wiki-graph-core 0.3.1

package/data/topologization/book_coherence_extraction.jinja

@@ -0,0 +1,223 @@
+You are reading a long text and need to extract essential connective tissue for maintaining the book's narrative coherence.
+
+---
+
+{% if user_language %}
+**Language Configuration:**
+
+**User interface language:** {{ user_language }}
+
+**Output Requirements:**
+- **label** and **content** fields: Must be in **{{ user_language }}** (translate from source text if needed)
+- **evidence**: Use the evidence selection protocol below in the **original language(s)** (NEVER translate)
+- Use **evidence** to identify the supporting source sentence.
+
+**Important Notes:**
+- The source text may be in any language or mix of languages
+- Your job is to:
+  1. Understand the text in whatever language(s) it uses
+  2. Translate only **label** and **content** to {{ user_language }} for user understanding
+  3. Keep **evidence** quotes in exact original language(s) - never translate them
+
+Example output structure:
+```json
+{
+  "importance_annotations": [
+    {"chunk_id": 19, "importance": "critical"}
+  ],
+  "chunks": [
+    {
+      "temp_id": "D",
+      "label": "{{ user_language }} translated label",
+      "content": "{{ user_language }} translated description",
+      "evidence": [
+        {
+          "sentence_id": "S1",
+          "quote": "Exact short original-language quote"
+        }
+      ],
+      "importance": "critical"
+    }
+  ]
+}
+```
+
+{% else %}
+**Language Note:** No user language specified. Keep all output fields in the **SAME LANGUAGE(S) as the source text**. Do not translate anything.
+
+**Output Requirements:**
+- **label**, **content**, and **evidence**: All in the source text's original language(s)
+- Use **evidence** to identify the supporting source sentence.
+- If the source text contains multiple languages, preserve them as they appear
+{% endif %}
+
+---
+
+## CONTEXT
+
+These user-focused chunks were already extracted from this text segment:
+
+{% if user_focused_chunks %}
+{% for chunk in user_focused_chunks %}
+{{ chunk.id }}. [{{ chunk.label }}] {{ chunk.content }}
+{% endfor %}
+{% else %}
+(No user-focused chunks were extracted from this segment)
+{% endif %}
+
+**Current Working Memory:**
+{{ working_memory }}
+
+---
+
+## TASK
+
+Your task has two parts:
+
+1. **Extract new book-coherence chunks** that are ESSENTIAL for understanding the book's logic but were NOT captured in the user-focused chunks above.
+
+2. **Annotate existing user-focused chunks with importance** if they are also critical for book coherence. Do NOT create duplicate chunks - instead, add their IDs to `importance_annotations`.
+
+**Be conservative**: Extract 1-3 new chunks ONLY if they meet the strict criteria below. Many segments need NO new book-coherence chunks at all - if the user-focused chunks already cover the essential connective tissue, output empty arrays. Don't extract just for the sake of extracting.
+
+---
+
+## EXTRACTION CRITERIA
+
+### What to extract (book-coherence chunks):
+
+Extract information essential for understanding the book's logic and narrative flow, REGARDLESS of the reader's personal interests. This includes:
+
+- **First introductions** of key concepts, characters, or settings that will be referenced later
+- **Causal relationships** that drive the plot or argument forward
+- **Context and background** that makes later content comprehensible
+- **Turning points or transitions** in the narrative structure
+- **Definitions, rules, or constraints** established by the author that affect subsequent content
+
+### What to avoid:
+
+- Information **already captured** in the user-focused chunks above (use `importance_annotations` instead)
+- Details that **won't affect comprehension** of later content
+- Generic filler, repetition, or tangential information
+- Content that's interesting but not structurally necessary
+
+### When to use importance_annotations:
+
+If a user-focused chunk represents a **major turning point, causal link, or structural necessity** for the book's narrative flow, add it to `importance_annotations` with an appropriate importance level. This allows the same information point to have both user retention level AND book importance level.
+
+### Link criteria:
+
+Connect book-coherence chunks when there are clear **logical dependencies**, **cause-and-effect relationships**, or **narrative continuity** between them.
+
+Also link to user-focused chunks if they provide essential context for each other (e.g., book-coherence chunk explains WHY something in user-focused chunk happened).
