feat: enhance local knowledge retrieval

src/main/knowledge/document-parser.test.ts

@@ -1,6 +1,10 @@
 import { strToU8, zipSync } from 'fflate'
 import { describe, expect, it } from 'vitest'
-import { chunkDocument, parseDocument } from './document-parser'
+import {
+  buildChunkContextPrefix,
+  chunkDocumentAdvanced,
+  parseDocument
+} from './document-parser'
 
 function createPdfFixture(text: string): Buffer {
   const stream = `BT /F1 18 Tf 50 100 Td (${text}) Tj ET`
@@ -35,7 +39,15 @@ describe('document parser', () => {
       'notes.md',
       Buffer.from(`# GoodBuddy\n\n${'知识内容。'.repeat(500)}`)
     )
-    const chunks = chunkDocument(parsed, 500, 50)
+    const chunks = chunkDocumentAdvanced(parsed, {
+      version: 1,
+      mode: 'fixed',
+      targetCharacters: 500,
+      overlapCharacters: 50,
+      parentCharacters: 4_800,
+      childCharacters: 900,
+      contextualIndexingEnabled: false
+    })
 
     expect(parsed.title).toBe('notes')
     expect(chunks.length).toBeGreaterThan(1)
@@ -103,9 +115,12 @@ describe('document parser', () => {
     expect(parsed.sections).toEqual([
       {
         locator: '第 1 页',
-        content: 'PDF body text'
+        content: 'PDF body text',
+        pageNumber: 1,
+        blockKind: 'text'
       }
     ])
+    expect(parsed.pageCount).toBe(1)
   })
 
   it('rejects unsupported or oversized content', async () => {
@@ -122,4 +137,162 @@ describe('document parser', () => {
       parseDocument('expanded.docx', Buffer.from(expandedArchive))
     ).rejects.toThrow('损坏')
   })
+
+  it('preserves headings and creates recall-only children with parent context', () => {
+    const chunks = chunkDocumentAdvanced(
+      {
+        title: 'Guide',
+        sourceFormat: '.md',
+        content: '# 安装\n' + '安装步骤和配置说明。'.repeat(250),
+        sections: [
+          {
+            locator: '全文',
+            content: '# 安装\n' + '安装步骤和配置说明。'.repeat(250)
+          }
+        ],
+        warnings: []
+      },
+      {
+        version: 1,
+        mode: 'parent-child',
+        targetCharacters: 1_600,
+        overlapCharacters: 100,
+        parentCharacters: 1_600,
+        childCharacters: 400,
+        contextualIndexingEnabled: false
+      }
+    )
+    const parents = chunks.filter((chunk) => chunk.role === 'parent')
+    const children = chunks.filter((chunk) => chunk.role === 'child')
+    expect(parents.length).toBeGreaterThan(0)
+    expect(children.length).toBeGreaterThan(parents.length)
+    expect(children.every((chunk) => chunk.heading === '安装')).toBe(true)
+    expect(
+      children.every((chunk) =>
+        parents.some((parent) => parent.position === chunk.parentPosition)
+      )
+    ).toBe(true)
+  })
+
+  it('tracks nested Markdown heading paths and resets deeper levels', () => {
+    const chunks = chunkDocumentAdvanced(
+      {
+        title: 'Guide',
+        sourceFormat: '.md',
+        content: '# A\none\n## B\ntwo\n### C\nthree\n## D\nfour\n# E\nfive',
+        sections: [
+          {
+            locator: '全文',
+            content:
+              '# A\none\n## B\ntwo\n### C\nthree\n## D\nfour\n# E\nfive'
+          }
+        ],
+        warnings: []
+      },
+      {
+        version: 1,
+        mode: 'structure',
+        targetCharacters: 500,
+        overlapCharacters: 0,
+        parentCharacters: 1_000,
+        childCharacters: 300,
+        contextualIndexingEnabled: false
+      }
+    )
+
+    expect(chunks.map((chunk) => chunk.headingPath)).toEqual([
+      ['A'],
+      ['A', 'B'],
+      ['A', 'B', 'C'],
+      ['A', 'D'],
+      ['E']
+    ])
+    expect(chunks.map((chunk) => chunk.heading)).toEqual([
+      'A',
+      'B',
+      'C',
+      'D',
+      'E'
+    ])
+    expect(chunks[1]?.content).toBe('## B\ntwo')
+  })
+
+  it('propagates page and table metadata without crossing section boundaries', () => {
+    const chunks = chunkDocumentAdvanced(
+      {
+        title: 'Workbook',
+        sourceFormat: '.xlsx',
+        content: '第一页\n\n表格行',
+        sections: [
+          {
+            locator: '第 1 页',
+            content: '第一页',
+            pageNumber: 1,
+            blockKind: 'text'
+          },
+          {
+            locator: '工作表 1',
+            content: '表格行'.repeat(200),
+            blockKind: 'table'
+          }
+        ],
+        warnings: []
+      },
+      {
+        version: 1,
+        mode: 'parent-child',
+        targetCharacters: 300,
+        overlapCharacters: 20,
+        parentCharacters: 300,
+        childCharacters: 100,
+        contextualIndexingEnabled: false
+      }
+    )
+
+    expect(
+      chunks
+        .filter((chunk) => chunk.locator === '第 1 页')
+        .every(
+          (chunk) =>
+            chunk.pageNumber === 1 && chunk.blockKind === 'text'
+        )
+    ).toBe(true)
+    expect(
+      chunks
+        .filter((chunk) => chunk.locator === '工作表 1')
+        .every(
+          (chunk) =>
+            chunk.pageNumber === undefined && chunk.blockKind === 'table'
+        )
+    ).toBe(true)
+    expect(
+      chunks.every((chunk) =>
+        chunk.locator === '第 1 页'
+          ? chunk.content.includes('第一页')
+          : !chunk.content.includes('第一页')
+      )
+    ).toBe(true)
+  })
+
+  it('builds deterministic bounded context without changing citation content', () => {
+    const chunk = {
+      position: 0,
+      locator: ' 第 2 页 \n 附录 ',
+      content: '## API\n原始引用内容',
+      headingPath: [' 指南 ', 'API'],
+      pageNumber: 2,
+      blockKind: 'table' as const
+    }
+    const originalContent = chunk.content
+    const first = buildChunkContextPrefix(' GoodBuddy \n 手册 ', chunk)
+    const second = buildChunkContextPrefix(' GoodBuddy \n 手册 ', chunk)
+
+    expect(first).toBe(second)
+    expect(first).toBe(
+      '[context title="GoodBuddy 手册" heading="指南 > API" page="2" locator="第 2 页 附录" block="table"]\n'
+    )
+    expect(first.length).toBeLessThanOrEqual(512)
+    expect(chunk.content).toBe(originalContent)
+    expect(first).not.toContain(originalContent)
+  })
 })