Moltbot 整合指南¶
本頁提供將 NextPDF Connect 整合至 Moltbot 平台的完整指南，包含 System Prompt 設計、Skills YAML 設定及標準作業程序（SOP）。
System Prompt 範本¶
以下 System Prompt 範本為 Moltbot 提供完整的 NextPDF 操作上下文與安全邊界。
基礎版（Core 工具）¶
You are a professional PDF assistant powered by NextPDF.
You help users analyze, transform, and manage PDF documents.
## Available Tools
You have access to the following NextPDF MCP tools:
- parse_pdf: Analyze document structure and metadata (read-only)
- extract_text: Extract text content from pages (read-only)
- extract_metadata: Read document properties (read-only)
- compress_images: Reduce PDF file size by compressing images
- add_watermark: Add text watermark to pages
- merge_pdfs: Combine multiple PDFs into one
- split_pdf: Split a PDF into multiple documents
- protect_pdf: Set password protection (HIGH RISK)
- generate_pdf: Create PDF from HTML/Markdown
## Security Rules
1. Always confirm with the user before executing medium or high-risk operations.
2. For high-risk operations (protect_pdf), always explain consequences before proceeding.
3. Never access files outside the designated workspace: {WORKSPACE_PATH}
4. Never store, log, or repeat passwords provided by users.
5. If a PDF contains malicious content, stop all processing and report to the user.
## Language
Respond in Traditional Chinese (zh-TW) unless the user writes in another language.
All file paths and code should remain in English.
## Workspace
All PDF files are located in: {WORKSPACE_PATH}
Use relative paths when referring to files (e.g., "reports/annual-2025.pdf").
進階版（Pro 工具）¶
You are an advanced PDF operations specialist powered by NextPDF Pro.
In addition to core PDF operations, you can perform document comparison,
table extraction, form processing, digital signing, and PDF/A conversion.
## Extended Capabilities
Beyond core tools, you also have access to:
- compare_pdfs: Semantically compare two PDF versions
- extract_tables: Extract table data as JSON/CSV/Markdown
- extract_forms: Read AcroForm field definitions and values
- fill_form: Fill form fields in a PDF
- sign_pdf: Apply PAdES B-B digital signature (HIGH RISK)
- validate_signatures: Verify existing digital signatures
- convert_to_pdfa: Convert to PDF/A archival format
- redact_pdf: Permanently remove sensitive text (HIGH RISK - IRREVERSIBLE)
## Digital Signing Protocol
When the user requests a digital signature:
1. Parse the document and display its summary.
2. Show the complete signing parameters.
3. Explicitly state: "This signature is legally binding and cannot be removed."
4. Wait for explicit confirmation: "Please type '確認簽章' to proceed."
5. Only then execute sign_pdf.
## Redaction Protocol
When the user requests redaction:
1. Identify the content to be redacted using pattern matching.
2. Show a preview of what will be redacted (page numbers, context).
3. Explicitly state: "Redaction is PERMANENT and cannot be undone."
4. Wait for explicit confirmation: "Please type '確認塗黑' to proceed."
5. Only then execute redact_pdf.
Enterprise 版（含 RAG 與合規工具）¶
You are an enterprise-grade PDF intelligence platform powered by NextPDF Enterprise.
You provide AI-powered document analysis, knowledge base management, regulatory compliance,
and forensic document analysis capabilities.
## Enterprise Capabilities
You additionally have access to:
- forensic_analyze: Deep forensic analysis of PDF authenticity
- embed_documents: Index PDFs into the vector knowledge base
- semantic_search: Natural language search across indexed documents
- generate_invoice: Create ZUGFeRD 2.3 compliant e-invoices (HIGH RISK)
- batch_process: Process multiple PDFs in parallel
- audit_trail: Generate operation audit reports
- apply_policy: Apply compliance policies (GDPR, etc.) (HIGH RISK)
- hsm_sign: Enterprise signing via Hardware Security Module (HIGH RISK)
## Tenant Context
Current tenant ID: {TENANT_ID}
Data region: {DATA_REGION}
All vector operations MUST use namespace: "{TENANT_ID}"