+
+---
+
+## IMPORTANCE LEVEL GUIDELINES
+
+Rate each book-coherence chunk (being extracted means it's already important, so use only 3 levels):
+
+- **critical**: Essential setup, major turning points, or causal links that cannot be skipped
+  - Use sparingly - only for structural necessities
+- **important**: Key context, definitions, or transitions that significantly aid understanding
+  - This should be the majority of book-coherence chunks
+- **helpful**: Supplementary background or minor connective tissue
+  - Use for nice-to-have context
+
+**Remember**: Book-coherence extraction should be conservative. If in doubt whether something is necessary for coherence, it probably isn't.
+
+---
+
+## LINK STRENGTH GUIDELINES
+
+Rate each link's dependency strength:
+
+- **critical**: Must understand FROM to understand TO (cannot skip FROM)
+- **important**: FROM provides essential context for TO (should not skip FROM)
+- **helpful**: FROM and TO are related but both can stand alone (optional connection)
+
+---
+
+**Output Format (JSON only, no other text):**
+
+The user will provide the same text segment in their message. Extract ONLY book-coherence chunks that supplement the user-focused extraction.
+
+CRITICAL: You MUST output in this EXACT structure with "importance_annotations" FIRST, "chunks" SECOND, then "links" THIRD:
+
+{
+  "importance_annotations": [
+    {"chunk_id": 19, "importance": "critical"},
+    {"chunk_id": 23, "importance": "important"}
+  ],
+  "chunks": [
+    {
+      "temp_id": "D",
+      "label": "brief topic label (5-15 chars)",
+      "content": "full description of the connective information",
+      "evidence": [
+        {
+          "sentence_id": "S1",
+          "quote": "exact short original-language quote"
+        }
+      ],
+      "importance": "critical"
+    }
+  ],
+  "links": [
+    {"from": "D", "to": 1, "strength": "important"},
+    {"from": 5, "to": "D", "strength": "helpful"}
+  ]
+}
+
+**If no book-coherence work is needed, output:**
+{
+  "importance_annotations": [],
+  "chunks": [],
+  "links": []
+}
+
+This is a VALID successful result.
+- Do NOT apologize
+- Do NOT say you cannot answer
+- Do NOT say no result was found
+- Do NOT output explanatory text outside the JSON
+- When nothing new is needed for coherence, return the empty JSON object above directly
+
+**Field Explanations:**
+- "importance_annotations": Array of importance labels for EXISTING user-focused chunks (0-5 annotations)
+  - Use this when a user-focused chunk is ALSO important for book coherence
+  - "chunk_id": The numeric ID of the user-focused chunk (from the list above)
+  - "importance": One of "critical", "important", or "helpful"
+  - This avoids duplicate extraction - the same information point gets both retention and importance labels
+  - **Process this FIRST**: Identify which user-focused chunks are also book-critical, then extract only what's missing
+
+- "chunks": Array of NEW book-coherence chunks (0-3 chunks, only if essential and NOT already in user-focused)
+  - "temp_id": Use letters continuing from user-focused chunks (if user had A,B,C then start with D,E,F)
+  - "type": MUST be "book_coherence"
+  - "label": 5-15 character summary
+  - "content": Full one-sentence description
+  - "evidence": Evidence selection for this chunk
+{{ evidence_selection_prompt | safe }}
+    - Keep quote in the original source language(s)
+    - Do NOT translate quote
+  - "importance": One of "critical", "important", or "helpful"
+  - **Process this SECOND**: After annotating existing chunks, extract only missing book-critical information
+
+- "links": Array of relationships between chunks
+  - "from" can be: working memory ID (number) OR temp_id (string from user-focused OR book-coherence chunks)
+  - "to" can be: working memory ID (number) OR temp_id (string from user-focused OR book-coherence chunks)
+  - "strength": One of "critical", "important", or "helpful"
+  - Can link book-coherence chunks to user-focused chunks (e.g., "D" to 1)
+  - Use empty array [] if no relationships exist
+
+**IMPORTANT:**
+1. The order matters! Output "importance_annotations" FIRST, "chunks" SECOND, then "links" THIRD.
+2. Think in this order: First identify which user-focused chunks are book-critical, then extract only missing information, finally link them.
+3. Every chunk MUST have a "type" field set to "book_coherence".
+4. Every chunk MUST have an "importance" field with one of the three valid values.
+5. Every importance_annotation MUST reference a valid user-focused chunk_id from the list above.
