and
+Chinese declares `key` alone with no suffix, which is the correct i18next shape and
+not a missing translation. This matters when pruning: a literal search for a key
+misses plural forms entirely, because `t('ankiCards', { count })` never renders the
+string `ankiCards_one` in source. Prune by hand for those, or keep them
+unconditionally.
+
+Two prunes have already found the same artefact: each locale file carried a block
+of English-only keys for surfaces that do not exist (an editor menubar, a notes /
+trash / tags feature). They were unreachable, and they made the app read as
+half-translated. Delete dead copy rather than translating it.

PRODUCT.md

@@ -113,6 +113,24 @@ Confirmed by the README and the code:
   `Verify` workflow is the gate; `en-US` and `zh-CN` locales move together
   (CONTRIBUTING.md).
 
+- **A notebook holds one growing conversation.** Confirmed by the maintainer. There
+  is no conversation picker and no "new chat": opening a notebook lands you in its
+  ongoing thread, and the store already creates the session on demand. When that
+  thread approaches the model's context budget, `SessionAutoSwitchService`
+  summarises it, archives the old row and continues in a new one — a rollover, not
+  a second conversation the user has to manage. Because the user experiences one
+  thread, a rollover must never be silent.
+- **Every answer records the passages it was built from.** Retrieval already knows
+  the `documentId`, `chunkId` and passage text; these are now persisted on the
+  assistant message and rendered as quotable evidence, with the retrieval outcome
+  recorded too (`used` / `none` / `failed`). What this does **not** yet provide is a
+  page or character span: the indexing path computes chunk offsets against a
+  preprocessed string rather than `documents.content`, and `parseResult.structure`
+  is still discarded, so a span-accurate citation would be a false claim until #82
+  lands. Inline per-sentence markers (the `[cite:id]` contract Cherry Studio and
+  Open Notebook both use) also need a system-prompt change and real model testing,
+  so they are deliberately not in place yet.
+
 Undecided, recorded rather than resolved:
 
 - The README says quiz generation, audio transcription and slide generation are
@@ -160,6 +178,9 @@ needed, ask for it.
 
 1. **Provenance before fluency.** An answer without a resolvable source location
    is an unverified claim. The citation is the product, not the chat bubble.
+   Today the answer carries the passages it used, and that is the first half of
+   this principle. It becomes the whole of it when a citation resolves to a page
+   and a span rather than to a passage's text (#82).
 2. **Local is the default, not a mode.** Retrieval has to work with the network
    off, and the only outbound request is the endpoint the user configured.
 3. **Say what is not there.** A claim ahead of the code is a bug, and the README

src/main/db/queries.ts

@@ -1,6 +1,7 @@
 import { eq, desc, and } from 'drizzle-orm'
 import { getDatabase, executeCheckpoint, dropNotebookVectorTable } from './index'
 import { chatSessions, chatMessages, notebooks, notes, documents, items } from './schema'
+import type { ChatMessageMetadata } from '../../shared/types/chat'
 
 // ==================== Chat Sessions ====================
 
@@ -205,6 +206,18 @@ export function updateMessageContent(
   db.update(chatMessages).set(updateData).where(eq(chatMessages.id, messageId)).run()
 }
 
+/**
+ * 更新消息的结构化元数据。
+ *
+ * 目前唯一的用途是把「这条回答基于哪些段落」写到助手消息上，
+ * 这样回答交付之后仍然可以回到原文（见 shared/types/chat.ts 的 AnswerSource）。
+ */
+export function updateMessageMetadata(messageId: string, metadata: ChatMessageMetadata) {
+  const db = getDatabase()
+
+  db.update(chatMessages).set({ metadata }).where(eq(chatMessages.id, messageId)).run()
+}
+
 // ==================== Notebooks ====================
 
