Update run_openai_public.sh and run_pipeline_json.py for new provider and model configurations; enhance style randomization in HTML output generation.

Author: srrk-GreenMan

generate_synthetic_table/flow.py

@@ -765,6 +765,7 @@ def build_synthetic_table_graph(
     llm: ChatOpenAI,
     provider: str = "openai",
     qa_only: bool = False,
+    skip_qa: bool = False,
 ) -> StateGraph:
     """
     Assemble the LangGraph pipeline.
@@ -773,6 +774,7 @@ def build_synthetic_table_graph(
         llm: LLM instance
         provider: LLM provider name
         qa_only: If True, generate QA directly from image without synthetic data generation
+        skip_qa: If True, skip QA generation after table generation (table only mode)
     """
 
     graph = StateGraph(TableState)
@@ -783,7 +785,7 @@ def build_synthetic_table_graph(
         graph.add_edge(START, "generate_qa_from_image")
         graph.add_edge("generate_qa_from_image", END)
     else:
-        # Full pipeline mode
+        # Full pipeline mode (or table-only mode if skip_qa=True)
         graph.add_node("image_to_html", image_to_html_node(llm))
         graph.add_node("pymupdf_parse", pymupdf_parse_node)
         graph.add_node("validate_parsed_table", validate_parsed_table_node(llm))
@@ -795,7 +797,9 @@ def build_synthetic_table_graph(
         graph.add_node("self_reflection", self_reflection_node(llm))
         graph.add_node("revise_synthetic_table", revise_synthetic_table_node(llm))
         graph.add_node("parse_synthetic_table", parse_synthetic_table_node(llm))
-        graph.add_node("generate_qa", generate_qa_node(llm))
+
+        if not skip_qa:
+            graph.add_node("generate_qa", generate_qa_node(llm))
 
         # Routing based on provider and input type
         def route_start(state: TableState) -> str:
@@ -842,8 +846,13 @@ def route_start(state: TableState) -> str:
         )
 
         graph.add_edge("revise_synthetic_table", "self_reflection")
-        graph.add_edge("parse_synthetic_table", "generate_qa")
-        graph.add_edge("generate_qa", END)
+
+        # Final edge: skip QA if requested
+        if skip_qa:
+            graph.add_edge("parse_synthetic_table", END)
+        else:
+            graph.add_edge("parse_synthetic_table", "generate_qa")
+            graph.add_edge("generate_qa", END)
 
     return graph
 
@@ -914,6 +923,7 @@ def run_synthetic_table_flow(
     azure_deployment: str | None = None,
     azure_endpoint: str | None = None,
     qa_only: bool = False,
+    skip_qa: bool = False,
     image_paths: List[str] | None = None,
     domain: str | None = None,
     # 체크포인팅 옵션
@@ -935,6 +945,7 @@ def run_synthetic_table_flow(
         azure_deployment: Azure OpenAI deployment name
         azure_endpoint: Azure OpenAI endpoint URL
         qa_only: If True, skip synthetic data generation and only generate QA from image
+        skip_qa: If True, generate table only without QA generation
         image_paths: Optional list of image paths for multi-image processing
         domain: Optional domain for prompt customization (e.g. 'public')
         enable_checkpointing: 체크포인팅 활성화 여부
@@ -955,7 +966,7 @@ def run_synthetic_table_flow(
         config_path=config_path,
     )
 
-    graph = build_synthetic_table_graph(llm, provider=provider, qa_only=qa_only)
+    graph = build_synthetic_table_graph(llm, provider=provider, qa_only=qa_only, skip_qa=skip_qa)
 
     # 체크포인팅 설정
     if enable_checkpointing:

generate_synthetic_table/prompts/academic.yaml

@@ -95,9 +95,8 @@ generate_qa_from_image: |
 generate_synthetic_table: |
   You are a Synthetic Data Generator specializing in Academic Data.
 