+6. Every link MUST have a "strength" field with one of the three valid values.
+7. Be conservative - only extract/annotate what's truly necessary for coherence.
+8. Do NOT create new chunks for information already in user-focused chunks - use importance_annotations instead.
+9. If no new annotations/chunks/links are needed, the correct output is still valid JSON with empty arrays.
+10. A "nothing to add" case is NOT a failure. Never respond with apology/refusal text such as "I cannot answer", "no relevant result", or similar wording.

package/data/topologization/chunk_extraction.jinja

@@ -0,0 +1,233 @@
+You are a reading intention analyzer. Your task is to convert a user's reading intention into a **precise extraction ruleset**.
+
+**CRITICAL: Use hybrid language approach:**
+- **Structure keywords**: Keep in English (Rule, Semantic pattern, Information type, Recognition criteria, Specificity, Retention level, Do NOT extract if, verbatim, detailed, focused, relevant)
+- **User intent descriptions**: Write in the user's language (Chinese if user writes in Chinese, English if user writes in English)
+- This ensures AI recognizes the structure while preserving semantic precision of user intent
+
+The user will tell you their reading intention in their message. Based on that intention, generate an extraction ruleset that defines:
+1. WHAT information patterns should be extracted
+2. WHAT retention level each pattern should have
+3. WHAT should be explicitly excluded
+
+---
+
+**Design Philosophy:**
+
+This framework is **genre-agnostic** - it works for narratives, academic papers, philosophy books, industry reports, and any other text type. Do NOT assume:
+- The text has "scenes" or "characters"
+- The text has a "protagonist"
+- The text distinguishes "narrator" from "author"
+- The text is a story
+
+Instead, focus on **semantic patterns** - what types of information does the user care about, regardless of text structure.
+
+---
+
+**Output Format (Structured rules, will be embedded in downstream prompts):**
+
+Generate a numbered list of extraction rules. **Keep all structure keywords in English, but write content descriptions in the user's language.** Each rule must follow this EXACT structure:
+
+## Extraction Rules
+
+### Rule [N]: [Brief rule name in user's language]
+
+**Semantic pattern (extract content matching this description):**
+- Information type: [What kind of information does user want? Use user's own words and concepts]
+- Recognition criteria: [How to identify this information? Describe semantic features, not structural positions]
+- Specificity: [Any constraints? e.g., "首次出现", "针对特定对象X", "包含具体数据", "direct quotations"]
+
+**Retention level:** [verbatim / detailed / focused / relevant] ← Keep these 4 terms in English
+
+**Do NOT extract if:**
+- [Exclusion conditions in user's language, based on user's "可以压缩/删节" hints]
+- [Use semantic descriptions, not structural rules]
+
+---
+
+**After all extraction rules, add a default rule:**
+
+### Default Rule
+
+**IF:** Content does NOT match any of the above rules
+
+**THEN:** Do NOT extract (leave for book-coherence phase only)
+
+---
+
+## Link Criteria
+
+[1-2 sentences explaining when two chunks should be connected. What relationships matter? Focus on semantic connections, not structural ones.]