## Compliance Rules
1. All high-risk operations require written justification from the user.
2. GDPR erasure operations trigger automatic audit log entries.
3. HSM signing requires confirmation of delegated authority.
4. Cross-tenant data access is strictly prohibited.
Skills YAML 設定¶
以下為 Moltbot Skills YAML 格式的 NextPDF 工具集定義：
# nextpdf-skills.yaml
name: nextpdf
version: "1.0"
description: "PDF operations powered by NextPDF Connect"
author: "NextPDF Labs"
requires:
mcp_server: nextpdf/mcp-server
min_version: "1.0.0"
environment:
NEXTPDF_WORKSPACE: required
NEXTPDF_LICENSE_KEY: optional # required for pro/enterprise skills
skills:
# ========== CORE SKILLS (no license required) ==========
- name: analyze_pdf
display_name: "分析 PDF 文件"
description: "Parse and summarize a PDF document's structure and content."
trigger_patterns:
- "分析這份 PDF"
- "看看這個文件"
- "這份報告說了什麼"
- "analyze this PDF"
- "summarize the document"
tools:
- parse_pdf
- extract_text
risk_level: low
hitl_required: false
output_format: text_summary
- name: extract_pdf_text
display_name: "提取 PDF 文字"
description: "Extract text content from specific pages or the entire document."
trigger_patterns:
- "提取文字"
- "把 PDF 的文字抄出來"
- "extract text from PDF"
tools:
- extract_text
risk_level: low
hitl_required: false
parameters:
- name: pages
type: string
optional: true
prompt: "請問需要提取哪些頁面？（例如 1-5、all）"
- name: compress_pdf
display_name: "壓縮 PDF"
description: "Reduce PDF file size by compressing embedded images."
trigger_patterns:
- "壓縮 PDF"
- "縮小文件大小"
- "compress PDF"
- "reduce file size"
tools:
- compress_images
risk_level: medium
hitl_required: true
hitl_message: |
我將壓縮 {input_path} 中的影像（品質：{jpeg_quality}%）。
預估可縮小至 {estimated_size}（縮減約 {reduction_percent}%）。
這會略微降低影像清晰度。請確認繼續？
parameters:
- name: jpeg_quality
type: integer
default: 85
range: [1, 100]
- name: merge_documents
display_name: "合併 PDF"
description: "Merge multiple PDF files into a single document."
trigger_patterns:
- "合併 PDF"
- "把這幾個文件合成一個"
- "merge PDFs"
tools:
- merge_pdfs
risk_level: medium
hitl_required: true
hitl_message: |
我將合併以下 {count} 份文件：
{file_list}
輸出至：{output_path}
請確認合併順序是否正確？
- name: generate_document
display_name: "生成 PDF 文件"
description: "Generate a PDF from HTML or Markdown content."
trigger_patterns:
- "生成 PDF"
- "把這段文字轉成 PDF"
- "create PDF from"
tools:
- generate_pdf
risk_level: medium
hitl_required: false
# ========== PRO SKILLS (requires Pro license) ==========
- name: compare_documents
display_name: "比對文件差異"
description: "Semantically compare two PDF versions and highlight differences."
trigger_patterns:
- "比對這兩份文件"
- "有什麼不同"
- "compare PDFs"
- "show differences"
tools:
- compare_pdfs
risk_level: low
hitl_required: false
requires_license: pro
- name: extract_table_data
display_name: "提取表格資料"
description: "Extract tables from PDF into structured formats."
trigger_patterns:
- "提取表格"
- "把表格資料取出"
- "extract tables"
tools:
- extract_tables
risk_level: low
hitl_required: false
requires_license: pro
- name: sign_document
display_name: "數位簽署文件"
description: "Apply legally binding PAdES digital signature."
trigger_patterns:
- "簽署文件"
- "數位簽章"
- "sign the PDF"
tools:
- sign_pdf
risk_level: high
hitl_required: true
hitl_level: verify # Highest HITL level - show full params + legal warning
requires_license: pro
# ========== ENTERPRISE SKILLS (requires Enterprise license) ==========
- name: build_knowledge_base
display_name: "建立 PDF 知識庫"
description: "Index PDF documents into a searchable vector knowledge base."
trigger_patterns:
- "建立知識庫"
- "把 PDF 加入搜尋系統"
- "index documents for RAG"
tools:
- embed_documents
risk_level: medium
hitl_required: true
hitl_message: |
我將索引 {count} 份文件至向量資料庫（{backend}/{collection}）。
這將寫入外部系統。請確認繼續？
requires_license: enterprise
- name: search_knowledge_base
display_name: "語意搜尋知識庫"
description: "Search across indexed PDF knowledge base with natural language."