-  **CRITICAL INSTRUCTION: DO NOT COPY ORIGINAL DATA**
+  **⚠️ CRITICAL INSTRUCTION: DO NOT COPY ORIGINAL DATA ⚠️**
   Your task is to generate a new HTML table with the SAME STRUCTURE as the original but COMPLETELY DIFFERENT academic data values.
-  The goal is to create realistic synthetic academic data that looks like it could come from the same domain, but with entirely different students, courses, and metrics.
 
   **Inputs:**
   1. **Original Table Structure (for structure reference ONLY - DO NOT copy the data values):**
@@ -107,55 +106,67 @@ generate_synthetic_table: |
   {summary}
 
   **Requirements:**
-  1. **Structure:** Keep the exact same HTML structure (rows, columns, headers, merges).
+  1. **Structure:** Keep the exact same HTML structure (rows, columns, headers, merges, rowspan, colspan).
   2. **Headers:** Keep header text the same (column names, category labels).
-  3. **Data Transformation - MANDATORY:**
+  3. **⚠️ Data Transformation - ABSOLUTELY MANDATORY ⚠️:**
      - **ALL data cell values MUST be replaced with completely new synthetic values.**
-     - **DO NOT copy any original data values** - generate fresh, realistic alternatives.
-     - For student names: Generate new Korean student names (e.g., "김철수" → "이영희", "학생A" → "학생B")
-     - For university names: Generate new Korean university names
-     - For course titles: Generate new course names
-     - For grades/scores: Generate new realistic values
-     - For model names (if research table): Generate new model/method names
-     - For dates: Generate new plausible dates
-  4. **Domain Consistency:**
-     - Ensure academic logic (credits sum correctly, GPA calculations valid)
-     - Use realistic Korean academic terminology
-     - Contexts: Transcripts, Research Papers, Enrollment Stats, Faculty Lists
-  5. **Output:** Return ONLY the raw HTML string starting with `<table>` and ending with `</table>`.
-
-  **Example Transformation:**
-  - Original: "서울대학교" → Synthetic: "고려대학교"
-  - Original: "학점 4.2" → Synthetic: "학점 3.8"
-  - Original: "BERT-Large" → Synthetic: "RoBERTa-Base"
-
-  Remember: The synthetic table should look like a completely different academic dataset from the same domain.
+     - **NEVER copy any original data values** - generate fresh, realistic alternatives.
+     - For student/model names: Generate DIFFERENT names
+     - For university names: Generate DIFFERENT names
+     - For grades/scores: Generate DIFFERENT realistic values
+     - For course/research topics: Generate DIFFERENT titles
+     - For dates: Generate DIFFERENT plausible dates
+  4. **Styling:** Use **Tailwind CSS** classes (NO inline styles). **Observe and mimic the original image's visual style:**
+      - Look at the original image's color scheme and design
+      - Use appropriate Tailwind color classes to match the original style
+      - Basic structure: `<table class="w-full border-collapse text-sm">`
+      - Headers/cells: Include `border`, `px-4 py-3`, appropriate colors
+      - Lists: `class="list-disc ml-5 space-y-1"`
+      - **DO NOT use inline style attributes**
+  5. **Domain Consistency:** Ensure academic logic (credits sum correctly, GPA valid)
+  6. **Output:** Return ONLY the raw HTML string starting with `<table>` and ending with `</table>`. No markdown code blocks.
+
+  **Example Transformation (Generic):**
+  - Original name: "학생A" → Synthetic: "학생B"
+  - Original score: "4.0" → Synthetic: "3.5"
+  - Original model: "모델X" → Synthetic: "모델Y"
+
+  ⚠️ If the generated content is identical or very similar to the original, the output is INVALID.
 
 generate_synthetic_table_from_image: |
   You are a Synthetic Data Generator specializing in Academic Data.
 
