Build Reflection-Based PDF Translation with Make.com

Overview

This is an advanced automation solution pursuing translation quality.

By introducing “faithfulness, expressiveness, elegance” reflection mechanism, eliminate robotic machine translation tone:

1. PDF Extraction - Jina Reader converts to Markdown
2. Smart Segmentation - Repeater divides by set granularity
3. Draft Translation - AI generates first version translation
4. Reflection Review - Proposes improvements based on translation principles
5. Generate Final - Integrates suggestions to output high-quality translation
6. Save Results - Write back to Notion database

PDF reflection translation automation workflow architecture

---

Core Decision Factors

When choosing AI translation solutions, focus on:

• Translation Quality - Can it achieve “faithfulness, expressiveness, elegance” standards, avoiding literal translation
• Model Flexibility - Can you switch between different AI models based on tasks
• Text Processing Capability - Segmentation and format handling for long texts
• Customization Capability - Support for glossaries and language pair settings
• Cost-effectiveness - Cost-performance ratio while ensuring quality

---

Technical Specifications

Specification	Value	Notes
Supported AI Models	GPT-4o Mini, Claude, Gemini, etc.	Switch via OpenRouter
Text Segmentation Granularity	100/300/600/1000 characters	Adjustable based on needs
Notion PDF Limit	5MB	Free account, can compress
Notion Text Limit	2000 characters	Need segmentation or Append save
PDF Extraction API	Jina Reader	Nearly free
GPT-4o Mini Cost	<$10/million tokens	Extremely cost-effective

---

Prerequisites

Before starting, ensure you have:

• Make.com account (free registration)
• OpenRouter API key (multi-model support)
• Jina Reader API key (PDF extraction)
• Notion account and database

Notion Database Structure

Database property settings

Create translation task database with these fields:

• Status (Select) - Pending/Processing/Completed
• Model Selection (Select) - GPT-4o Mini/Claude/Gemini
• Text Granularity (Select) - 100/300/600/1000
• Source Language (Select) - English/Japanese/French, etc.
• Target Language (Select) - Chinese/English, etc.
• Glossary (Text) - Custom technical terms
• PDF File (Files) - Document to translate
• Translation Result (Text) - Output translation

---

Workflow Architecture

Stage 1: PDF Extraction

Use Jina Reader to convert PDF to Markdown:

Configuration Points:

• API nearly free
• Outputs Markdown format
• Preserves document structure

Note: Ensure correct credentials when configuring API Key, otherwise 401 errors occur.

---

Stage 2: Smart Segmentation

Text segmentation formula setup

Use Repeater module for long text segmentation:

Segmentation Logic:

• Based on user-set granularity (e.g., 1000 characters)
• Use Substring function for dynamic splitting
• Calculate total segments: text length / granularity

Why Segment?

• Avoid exceeding model token limits
• Improve translation accuracy
• Facilitate segment-by-segment review

---

Stage 3: Draft Translation

Call AI model to generate first version translation:

Prompt Design Points:

You are a professional translator fluent in {{source_lang}} and {{target_lang}}.

Please translate the following text to {{target_lang}}:
- Maintain accuracy of original meaning
- Use natural, fluent expressions
- Reference glossary: {{glossary}}

Original text:
{{text_segment}}

---

Stage 4: Reflection Review (Core Innovation)

Original text, improved translation, and suggestion comparison

This is the workflow’s core highlight—reflection mechanism based on “faithfulness, expressiveness, elegance” principles:

Reflection Prompt Design:

You are a translation quality reviewer. Please review the following translation based on "faithfulness, expressiveness, elegance" principles:

- Faithfulness: Does it accurately convey original meaning
- Expressiveness: Is it smooth and fluent, conforming to target language habits
- Elegance: Is it refined and appropriate, avoiding robotic translation tone

Original: {{original}}
Translation: {{translation}}

Please propose {{suggestion_count}} specific improvement suggestions.

Suggestion Count Calculation:

• 1000-character text: ~26 suggestions
• 100-character text: ~5-7 suggestions
• Formula: text granularity / 50 + 6

Warning: Don’t blindly request too many suggestions (e.g., 100); causes AI hallucination with meaningless suggestions.

---

Stage 5: Generate Final Translation

Quality improvement from reflection translation

Integrate review suggestions to generate final translation:

Effect Comparison:

• “Build” vs “Construct”
• “AI Programming” vs “Write AI Code”
• More natural, more “human-like”

---

Gotchas

Common pitfalls in practice:

1. Long Document Time - Full translation of long documents takes considerable time, possibly tens of minutes

2. Notion Limits - PDF limited to 5MB, text properties to 2000 characters

3. JSON Format Issues - Special characters like quotes in text may break JSON structure

4. API Authorization Errors - Ensure correct API Key configuration to avoid 401 errors

5. Suggestion Count Control - Too many suggestions cause AI hallucination; need reasonable settings

---

Use Cases

Recommended Users

• Professional Translators - Need high-quality translation handling large volumes of PDF documents
• Content Creators - Need polished translation eliminating “AI flavor”
• International Enterprises/Researchers - Need to translate internal documents and research reports

May Not Suit

• Complete beginners unfamiliar with Make.com and API configuration
• Regular users only occasionally needing simple translations
• Scenarios with extreme timeliness requirements for translation

---

FAQ

What is reflection translation?

After initial draft translation, AI self-reviews based on “faithfulness, expressiveness, elegance” principles and proposes improvements, ultimately generating higher quality final translation that eliminates robotic AI tone.

Why use OpenRouter?

OpenRouter supports multi-model switching: use GPT-4o Mini for simple tasks to save money, use Claude for high-quality requirements.

How to handle long texts?

Use Repeater module with Substring function for intelligent segmentation, supporting 100/300/600/1000 character granularities to avoid token limits.

What are Notion’s limitations?

Free account has 5MB PDF upload limit and 2000 character text property limit. Large PDFs need compression; long texts need segmentation or Append module saving.

---

Next Steps

After mastering the basic workflow, you can try:

• Adding more language pair support
• Integrating professional terminology databases
• Adding translation memory functionality
• Outputting to Google Docs for saving

Questions? Feel free to leave comments!