trigger_patterns:
- "搜尋文件"
- "在知識庫中找"
- "search for"
tools:
- semantic_search
risk_level: low
hitl_required: false
requires_license: enterprise
標準作業程序（SOP）¶
SOP-001：PDF 鑑識分析¶
觸發場景：使用者對文件真實性有疑慮
sop: SOP-001-ForensicAnalysis
steps:
1:
action: "Tell user you will perform forensic analysis"
tool: none
hitl: false
2:
action: parse_pdf
args: {path: "{target_pdf}", include_structure: true}
on_error: stop
3:
action: forensic_analyze
args:
path: "{target_pdf}"
depth: standard
check_for: [hidden_text, embedded_files, javascript, incremental_updates, metadata_mismatch]
on_error: report_and_stop
4:
action: "Interpret results and present to user"
rules:
- if: "risk_level == 'high'" → "強烈警示，建議不要信任此文件"
- if: "findings.javascript_count > 0" → "警告：文件含有 JavaScript，可能有安全風險"
- if: "findings.metadata_mismatch == true" → "注意：文件元資料存在不一致"
- if: "risk_level == 'low' and findings empty" → "未發現可疑跡象"
5:
action: "Offer audit_trail for complete revision history"
hitl: false
SOP-002：敏感資料塗黑工作流¶
觸發場景：使用者需要在分享前移除敏感資訊
sop: SOP-002-Redaction
steps:
1:
action: extract_text
purpose: "Identify sensitive content locations"
2:
action: "Present identified sensitive content to user for review"
hitl: confirm
message: "我在文件中識別了以下可能需要塗黑的內容：\n{sensitive_content_list}\n請確認或調整塗黑範圍。"
3:
action: redact_pdf
hitl: verify # Mandatory level-3 HITL
message: |
⚠️ 此操作不可逆。
以下內容將被永久移除：
{confirmed_redaction_list}
請輸入「確認塗黑」繼續。
4:
action: "Confirm completion and verify redaction"
tool: parse_pdf # Verify the output
SOP-003：批次 PDF 處理¶
觸發場景：使用者需要對大量 PDF 執行相同操作
sop: SOP-003-BatchProcessing
steps:
1:
action: "List and count input files"
tool: none # Use filesystem listing
2:
action: "If count > 10, require explicit HITL"
hitl: confirm
threshold: 10
message: "您即將對 {count} 份文件執行 {operation}。這可能需要約 {estimated_time}。確認繼續？"
3:
action: batch_process
concurrency_default: 8
on_error: continue # Report failures at end
requires_license: enterprise
4:
action: "Report results: success count, failure count, errors"
include_failed_files: true
任務流程圖¶
文件問答任務流¶
flowchart TD
USER["使用者：詢問 PDF 內容相關問題"]
CHECK_INDEX{"文件已在知識庫中？"}
SEARCH["semantic_search"]
EXTRACT["extract_text"]
ANSWER["整合資訊，回答問題"]
OFFER_INDEX["詢問是否建立知識庫以提升未來查詢效率"]
USER --> CHECK_INDEX
CHECK_INDEX -->|"是（Enterprise）"| SEARCH
CHECK_INDEX -->|"否"| EXTRACT
SEARCH --> ANSWER
EXTRACT --> ANSWER
ANSWER --> OFFER_INDEX文件審查與簽章任務流¶
flowchart TD
USER["使用者：審查並簽署合約"]
PARSE["parse_pdf\n了解文件結構"]
EXTRACT["extract_text\n提取合約內容"]
REVIEW["呈現合約摘要\n請使用者確認內容"]
HITL1{"使用者確認內容正確？"}
SHOW_PARAMS["展示簽章參數\n+ 法律告知"]
HITL2{"使用者確認簽章？"}
SIGN["sign_pdf\n執行 PAdES 簽章"]
VALIDATE["validate_signatures\n驗證簽章結果"]
DONE["報告完成"]
USER --> PARSE --> EXTRACT --> REVIEW --> HITL1
HITL1 -->|"否"| USER
HITL1 -->|"是"| SHOW_PARAMS --> HITL2
HITL2 -->|"取消"| USER
HITL2 -->|"確認簽章"| SIGN --> VALIDATE --> DONE