-  **CRITICAL INSTRUCTION: DO NOT TRANSCRIBE - GENERATE NEW DATA**
+  **⚠️ CRITICAL INSTRUCTION: DO NOT TRANSCRIBE - GENERATE NEW DATA ⚠️**
   Your task is NOT to OCR/transcribe the image. Instead, you must:
   1. Understand the table's STRUCTURE from the image
   2. Understand it's an ACADEMIC table
-  3. Generate COMPLETELY NEW synthetic academic data that fits the domain but uses different values
+  3. Generate COMPLETELY NEW synthetic academic data that fits the domain but uses ENTIRELY DIFFERENT values
 
   **Inputs:**
   1. **Image:** An image of an academic table. Use this to understand structure and domain ONLY.
 
   **Requirements:**
-  1. **Structure Preservation:** Accurately reconstruct the table structure.
-  2. **Headers:** Keep header text (column names, category labels) the same as in the image.
-  3. **Data Generation - CRITICAL:**
-     - **DO NOT copy the data values from the image** - this is NOT an OCR task
-     - Generate COMPLETELY NEW synthetic academic values for all data cells
-     - For student/model names: Generate new names (different from what you see)
-     - For grades/scores: Generate new realistic values
-     - For course/research topics: Generate new titles
-  4. **Styling:** Use **Tailwind CSS** classes (same as default).
-      - `class="border-collapse border border-slate-400 w-full text-sm text-left rtl:text-right text-gray-500"` on `<table>`.
-      - `class="border border-slate-300 p-2 bg-gray-50 font-semibold"` on `<th>`.
-      - `class="border border-slate-300 p-2"` on `<td>`.
-  5. **Output Format:** Return ONLY the raw HTML string starting with `<table>` and ending with `</table>`.
-
-  Remember: The output should be a new synthetic academic dataset, not a transcription of the original.
+  1. **Structure Preservation:** Accurately reconstruct the table structure, including rowspan/colspan.
+  2. **Headers:** Keep header text the same as in the image.
+  3. **⚠️ Data Generation - ABSOLUTELY CRITICAL ⚠️:**
+     - **NEVER copy the data values from the image** - this is NOT an OCR task
+     - **ALL cell content must be completely NEW and DIFFERENT**
+     - For student/model names: Generate DIFFERENT names
+     - For grades/scores: Generate DIFFERENT values
+     - For course/research topics: Generate DIFFERENT titles
+  4. **Styling:** Use **Tailwind CSS** classes exclusively (NO inline styles).
+      - `<table>`: `class="w-full border-collapse text-sm"`
+      - `<thead>`: `class="bg-gradient-to-r from-indigo-700 to-indigo-800 text-white"`
+      - `<th>`: `class="border border-indigo-300 px-4 py-3 font-semibold text-left"`
+      - `<tbody>`: `class="divide-y divide-slate-200"`
+      - `<tr>` (body rows): `class="hover:bg-indigo-50 transition-colors"`
+      - `<td>`: `class="border border-slate-200 px-4 py-3 text-slate-700"`
+      - `<ul>`: `class="list-disc ml-5 space-y-1 text-slate-600"`
+      - **DO NOT use inline style attributes**
+  5. **Output Format:** Return ONLY the raw HTML string starting with `<table>` and ending with `</table>`. No markdown code blocks.
+
+  **Example (Generic):**
+  - Name in image: "이름X" → Generate: "이름Y"
+  - Score in image: "점수A" → Generate: "점수B"
+
+  ⚠️ If the generated content is identical or very similar to the image, the output is INVALID.

generate_synthetic_table/prompts/business.yaml

@@ -95,7 +95,7 @@ generate_qa_from_image: |
 generate_synthetic_table: |
   You are a Synthetic Data Generator specializing in Business Data.
 
-  **CRITICAL INSTRUCTION: DO NOT COPY ORIGINAL DATA**
+  **⚠️ CRITICAL INSTRUCTION: DO NOT COPY ORIGINAL DATA ⚠️**
   Your task is to generate a new HTML table with the SAME STRUCTURE as the original but COMPLETELY DIFFERENT business data values.
   The goal is to create realistic synthetic business data that looks like it could come from the same domain, but with entirely different companies, employees, products, and metrics.
 