 /**

src/main/ipc/chatHandlers.ts

@@ -2,35 +2,50 @@ import { ipcMain, IpcMainInvokeEvent } from 'electron'
 import * as queries from '../db/queries'
 import { ConnectionManager } from '../models/ConnectionManager'
 import { SessionAutoSwitchService } from '../services/SessionAutoSwitchService'
-import { KnowledgeService } from '../services/KnowledgeService'
+import { KnowledgeService, type SearchResult } from '../services/KnowledgeService'
 import { validateAndCleanMessages } from '../utils/messageValidator'
 import Logger from '../../shared/utils/logger'
+import type { AnswerSource, RetrievalStatus } from '../../shared/types/chat'
 import { ChatSchemas, validate } from './validation'
 
 // 管理活跃的流式请求
 const activeStreams = new Map<string, AbortController>()
 
 /**
- * 构建 RAG 上下文 prompt
+ * 构建 RAG 上下文 prompt，并把「这段回答基于哪些段落」一起交出来。
+ *
+ * 之前这里只取 `documentTitle` / `content` / `score` 三个字段，
+ * `chunkId`、`documentId`、`chunkIndex` 全部被丢掉 —— 于是回答交付之后，
+ * 界面上再也没有回到原文的路。prompt 文本保持不变，这里只是不再丢弃身份。
  */
-function buildRAGContext(
-  searchResults: Array<{
-    documentTitle: string
-    content: string
-    score: number
-  }>
-): string {
-  if (searchResults.length === 0) return ''
-
-  const contextParts = searchResults.map((result, index) => {
-    return `[来源 ${index + 1}: ${result.documentTitle}]\n${result.content}`
-  })
+function buildRAGContext(searchResults: SearchResult[]): {
+  context: string
+  sources: AnswerSource[]
+} {
+  if (searchResults.length === 0) return { context: '', sources: [] }
+
+  const sources: AnswerSource[] = searchResults.map((result, index) => ({
+    index: index + 1,
+    documentId: result.documentId,
+    documentTitle: result.documentTitle,
+    documentType: result.documentType,
+    chunkId: result.chunkId,
+    chunkIndex: result.chunkIndex,
+    content: result.content,
+    score: result.score
+  }))
+
+  const contextParts = sources.map(
+    (source) => `[来源 ${source.index}: ${source.documentTitle}]\n${source.content}`
+  )
 
-  return `以下是与用户问题相关的背景知识，请参考这些信息来回答：
+  const context = `以下是与用户问题相关的背景知识，请参考这些信息来回答：
 
 ${contextParts.join('\n\n---\n\n')}
 
 请基于以上背景知识回答用户的问题。如果背景知识不足以回答问题，请说明并尽力提供有帮助的回答。`
+
+  return { context, sources }
 }
 
 /**
@@ -125,7 +140,11 @@ export function registerChatHandlers(
     }
 
     // 3.2 RAG 增强：检索相关知识并注入上下文
-    // 只有在配置了 embedding connection 时才启用 RAG
+    // 只有在配置了 embedding connection 时才启用 RAG。
+    // 检索结果同时记录到消息上：以前检索失败只留一行日志，
+    // 于是「没有依据的回答」和「有依据的回答」在界面上完全无法区分。
+    let retrieval: RetrievalStatus = 'none'
+    let answerSources: AnswerSource[] = []
     try {
       const embeddingClient = await connectionManager.getEmbeddingClient()
 
@@ -138,7 +157,9 @@ export function registerChatHandlers(
           })
 
           if (searchResults.length > 0) {
-            const ragContext = buildRAGContext(searchResults)
+            const { context, sources } = buildRAGContext(searchResults)
+            retrieval = 'used'
+            answerSources = sources
             Logger.debug(
               'ChatHandlers',
               `RAG: Found ${searchResults.length} relevant chunks for query`
@@ -147,18 +168,25 @@ export function registerChatHandlers(
             // 将 RAG 上下文作为 system message 插入到消息列表开头
             messages.unshift({
               role: 'system',
-              content: ragContext
+              content: context
             })
           }
         }
       } else {
         Logger.debug('ChatHandlers', 'RAG disabled: No embedding model configured')
       }
     } catch (error) {
-      // RAG 失败不应该阻止对话，只记录警告
+      // RAG 失败不应该阻止对话
+      retrieval = 'failed'
       Logger.warn('ChatHandlers', 'RAG search failed:', error)
     }
 
