fix(chat): bound PDF context with retrieved chunks

apps/electron/src/main/lib/chat-service.ts

@@ -26,6 +26,7 @@ import { listChannels, resolveChannelRuntimeApiKey } from './channel-manager'
 import { appendMessage, updateConversationMeta, getConversationMessages } from './conversation-manager'
 import { readAttachmentAsBase64, isImageAttachment } from './attachment-service'
 import { extractTextFromAttachment, isDocumentAttachment } from './document-parser'
+import { estimatePromptTokens, retrieveDocumentContext } from './document-context'
 import { getFetchFn } from './proxy-fetch'
 import { getEffectiveProxyUrl } from './proxy-settings-service'
 import { getEnabledTools } from './chat-tool-registry'
@@ -85,7 +86,8 @@ async function enrichMessageWithDocuments(
     try {
       const text = await extractTextFromAttachment(att.localPath)
       if (text.trim()) {
-        parts.push(`\n<file name="${att.filename}">\n${text}\n</file>`)
+        const context = retrieveDocumentContext(text, messageText)
+        parts.push(`\n<file name="${att.filename}" mode="retrieved-chunks">\n${context}\n</file>`)
       } else {
         parts.push(`\n<file name="${att.filename}">\n[文件内容为空]\n</file>`)
       }
@@ -102,7 +104,7 @@ async function enrichMessageWithDocuments(
 /**
  * 为历史消息列表注入文档附件文本
  *
- * 遍历历史消息，对包含文档附件的用户消息进行文本增强。
+ * 历史附件只保留引用，禁止每一轮重新注入整篇文档。
  * 返回新的消息数组（不修改原始消息）。
  */
 async function enrichHistoryWithDocuments(
@@ -115,8 +117,8 @@ async function enrichHistoryWithDocuments(
     if (msg.role === 'user' && msg.attachments && msg.attachments.length > 0) {
       const hasDocuments = msg.attachments.some((att) => isDocumentAttachment(att.mediaType))
       if (hasDocuments) {
-        const enrichedContent = await enrichMessageWithDocuments(msg.content, msg.attachments)
-        enriched.push({ ...msg, content: enrichedContent })
+        const names = msg.attachments.filter((att) => isDocumentAttachment(att.mediaType)).map((att) => att.filename)
+        enriched.push({ ...msg, content: `${msg.content}\n[历史附件引用：${names.join('、')}]` })
         continue
       }
     }
@@ -251,6 +253,19 @@ export async function sendMessage(
   const filteredHistory = filterHistory(fullHistory, contextDividers, contextLength)
   const enrichedHistory = await enrichHistoryWithDocuments(filteredHistory)
   const enrichedUserMessage = await enrichMessageWithDocuments(userMessage, attachments)
+  const estimatedPromptTokens = estimatePromptTokens([
+    systemMessage ?? '',
+    ...enrichedHistory.map((message) => message.content),
+    enrichedUserMessage,
+  ])
+  // 留出模型回复和工具调用空间。超限必须由客户端明确提示，不能让 A2 静默截断。
+  if (estimatedPromptTokens > 180_000) {
+    webContents.send(CHAT_IPC_CHANNELS.STREAM_ERROR, {
+      conversationId,
+      error: `当前对话预计 ${estimatedPromptTokens.toLocaleString()} tokens，超过 Kiro 安全输入上限 180,000。请新建会话、缩小 PDF 范围或使用文档摘要。`,
+    })
+    return
+  }
 
   // 6. 创建 AbortController
   const controller = new AbortController()

apps/electron/src/main/lib/document-context.test.ts

@@ -0,0 +1,15 @@
+import { describe, expect, test } from 'bun:test'
+import { chunkDocument, DOCUMENT_CONTEXT_MAX_TOKENS, estimatePromptTokens, retrieveDocumentContext } from './document-context'
+
+describe('文档上下文预算', () => {
+  test('长文档按重叠片段切分并只检索相关内容', () => {
+    const text = `${'普通内容 '.repeat(2000)}关键结论：财政政策有效。${'尾部 '.repeat(2000)}`
+    expect(chunkDocument(text).length).toBeGreaterThan(1)
+    expect(retrieveDocumentContext(text, '财政政策')).toContain('财政政策有效')
+  })
+
+  test('注入内容永远不超过六万 token 预算', () => {
+    const selected = retrieveDocumentContext('研究资料 '.repeat(200000), '研究资料')
+    expect(estimatePromptTokens([selected])).toBeLessThanOrEqual(DOCUMENT_CONTEXT_MAX_TOKENS)
+  })
+})

apps/electron/src/main/lib/document-context.ts

@@ -0,0 +1,54 @@
+import { createHash } from 'node:crypto'
+
+const CHUNK_CHARS = 4_000
+const CHUNK_OVERLAP_CHARS = 600
+export const DOCUMENT_CONTEXT_MAX_TOKENS = 60_000
+// 预留约 2% 给片段编号、XML 包装及 tokenizer 误差。
+const DOCUMENT_CONTEXT_MAX_CHARS = Math.floor(DOCUMENT_CONTEXT_MAX_TOKENS * 4 * 0.98)
+
+const chunkCache = new Map<string, string[]>()
+
+function documentKey(text: string): string {
+  return createHash('sha256').update(text).digest('hex')
+}
+
+export function chunkDocument(text: string): string[] {
+  const key = documentKey(text)
+  const cached = chunkCache.get(key)
+  if (cached) return cached
+  const normalized = text.replace(/\r\n/g, '\n').trim()
+  const chunks: string[] = []
+  for (let start = 0; start < normalized.length; start += CHUNK_CHARS - CHUNK_OVERLAP_CHARS) {
+    chunks.push(normalized.slice(start, start + CHUNK_CHARS))
+  }
+  chunkCache.set(key, chunks)
+  return chunks
+}
+
+function queryTerms(query: string): Set<string> {
+  return new Set((query.toLowerCase().match(/[\p{L}\p{N}_-]{2,}/gu) ?? []).slice(0, 200))
+}
+
+export function retrieveDocumentContext(text: string, query: string): string {
+  const terms = queryTerms(query)
+  const ranked = chunkDocument(text).map((chunk, index) => {
+    const lower = chunk.toLowerCase()
+    let score = 0
+    for (const term of terms) if (lower.includes(term)) score++
+    return { chunk, index, score }
+  }).sort((a, b) => b.score - a.score || a.index - b.index)
+
+  const selected: string[] = []
+  let chars = 0
+  for (const item of ranked) {
+    if (chars + item.chunk.length > DOCUMENT_CONTEXT_MAX_CHARS) continue
+    selected.push(`[片段 ${item.index + 1}]\n${item.chunk}`)
+    chars += item.chunk.length
+    if (chars >= DOCUMENT_CONTEXT_MAX_CHARS) break
+  }
+  return selected.join('\n\n')
+}
+
+export function estimatePromptTokens(parts: string[]): number {
+  return Math.ceil(parts.reduce((sum, value) => sum + value.length, 0) / 4)
+}