@@ -107,54 +107,77 @@ generate_synthetic_table: |
   {summary}
 
   **Requirements:**
-  1. **Structure:** Keep the exact same HTML structure (rows, columns, headers, merges).
-  2. **Headers:** Keep header text the same (column names, category labels).
-  3. **Data Transformation - MANDATORY:**
+  1. **Structure:** Keep the exact same HTML structure (rows, columns, headers, merges, rowspan, colspan).
+  2. **Headers:** Keep header text the same (column names, category labels like 기업경쟁력, 시장경쟁력).
+  3. **⚠️ Data Transformation - ABSOLUTELY MANDATORY ⚠️:**
      - **ALL data cell values MUST be replaced with completely new synthetic values.**
-     - **DO NOT copy any original data values** - generate fresh, realistic alternatives.
-     - For company names: Generate new Korean company names (e.g., "삼성물산" → "현대상사", "A팀" → "B팀")
-     - For employee names: Generate new Korean names
-     - For product names: Generate new product line names
-     - For revenue/sales figures: Generate new realistic amounts (different values)
-     - For dates: Generate new plausible dates
-  4. **Domain Consistency:**
+     - **NEVER copy any original data values** - generate fresh, realistic alternatives.
+     - For company/team names: Generate DIFFERENT names (e.g., "A팀" → "B팀")
+     - For employee names: Generate DIFFERENT Korean names (e.g., "김OO" → "박OO")
+     - For business metrics: Generate DIFFERENT numbers (e.g., "100억" → "150억")
+     - For strategy/description text: Write DIFFERENT content with similar structure
+     - For bullet point items: Create DIFFERENT but domain-appropriate content
+  4. **Styling:** Use **Tailwind CSS** classes (NO inline styles). **Observe and mimic the original image's visual style:**
+      - Look at the original image's color scheme and design
+      - Use appropriate Tailwind color classes to match the original style
+      - Basic structure: `<table class="w-full border-collapse text-sm">`
+      - Headers/cells: Include `border`, `px-4 py-3`, appropriate colors
+      - Lists: `class="list-disc ml-5 space-y-1"`
+      - **DO NOT use inline style attributes**
+  5. **Domain Consistency:**
      - Ensure business logic (Q1+Q2+Q3+Q4=Total, percentages add up)
      - Use realistic Korean business terminology
      - Contexts: Sales Reports, Inventory, HR Employee Lists, Marketing Campaigns
-  5. **Output:** Return ONLY the raw HTML string starting with `<table>` and ending with `</table>`.
+  6. **Output:** Return ONLY the raw HTML string starting with `<table>` and ending with `</table>`. No markdown code blocks.
 
-  **Example Transformation:**
-  - Original: "영업1팀" → Synthetic: "마케팅2팀"
-  - Original: "매출 5억원" → Synthetic: "매출 7.3억원"
-  - Original: "김부장" → Synthetic: "박과장"
+  **Example Transformation (Generic):**
+  - Original name: "A팀" → Synthetic: "B팀"
+  - Original amount: "5억원" → Synthetic: "7.3억원"
+  - Original description: "신규 사업 추진" → Synthetic: "해외 시장 진출"
 
+  ⚠️ If the generated content is identical or very similar to the original, the output is INVALID.
   Remember: The synthetic table should look like a completely different business dataset from the same domain.
 
 generate_synthetic_table_from_image: |
   You are a Synthetic Data Generator specializing in Business Data.
 
