document-summarizer

name	document-summarizer
description	Summarize and structure raw document text (PDF/Word) into a format optimized for PPT generation. Use when the user uploads a document and wants to generate a presentation from it.

Document Summarizer for PPT Generation

Purpose

Transform raw extracted text from PDF or Word documents into a structured summary that feeds directly into the PPT generation pipeline. The output is designed so that each section maps naturally to one or more PPT slides, preserving document hierarchy, extracting key facts and data tables, and providing actionable hints for slide planning.

When to Use

• User uploads a PDF or Word document and wants to generate a PPT from it
• Need to understand document structure before creating a PPT outline
• This step runs before the PPT planning stage — its output becomes the input context for the PPT outline planner

How It Integrates With the PPT Pipeline

PDF/Word → extract text → this summarizer → structured_summary
                                                    ↓
                                          PPT outline planner
                                                    ↓
                                          逐页生成 + QA (PPT runtime)

The sections array maps to PPT slide groups (each section typically covers 1-3 slides). The ppt_generation_hints directly seeds parameters for the downstream PPT planner: audience, style_preference, and suggested_total_slides become planner inputs.

Input

Raw text extracted from PDF or Word document, optionally with:

• Extracted tables (list of 2D arrays from pdfplumber)
• Document metadata ({ title, author, date, page_count })
• User's stated goal, target audience, or style preference

If the document has a clear hierarchical structure (chapters → sections → subsections), preserve it. If the document is narrative or prose-heavy, infer logical breaking points.

Output Format

Only output JSON — no markdown code fences, no explanations, no preamble.

Schema

{
  "document_title": "string (required) — original title or inferred",
  "document_summary": "string (required) — 2-3 sentence overview of the document's purpose and scope",
  "sections": [
    {
      "section_title": "string (required)",
      "section_summary": "string (required) — 1-3 sentences, what this section covers and why it matters",
      "key_points": [
        "string — complete, standalone fact or idea. No vague language, no references to other sections."
      ],
      "tables": [
        {
          "description": "string (required) — what this table shows and why it is important",
          "headers": ["string"],
          "rows": [["string"]]
        }
      ],
      "important_data": [
        "string — specific number, statistic, date, or ranking. No invented data."
      ],
      "suggested_slide_count": "integer (required) — 1 for brief sections, 2-3 for substantive ones",
      "narrative_transition": "string (optional) — one sentence explaining how this section connects to the next"
    }
  ],
  "metadata": {
    "source_file": "string (optional)",
    "total_pages": "integer (optional)",
    "language": "string — zh-CN | en-US | auto-detect",
    "estimated_reading_time_minutes": "integer (optional)"
  },
  "ppt_generation_hints": {
    "suggested_total_slides": "integer — min 4, max 20, target 6-12",
    "audience": "string — general | student | technical | business | academic",
    "style_preference": "string — informative | executive | creative | academic",
    "recommended_visual_elements": ["string — diagrams | statistics_callouts | timeline | comparison | process_flow"],
    "content_focus": "string — knowledge_sharing | proposal | report | training | introduction",
    "special_requirements": "string (optional) — any explicit user request or constraint noted in the document"
  }
}

Required vs Optional Fields

Field	Required
document_title	Yes
document_summary	Yes
sections (non-empty)	Yes
sections[].section_title	Yes
sections[].section_summary	Yes
sections[].key_points (non-empty)	Yes
sections[].suggested_slide_count	Yes
sections[].tables	No (omit if none)
sections[].important_data	No (omit if none)
sections[].narrative_transition	No
metadata.*	No
ppt_generation_hints.*	No (but recommended)

Rules

1. Map Document Structure to Sections

• Each major chapter/heading in the source becomes a section
• Sub-sections under the same heading can be merged if they share a similar theme
• Never skip a major section — every significant heading must appear
• If the document has no clear headings (prose-heavy), infer logical breaks: introduction → core argument → evidence/examples → conclusion
• For long documents (> 30k characters): mark core sections distinctly, group minor supplementary sections together, and note this in special_requirements

2. Write Key Points That Stand Alone

Each entry in key_points must be:

• Self-contained: understandable without reading surrounding text or other sections
• Specific: include concrete details (names, numbers, dates, comparisons), not generic descriptions
• Concise: 1-2 sentences maximum
• Distinct: never repeat the same fact with different wording

Never include:

• Vague phrases ("various factors", "etc.", "many aspects")
• Introductory/conclusion fillers ("In summary...", "To conclude...")
• References forward or backward ("As discussed earlier...", "This will be covered later...")
• Multi-part ideas split across bullets

3. Extract Tables Selectively

Include a table only if it contains data that benefits from visual presentation:

• Comparative data (features, specs, prices, metrics across items)
• Process/role/architecture information (components, steps, responsibilities)
• Timeline or ranking data

Skip tables with < 2 columns or < 2 rows. For tables > 6 rows, keep first 5 and append rows: [["..."]].

Every table must have a description explaining what it shows and why it matters for the PPT.