+
+---
+
+**Important Guidelines:**
+
+**Understanding trigger phrases for retention levels:**
+- **Verbatim triggers** (entire passages): "一字不漏", "一字不改", "原文照抄", "原样引用", "verbatim", "word-for-word"
+- **Detailed triggers**: "必须保留", "完整保留", "保留术语", "精确定义", "must preserve", "keep complete"
+- **Focused**: User's PRIMARY interest (未明确强调但核心关注)
+- **Relevant**: User's SECONDARY interest (可有可无的辅助信息)
+
+**How to write "Information type":**
+- Use the EXACT concepts and terminology from user's intention
+- Examples:
+  - If user says "别人对朱元璋的看法" → write "他人对朱元璋的评价、态度或行为"
+  - If user says "model architecture" → write "模型的结构、层次或参数设计"
+  - If user says "核心论证" → write "作者的主要论证步骤或逻辑推理"
+  - If user says "market data" → write "市场规模、增长率或份额数据"
+
+**How to write "Recognition criteria":**
+- Describe SEMANTIC features, not structural positions
+- Focus on "what does this information look like" rather than "where is it located"
+- Examples:
+  - ✅ GOOD: "描述具体人物针对朱元璋的言语、动作或态度表现"
+  - ❌ BAD: "出现在与朱元璋互动的场景中"
+  - ✅ GOOD: "包含具体的数值、百分比或统计结果"
+  - ❌ BAD: "出现在数据章节"
+  - ✅ GOOD: "作者明确陈述的推理步骤或概念定义"
+  - ❌ BAD: "论证部分的内容"
+
+**How to write "Specificity" (optional, use when user has special requirements):**
+- Temporal constraints: "首次出现"、"最后一次"、"贯穿全文的"
+- Targeting constraints: "针对主角X"、"关于概念Y"、"涉及公司Z"
+- Format constraints: "直接引语"、"包含数据"、"带有作者评价"
+- Completeness constraints: "完整语句"、"上下文完整"
+
+**How to write "Do NOT extract if":**
+- Based on user's explicit exclusion hints: "社会背景可以压缩" → "内容主要描述社会背景、历史事件或环境"
+- Based on common distinctions:
+  - For narratives: External actions vs internal thoughts (外在行为 vs 内心活动)
+  - For academic texts: Core claims vs background review (核心观点 vs 背景综述)
+  - For reports: Primary data vs contextual description (关键数据 vs 背景描述)
+- Use SEMANTIC criteria: "内容是概括性陈述而非具体事例" NOT "内容是作者而非角色所说"
+
+**CRITICAL: Avoid narrative-specific assumptions:**
+- ❌ DO NOT write: "Speaker/Actor: 故事中的人物（非叙述者/作者）"
+  - Problem: Assumes narrator/character separation, which doesn't exist in many texts
+- ✅ WRITE INSTEAD: Recognition criteria: "描述具体人物的行为或言语（而非概括性评论）"
+  - Semantic distinction: Specific instance vs. general commentary
+
+- ❌ DO NOT write: "Scene/Context: 第一次见面"
+  - Problem: Assumes text has "scenes"
+- ✅ WRITE INSTEAD: Specificity: "文本明确标识为'首次'、'初次'或'第一次'的相关描述"
+  - Semantic distinction: First occurrence vs. subsequent occurrences
+
+**Examples of good rule structures:**
+
+**Example 1: Narrative text (Chinese user intention)**
+
+User says: "压缩此书，重点关注别人对朱元璋的看法和对待方式，第一次见面的评价必须一字不漏。社会背景可以压缩。"
+
+✅ GOOD:
+```
+### Rule 1: 他人对朱元璋的首次评价
+
+**Semantic pattern (extract content matching this description):**
+- Information type: 他人对朱元璋的评价、态度或反应
+- Recognition criteria: 描述具体人物针对朱元璋的言语、表情、动作或判断（而非概括性的历史评论）
+- Specificity: 文本明确标识为"第一次见面"、"初次相识"、"初见"或类似表述的场合
+
+**Retention level:** verbatim
+
+**Do NOT extract if:**
+- 内容描述朱元璋自己的内心想法或感受
+- 内容是作者对历史的概括性评论（如"元末地主普遍如此"）
+
+### Rule 2: 他人对待朱元璋的行为和态度
+
+**Semantic pattern (extract content matching this description):**
+- Information type: 他人对朱元璋的行为、态度或处置方式
+- Recognition criteria: 描述具体人物针对朱元璋的具体行动或态度表现（而非对群体的泛泛描述）
+- Specificity: 任何时间，但非"第一次见面"（已被Rule 1覆盖）
+
+**Retention level:** detailed
+
+**Do NOT extract if:**
+- 内容主要描述社会背景、历史事件或环境，而非具体人物的行为
+- 内容描述朱元璋自己的内心想法、决策过程或心理变化
+```
+
+**Example 2: Academic paper (English user intention)**
+
+User says: "Compress this ML paper, focus on model architecture and experimental results. Background can be compressed."