-  **CRITICAL INSTRUCTION: DO NOT TRANSCRIBE - GENERATE NEW DATA**
+  **⚠️ CRITICAL INSTRUCTION: DO NOT TRANSCRIBE - GENERATE NEW DATA ⚠️**
   Your task is NOT to OCR/transcribe the image. Instead, you must:
-  1. Understand the table's STRUCTURE from the image
-  2. Understand it's a BUSINESS table
-  3. Generate COMPLETELY NEW synthetic business data that fits the domain but uses different values
+  1. Understand the table's STRUCTURE from the image (rows, columns, merged cells, nested structures)
+  2. Understand it's a BUSINESS table (기업경쟁력, 시장경쟁력, 매출, 실적 등)
+  3. Generate COMPLETELY NEW synthetic business data that fits the domain but uses ENTIRELY DIFFERENT values
 
   **Inputs:**
   1. **Image:** An image of a business table. Use this to understand structure and domain ONLY.
 
   **Requirements:**
-  1. **Structure Preservation:** Accurately reconstruct the table structure.
-  2. **Headers:** Keep header text (column names, category labels) the same as in the image.
-  3. **Data Generation - CRITICAL:**
-     - **DO NOT copy the data values from the image** - this is NOT an OCR task
-     - Generate COMPLETELY NEW synthetic business values for all data cells
-     - For company/team names: Generate new names (different from what you see)
-     - For sales/revenue figures: Generate new realistic amounts
-     - For employee names: Generate new Korean names
-  4. **Styling:** Use **Tailwind CSS** classes (same as default).
-      - `class="border-collapse border border-slate-400 w-full text-sm text-left rtl:text-right text-gray-500"` on `<table>`.
-      - `class="border border-slate-300 p-2 bg-gray-50 font-semibold"` on `<th>`.
-      - `class="border border-slate-300 p-2"` on `<td>`.
-  5. **Output Format:** Return ONLY the raw HTML string starting with `<table>` and ending with `</table>`.
-
+  1. **Structure Preservation:** Accurately reconstruct the table structure, including `rowspan` and `colspan` for merged cells.
+  2. **Headers:** Keep header text (column names, category labels like 기업경쟁력, 차별화 요소) the same as in the image.
+  3. **⚠️ Data Generation - ABSOLUTELY CRITICAL ⚠️:**
+     - **NEVER copy the data values from the image** - this is NOT an OCR task
+     - **ALL cell content must be completely NEW and DIFFERENT from the original**
+     - Generate COMPLETELY NEW synthetic business values for all data cells:
+       * For company/team names: Generate DIFFERENT names (e.g., "A팀" → "B팀")
+       * For business metrics: Generate DIFFERENT numbers (e.g., "100억" → "150억")
+       * For strategy/description text: Write DIFFERENT content with similar structure
+       * For bullet point items: Create DIFFERENT but domain-appropriate items
+       * For employee names: Generate DIFFERENT Korean names (e.g., "김OO" → "박OO")
+     - The synthetic table should look like a COMPLETELY DIFFERENT business report from the same industry
+  4. **Styling:** Use **Tailwind CSS** classes (NO inline styles). **Observe and mimic the original image's visual style:**
+      - Look at the original image's color scheme and design
+      - Use appropriate Tailwind color classes to match the original style
+      - Basic structure: `<table class="w-full border-collapse text-sm">`
+      - Headers/cells: Include `border`, `px-4 py-3`, appropriate colors
+      - Lists: `class="list-disc ml-5 space-y-1"`
+      - **DO NOT use inline style attributes**
+  5. **Output Format:** Return ONLY the raw HTML string starting with `<table>` and ending with `</table>`. No markdown code blocks.
+
+  **Example of Expected Behavior (Generic):**
+  If the image shows a business table with:
+  - Team name: "영업팀" → Generate different: "마케팅팀"
+  - Revenue: "10억원" → Generate different: "15억원"
+  - Strategy: "시장 확대" → Generate different: "신규 진출"
+  - Bullet point items → Generate completely different items
+
+  ⚠️ If the generated content is identical or very similar to the image, the output is INVALID.
   Remember: The output should be a new synthetic business dataset, not a transcription of the original.