document-extraction

文档提取技能

从现有文档源中提取需求，以便系统地挖掘需求。

何时使用此技能

关键词： 提取需求、文档挖掘、PDF 需求、转录分析、解析文档、现有文档、遗留需求、竞品分析

在以下情况调用此技能：

• 从现有文档中挖掘需求
• 处理会议转录内容以提取需求
• 从竞品中提取需求
• 分析监管文件以提取合规要求
• 将遗留文档转换为结构化需求

支持的文档类型

类型	扩展名	提取方法
Markdown	.md	直接读取
文本	.txt	直接读取
PDF	.pdf	Read 工具（支持 PDF）
Word	.docx	Read 工具
网页	URL	WebFetch 工具
会议记录	.md, .txt	转录模式
规范	.md, .docx	需求模式

提取工作流

第 1 步：文档评估

分析文档以确定提取策略：

document_assessment:
  path: "{file path or URL}"
  type: "{detected document type}"
  size: "{approximate size}"
  structure:
    has_sections: true|false
    has_lists: true|false
    has_tables: true|false
  quality:
    formal_language: true|false
    clear_requirements: true|false
    needs_interpretation: true|false

第 2 步：模式匹配

应用需求检测模式：

显式需求标记：

- "The system shall..."
- "The system must..."
- "Users should be able to..."
- "REQ-XXX:"
- Numbered requirements (1.1, 1.2, etc.)

EARS 模式：

- "When [trigger], the [system] shall [response]"
- "While [state], the [system] shall [behavior]"
- "Where [feature], the [system] shall [behavior]"
- "If [condition], then the [system] shall [response]"

隐式需求指示词：

- "It is important that..."
- "We need..."
- "The goal is to..."
- "Users expect..."
- "Performance should..."

第 3 步：需求提取

对于每项已识别的需求：

extracted_requirement:
  id: REQ-{sequence}
  text: "{cleaned requirement statement}"
  source: document
  source_file: "{file path}"
  source_location: "{section/page/line}"
  original_text: "{exact text from document}"
  type: functional|non-functional|constraint|assumption
  confidence: high|medium|low
  extraction_method: explicit|pattern|inferred
  needs_review: true|false
  review_notes: "{why review needed}"

第 4 步：分类

对提取的需求进行分类：

categories:
  functional:
    - features
    - behaviors
    - interactions
  non_functional:
    - performance
    - security
    - usability
    - reliability
    - scalability
  constraints:
    - technical
    - business
    - regulatory
  assumptions:
    - environmental
    - user_behavior
    - dependencies

第 5 步：去重

识别并合并重复的需求：

deduplication:
  strategy: semantic_similarity
  threshold: 0.8
  action: merge|flag_for_review
  merged_requirements:
    - id: REQ-merged-001
      sources: [REQ-001, REQ-015]
      text: "{consolidated requirement}"

特定文档策略

会议转录

transcript_extraction:
  focus_on:
    - Action items
    - Decisions made
    - Requirements discussed
    - Concerns raised
  patterns:
    - "We decided that..."
    - "The requirement is..."
    - "Action item:"
    - "TODO:"
    - "Need to..."
  speaker_context:
    - Note who said what
    - Weight by speaker role

监管文件

regulatory_extraction:
  focus_on:
    - Mandatory requirements ("shall", "must")
    - Prohibited actions ("shall not", "must not")
    - Conditional requirements ("if...then")
  compliance_mapping:
    - Reference section numbers
    - Note effective dates
    - Track version/revision

竞品分析

competitor_extraction:
  focus_on:
    - Feature descriptions
    - User capabilities
    - Unique selling points
  output:
    - Feature requirements
    - Differentiation opportunities
    - Gap identification
  confidence: low  # Based on external observation

遗留规范

legacy_extraction:
  focus_on:
    - Existing requirements
    - System behaviors
    - Integration points
  modernization:
    - Update terminology
    - Convert to EARS format
    - Flag deprecated requirements

输出格式

单文档输出

extraction_result:
  source:
    file: "{path or URL}"
    type: "{document type}"
    extraction_date: "{ISO-8601}"
    confidence: high|medium|low

  statistics:
    total_candidates: {number}
    extracted: {number}
    filtered: {number}
    needs_review: {number}

  requirements:
    - id: REQ-{number}
      text: "{requirement}"
      type: functional|non-functional|constraint
      source_location: "{section/page}"
      confidence: high|medium|low
      original_text: "{exact source text}"

  review_items:
    - requirement_id: REQ-{number}
      reason: "{why review needed}"
      suggestion: "{proposed action}"

  metadata:
    sections_processed: {number}
    extraction_patterns_used: ["{pattern names}"]

自主性级别

引导模式

guided_behavior:
  document_selection: Human selects
  extraction_strategy: AI suggests, human approves
  each_requirement: AI highlights, human confirms
  categorization: AI suggests, human validates

半自主模式

semi_auto_behavior:
  document_selection: AI suggests priority, human approves list
  extraction_strategy: AI chooses autonomously
  requirements: AI extracts all, human reviews in batches
  categorization: AI categorizes, human spot-checks

全自主模式

full_auto_behavior:
  document_selection: AI processes all relevant
  extraction_strategy: AI optimizes per document
  requirements: AI extracts, deduplicates, categorizes
  output: Full extraction report for final review

质量指标

高置信度提取

• 显式需求标记（“shall”“must”）
• 匹配 EARS 模式
• 带编号的需求列表
• 清晰的祈使句陈述

中置信度提取

• 隐式指示词（“should”“needs to”）
• 依赖上下文的解释
• 部分匹配模式
• 需要领域知识

低置信度提取

• 根据描述推断
• 解释叙述文本
• 竞品分析
• 基于上下文的假设

委派

对于相关任务，委派给：

• gap-analysis：检查提取的需求是否完整
• domain-research：研究不熟悉的术语或概念
• elicitation-methodology：返回以选择技术

输出位置

将提取结果保存到：

.requirements/{domain}/documents/DOC-{filename}-{timestamp}.yaml

相关内容

• elicitation-methodology - 父级中心技能
• gap-analysis - 提取后的完整性检查
• interview-conducting - 与利益相关者澄清提取的需求