+
+✅ GOOD:
+```
+### Rule 1: Model architecture description
+
+**Semantic pattern (extract content matching this description):**
+- Information type: Model structure, layers, parameters, or computational flow
+- Recognition criteria: Technical descriptions of how the model is constructed or operates, including specific layer types, dimensions, or architectural innovations
+- Specificity: Core architectural components (not historical background of similar models)
+
+**Retention level:** detailed
+
+**Do NOT extract if:**
+- Content is background review of related work
+- Content is general motivation without technical specifics
+
+### Rule 2: Experimental results and comparisons
+
+**Semantic pattern (extract content matching this description):**
+- Information type: Performance metrics, comparison with baselines, or ablation study results
+- Recognition criteria: Specific numerical results, tables, or quantitative comparisons showing model performance
+- Specificity: Primary results (not auxiliary experiments unless directly compared to baselines)
+
+**Retention level:** detailed
+
+**Do NOT extract if:**
+- Content is experimental setup description without results
+- Content is background discussion of evaluation methods
+```
+
+**Example 3: Philosophy book (Chinese user intention)**
+
+User says: "压缩康德《纯粹理性批判》，重点保留先验感性论和先验逻辑的核心论证，'时空是先天直观形式'的论述一字不漏。历史背景可删。"
+
+✅ GOOD:
+```
+### Rule 1: 时空先天性论述
+
+**Semantic pattern (extract content matching this description):**
+- Information type: 关于时空作为先天直观形式的论证或陈述
+- Recognition criteria: 康德关于时空的先天性、直观性或形式性的直接论述或论证步骤
+- Specificity: 包含"先天"、"直观"、"形式"等核心概念，且专门论述时空的段落
+
+**Retention level:** verbatim
+
+**Do NOT extract if:**
+- 内容是对其他哲学家观点的引述或批判（非康德自己的论证）
+
+### Rule 2: 先验感性论和先验逻辑的核心论证
+
+**Semantic pattern (extract content matching this description):**
+- Information type: 先验感性论或先验逻辑中的主要论证步骤、概念定义或推理链条
+- Recognition criteria: 康德阐述核心观点、定义关键概念、或进行逻辑推导的段落（而非举例说明或历史回顾）
+- Specificity: 排除时空论述（已被Rule 1覆盖）
+
+**Retention level:** detailed
+
+**Do NOT extract if:**
+- 内容是历史背景介绍或对前人观点的综述
+- 内容是辅助性的例子说明（非核心论证）
+```
+
+❌ BAD (uses narrative-specific structure):
+```
+### Rule 1: 核心论证
+
+**IF (Content matches ALL of these conditions):**
+- Content type: 论证步骤
+- Scene/Context: 先验感性论章节  ← 这是结构位置，不是语义特征
+- Speaker/Actor: 康德（非叙述者/作者）  ← 康德就是作者！矛盾！
+
+**THEN:**
+- Retention level: detailed
+```
+
+**Number of rules:**
+- Typically 3-6 rules depending on complexity
+- Each rule should be mutually exclusive where possible
+- Rules are checked in order - first match wins
+- More specific rules should come before more general rules

package/data/topologization/evidence_choice.jinja

@@ -0,0 +1,38 @@
+You previously extracted a chunk from the source text.
+That chunk is already written and should not be changed.
+
+One evidence field in that chunk is ambiguous, so you need to identify which candidate occurrence in the source text
+best matches what the chunk was referring to.
+
+This chunk originally came from the following extraction context:
+
+{% if metadata_field == "retention" %}
+User-focused extraction rules:
+{{ extraction_guidance }}
+{% else %}
+Book-coherence extraction rules:
+- Extract only information essential for understanding the text's logic and flow.
+- Prefer first introductions, causal links, critical context, turning points, and definitions.
+- Do not duplicate information already covered by user-focused chunks.
+- Use this stage for connective tissue that is structurally necessary for comprehension.
+{% endif %}
+
+Your job:
+1. Read the previously generated chunk
+2. Read the candidate occurrences from the source text
+3. Use the extraction context above to infer why this chunk was selected
+4. Choose exactly one occurrence ID for the requested evidence field
+
+Output format:
+```json
+{"choice": "S2"}
+```
+
+Rules:
+- Choose exactly one candidate occurrence ID from the provided list
+- Do not output any other fields
+- Do not explain your reasoning
+- Do not rewrite the chunk
+- Use the extraction context above as part of your judgment
+- Use the candidate context (`prev`, `text`, `next`) to disambiguate duplicates
+- If several candidates look similar, still choose the single best match from the provided list

package/data/topologization/translate_chunks.jinja

@@ -0,0 +1,138 @@
+You are a professional translator specialized in translating extracted information chunks.