+    queries.updateMessageMetadata(assistantMessage.id, {
+      ...(assistantMessage.metadata ?? {}),
+      retrieval,
+      sources: answerSources
+    })
+
     // 4. 调用 Model Connection 流式生成
     const client = await connectionManager.getChatClient()
     if (!client) {

src/renderer/src/components/notebook/ProcessPanel.tsx

@@ -312,14 +312,8 @@ function ProcessPanel({
             value={input}
             onChange={handleInputChange}
             onKeyDown={handleKeyDown}
-            placeholder={
-              !currentSession
-                ? t('selectSession')
-                : !hasChatModel
-                  ? t('noProviderConfigured')
-                  : t('inputMessage')
-            }
-            disabled={!currentSession || !hasChatModel}
+            placeholder={!hasChatModel ? t('noProviderConfigured') : t('inputMessage')}
+            disabled={!hasChatModel}
             rows={1}
             className="w-full bg-transparent border-0 pl-4 pr-14 py-3 text-sm text-foreground placeholder-muted-foreground resize-none focus-visible:ring-0 focus-visible:ring-offset-0 overflow-y-auto min-h-[84px] max-h-[280px] themed-scrollbar select-text"
           />

src/renderer/src/components/notebook/SourcePanel.tsx

@@ -304,6 +304,20 @@ export default function SourcePanel(): ReactElement {
     setSelectedDocument(null)
   }
 
+  // 对话里的来源被点击时，在这里把对应文档打开。
+  // 右栏的对话和左栏的知识库是兄弟节点，所以这个请求走 uiStore 而不是 props。
+  //
+  // 纯派生，不用 effect：在 effect 里同步 setState 会引发级联渲染（本文件上面
+  // 那个 notebook 切换的复位用的是同一套“render 期间运算”的思路），而在 render
+  // 里写 store 会更糟。列表里的点击会把 store 里这个请求清掉，所以两者不会打架。
+  const focusedSourceDocumentId = useUIStore((state) => state.focusedSourceDocumentId)
+  const focusSourceDocument = useUIStore((state) => state.focusSourceDocument)
+
+  const focusedDocument = focusedSourceDocumentId
+    ? (documents.find((doc) => doc.id === focusedSourceDocumentId) ?? null)
+    : null
+  const openDocument = selectedDocument ?? focusedDocument
+
   // 处理文件上传
   const handleFileUpload = useCallback(async () => {
     if (!notebookId) return
@@ -419,18 +433,21 @@ export default function SourcePanel(): ReactElement {
       // 文本和笔记类型可以预览，直接显示预览页面
       if (document.type === 'text' || document.type === 'note') {
         setSelectedDocument(document)
+        // 列表里的选择优先，清掉对话那边可能还挂着的请求
+        focusSourceDocument(null)
       } else {
         // 其他类型（文件、URL）直接打开
         handleOpenSource(document.id)
       }
     },
-    [handleOpenSource]
+    [handleOpenSource, focusSourceDocument]
   )
 
   // 返回列表
   const handleBack = useCallback(() => {
     setSelectedDocument(null)
-  }, [])
+    focusSourceDocument(null)
+  }, [focusSourceDocument])
 
   // 处理弹窗提交
   const handleModalSubmit = useCallback(
@@ -448,13 +465,9 @@ export default function SourcePanel(): ReactElement {
 
   return (
     <Card className="flex h-full flex-col overflow-hidden">
-      {selectedDocument ? (
+      {openDocument ? (
         // 文档预览页面
-        <DocumentViewerPanel
-          key={selectedDocument.id}
-          document={selectedDocument}
-          onBack={handleBack}
-        />
+        <DocumentViewerPanel key={openDocument.id} document={openDocument} onBack={handleBack} />
       ) : (
         // 文档列表页面
         <>