4. Extract Important Data Precisely

• Only numbers, statistics, dates, percentages, rankings explicitly stated in the document
• Do not calculate or infer — if the document says "increased by 20%", include exactly that
• If no significant data exists in a section, omit important_data entirely
• Useful for stat callout slides: one strong number can anchor an entire page

5. Estimate Slide Count Honestly

suggested_slide_count per section:

• 1: pure overview, introduction, or very brief (under 200 words of content)
• 2: standard section with 3-5 key points and moderate detail
• 3+: complex section with multiple sub-topics, heavy data, or a diagram worth its own slide

suggested_total_slides (top-level):

• Range: 4 to 20, target 6-12 for most documents
• Formula: sum of section slide counts + 2 (cover + closing)
• If the sum exceeds 20, prioritize the most substantive sections and merge or drop minor ones

6. Maintain Content Integrity

• Never fabricate — every key_point and summary must be derivable from the source text
• If a section is too thin to extract meaningful points, write a brief descriptive note instead of inventing content
• Preserve technical terminology accurately — do not simplify technical terms for general audiences unless instructed
• Keep language consistent — all text in the same language as the source document

7. Guide PPT Visual Strategy

ppt_generation_hints.recommended_visual_elements should match what the document actually contains:

• statistics_callouts — if important_data has strong numbers
• diagrams — if the document describes processes, structures, or relationships
• timeline — if chronological data or milestones are present
• comparison — if tables contain comparative data
• process_flow — if steps or workflows are described
• map — if geographic data is present

content_focus should reflect the document's primary purpose:

• knowledge_sharing — training materials, educational content
• proposal — project or business proposals
• report — status reports, research summaries
• training — onboarding, procedure manuals
• introduction — topic overviews, company/product introductions

Complete Example

Input (abbreviated source text)

Chapter 1: 大模型技术概述 大语言模型（Large Language Model, LLM）是基于深度学习的自然语言处理技术... 2020年GPT-3发布后，LLM进入爆发期，参数量从亿级增长到万亿级...

Chapter 2: 核心技术架构 Transformer架构是LLM的基础，由Vaswani等人于2017年提出... 核心组件包括：自注意力机制（Self-Attention）、前馈神经网络（FFN）、位置编码...

Chapter 3: 训练与优化 预训练阶段使用海量无标注文本进行自监督学习...

Output JSON

{
  "document_title": "大模型技术概述",
  "document_summary": "本文件系统性介绍大语言模型（LLM）的发展历程、核心技术架构与训练方法。内容涵盖从GPT-3到当前主流模型的演进、Transformer架构的关键组件，以及预训练与微调的技术细节。",
  "sections": [
    {
      "section_title": "大模型技术概述",
      "section_summary": "本节介绍LLM的定义及其发展脉络。GPT-3于2020年发布标志着LLM进入爆发期，参数量从亿级跃升至万亿级，引发学术界与产业界的广泛关注。",
      "key_points": [
        "大语言模型（LLM）是基于深度学习的自然语言处理技术，通过海量文本数据学习语言规律与知识表示",
        "2020年GPT-3发布后LLM进入爆发期，参数量从亿级增长到万亿级，模型能力出现质的飞跃"
      ],
      "tables": [],
      "important_data": ["参数量：亿级 → 万亿级", "标志性节点：2020年 GPT-3发布"],
      "suggested_slide_count": 1,
      "narrative_transition": "理解了LLM的基本概念后，接下来深入其底层技术基础——Transformer架构"
    },
    {
      "section_title": "核心技术架构",
      "section_summary": "Transformer架构由Vaswani等人于2017年提出，是所有主流LLM的共同基础。其核心在于自注意力机制，使模型能够并行处理序列中的任意位置信息，大幅提升训练效率。",
      "key_points": [
        "Transformer架构由Vaswani等人于2017年提出，是GPT及后续所有LLM的共同基础",
        "自注意力机制（Self-Attention）允许模型并行处理任意距离的token，突破RNN的序列限制",
        "前馈神经网络（FFN）与残差连接构成Transformer的基本计算单元",
        "位置编码（Positional Encoding）为序列中的词序信息提供可学习的表示"
      ],
      "tables": [
        {
          "description": "Transformer的三大核心组件及其作用",
          "headers": ["组件", "英文名", "核心作用"],
          "rows": [
            ["自注意力", "Self-Attention", "建模token间的依赖关系"],
            ["前馈网络", "FFN", "逐位非线性变换"],
            ["位置编码", "Positional Encoding", "注入序列位置信息"]
          ]
        }
      ],
      "important_data": ["架构提出年份：2017年"],
      "suggested_slide_count": 2,
      "narrative_transition": "掌握架构原理后，下面来看模型是如何通过大规模训练获得能力的"
    },
    {
      "section_title": "训练与优化",
      "section_summary": "LLM的训练分为预训练和微调两个阶段。预训练使用海量无标注文本进行自监督学习，微调则通过少量标注数据使模型适应特定任务。",
      "key_points": [
        "预训练阶段：使用海量无标注文本进行自监督学习，目标是最小化语言建模损失",
        "微调阶段：通过少量标注数据调整模型参数，使模型适应特定下游任务",
        "主流微调方法包括：Prompt Tuning、LoRA等参数高效微调技术，大幅降低微调成本"
      ],
      "tables": [],
      "important_data": [],
      "suggested_slide_count": 1
    }
  ],
  "metadata": {
    "language": "zh-CN",
    "estimated_reading_time_minutes": 15
  },
  "ppt_generation_hints": {
    "suggested_total_slides": 7,
    "audience": "technical",
    "style_preference": "informative",
    "recommended_visual_elements": ["diagrams", "statistics_callouts"],
    "content_focus": "knowledge_sharing",
    "special_requirements": "第二章技术架构内容较深，建议为自注意力机制单独配一张示意图"
  }
}

Constraints

1. Output only JSON — no fences, no preamble, no commentary
2. JSON must be valid — proper escaping, no trailing commas, arrays/objects closed correctly
3. No truncation — every section gets its full key_points, no "..."
4. No fabrication — all content traceable to source text
5. Fields must be meaningful — null, empty strings, and placeholder text are prohibited in required fields