+
+---
+
+**Target Language:** {{ user_language }}
+
+---
+
+## TRANSLATION STRATEGY
+
+**Understand the hierarchical structure:**
+- Each chunk has three fields with a clear hierarchical relationship:
+  1. **sourceSentences**: Original text excerpts (the source)
+  2. **content**: A summary/distillation of the sourceSentences
+  3. **label**: A condensed summary of the content (5-15 characters)
+- When translating, maintain this hierarchy:
+  - Translate **content** based on the meaning in sourceSentences
+  - Translate **label** as a concise summary of the translated content
+
+**Use sourceSentences as translation reference:**
+- sourceSentences provides the full context and original meaning
+- This is crucial for understanding proper nouns, technical terms, and ambiguous phrases
+- Especially important for short labels which may be unclear without context
+
+**Handle already-translated content:**
+- Some chunks may already be partially or fully in the target language
+- **Check each field (label, content) individually** before translating
+- **If a field is already in {{ user_language }}, keep it unchanged** - do not rephrase or rewrite
+- Only translate fields that are NOT in {{ user_language }}
+
+**Output format:**
+- **Do NOT include sourceSentences in your output** - it's for reference only
+- Only output: id, label, and content
+
+---
+
+## OUTPUT REQUIREMENTS
+
+**Format:** Return a JSON array with the exact same structure as the input.
+
+**Example:**
+
+Input:
+```json
+[
+  {
+    "id": 1,
+    "label": "新材料发现",
+    "content": "研究团队在实验室中偶然发现了一种具有超导特性的新型合金材料。",
+    "sourceSentences": ["在一次常规的材料测试中,科学家们意外地观察到一种金属混合物在室温下表现出零电阻现象。"]
+  },
+  {
+    "id": 2,
+    "label": "应用前景",
+    "content": "这种材料有望在能源传输和量子计算领域带来革命性变化。",
+    "sourceSentences": ["专家们认为,如果能够实现大规模生产,这项发现将彻底改变电力传输效率,并为下一代计算机技术铺平道路。"]
+  }
+]
+```
+
+Output (if target language is English):
+```json
+[
+  {
+    "id": 1,
+    "label": "New Material Found",
+    "content": "Research team accidentally discovered a new alloy with superconducting properties in the lab."
+  },
+  {
+    "id": 2,
+    "label": "Future Applications",
+    "content": "This material promises revolutionary changes in energy transmission and quantum computing."
+  }
+]
+```
+
+**Notice:**
+- Labels are concise translations informed by the source context
+- Content accurately reflects the meaning from sourceSentences
+- sourceSentences are NOT included in the output (they were only for translation reference)
+
+---
+
+**Example 2 - Mixed language input (some already translated):**
+
+Input:
+```json
+[
+  {
+    "id": 3,
+    "label": "New Discovery",
+    "content": "科学家发现了一种新的治疗方法。",
+    "sourceSentences": ["在最新的临床试验中，研究人员成功验证了一种创新的基因疗法。"]
+  },
+  {
+    "id": 4,
+    "label": "临床试验",
+    "content": "Clinical trials showed 85% success rate.",
+    "sourceSentences": ["临床试验结果显示，在参与的200名患者中，有170人出现了显著改善。"]
+  }
+]
+```
+
+Output (if target language is English):
+```json
+[
+  {
+    "id": 3,
+    "label": "New Discovery",
+    "content": "Scientists discovered a new treatment method."
+  },
+  {
+    "id": 4,
+    "label": "Clinical Trials",
+    "content": "Clinical trials showed 85% success rate."
+  }
+]
+```
+
+**Notice:**
+- Chunk 3: label was already in English (kept unchanged), only content was translated
+- Chunk 4: content was already in English (kept unchanged), only label was translated
+- No unnecessary rewording or rephrasing of already-translated text
+
+---
+
+## CRITICAL RULES
+
+1. **Output ONLY the JSON array** - no explanations, no markdown code blocks, no extra text
+2. **Every chunk** in the input MUST appear in the output with the same id
+3. **Preserve array order** - output chunks in the same order as input
+4. **Never modify id values** - copy them exactly as provided
+5. **Check each field's language individually**:
+   - If label is already in {{ user_language }}, output it unchanged
+   - If content is already in {{ user_language }}, output it unchanged
+   - Only translate fields that are NOT in {{ user_language }}
+6. **Never rephrase or rewrite** content that's already in the target language
+7. **Do NOT output sourceSentences** - they are for reference only, not part of the output