Update app.py

app.py

@@ -5,7 +5,7 @@ from pathlib import Path
 import google.generativeai as genai
 import tempfile
 import base64
-from concurrent.futures import ThreadPoolExecutor, as_completed  # as_completed を追加
+from concurrent.futures import ThreadPoolExecutor, as_completed
 import logging
 import time
 
@@ -14,7 +14,7 @@ logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(
 
 # 環境変数から設定を読み込む
 GOOGLE_API_KEY = os.environ.get("GOOGLE_API_KEY")
-MODEL_NAME = os.environ.get("MODEL_NAME", "gemini-1.5-pro")  # デフォルトモデルも設定
+MODEL_NAME = os.environ.get("MODEL_NAME", "gemini-1.5-pro")
 
 if not GOOGLE_API_KEY:
     raise ValueError("環境変数 'GOOGLE_API_KEY' が設定されていません。")
@@ -46,7 +46,7 @@ def split_pdf(pdf_path, output_dir, pages_per_chunk=5):
         output_pdf.save(output_path)
         output_pdf.close()
 
-        split_pdfs.append((start_page, output_path))  # (開始ページ, パス) のタプルで保存
+        split_pdfs.append((start_page, output_path))
 
     pdf_document.close()
     return split_pdfs
@@ -60,43 +60,69 @@ def encode_pdf_to_base64(pdf_path):
 
 def ocr_pdf_with_gemini(pdf_path, model_name):
     """GeminiモデルでPDFをOCRしてマークダウンに変換する関数"""
-    # PDFをbase64エンコード
     pdf_base64 = encode_pdf_to_base64(pdf_path)
-
-    # Geminiモデルの設定
     model = genai.GenerativeModel(model_name)
 
-    # プロンプトの設定
     prompt = """
-    Extract the text content from this PDF using OCR and output it in a well-structured Markdown format.
-    Focus solely on the text extraction; do not include any conversational elements, greetings, or additional explanations.
-    Only provide the extracted text in Markdown.
-
-    Pay attention to the following:
-    - Use appropriate Markdown heading syntax (#, ##, ###, etc.) for headings.
-    - Convert bulleted lists correctly.
-    - Convert tables to Markdown table format.
-    - Maintain the paragraph structure.
-    - Clean up any extra line breaks or spaces.
-    - For images, use the format [Image: Description of the content].
+    You are an expert document processing assistant. Your task is to extract text from the provided PDF using OCR and convert it into a highly readable and visually appealing Markdown format.
+
+    **Crucial Instructions:**
+
+    1.  **Consistency:** Maintain a consistent Markdown style (headings, lists, tables, font sizes, etc.) throughout the entire output, even if the PDF is split into multiple parts. Avoid variations in formatting between different sections.
+    2.  **Visual Fidelity:** Strive to reproduce the original visual appearance of the PDF as closely as possible in Markdown. Pay close attention to:
+        *   **Tables:** Recreate tables with accurate borders, cell alignment, and row/column spans.  Use extended Markdown table syntax if needed.
+        *   **Lists:** Preserve the original list style (numbered, bulleted, nested).
+        *   **Font Sizes:**  Use Markdown heading levels (#, ##, ###) and relative font size adjustments (if possible in your Markdown flavor) to approximate the visual hierarchy of the original text.  Larger text should generally correspond to higher-level headings.
+        *   **Layout:** Try to maintain the spatial relationships between elements (e.g., paragraphs, images). If there are multiple columns, consider representing that structure in Markdown, perhaps using tables.
+    3.  **OCR Correction:** The OCR process may introduce errors (typos, misspellings, incorrect characters).  Apply your language understanding capabilities to correct these errors and produce grammatically correct and semantically meaningful text.  Do *not* blindly output the raw OCR result if it contains obvious mistakes.
+    4.  **Content Filtering (Screenshots):** If the PDF primarily contains screenshots (e.g., of software interfaces), focus *exclusively* on extracting text from the *main content area* of the screenshots.  *Do not* include text from:
+        *   Window title bars
+        *   Operating system toolbars (e.g., Windows taskbar)
+        *   Menu bars *unless* they are directly related to the primary content (e.g. describing a software's menu options)
+        *   Any other UI elements that are not part of the core content being displayed.
+    5. **Accuracy and Clarity:** Prioritize providing accurate and clear information to the user. Do not simply reproduce OCR output verbatim if it is nonsensical or misleading. Use your understanding of the content to present information in a user-friendly way.
+    6. **Output:** Only provide the extracted text in Markdown.
+
+    **Example (Illustrative - Adapt to the specific PDF):**
+
+    **Input PDF (Screenshot of a webpage):**
+
+    ```
+    [Screenshot of a webpage with a large heading "Welcome", a paragraph of text, a bulleted list, and a table.]
+    ```
+
+    **Desired Markdown Output:**
+
+    ```markdown
+    # Welcome
+
+    This is a paragraph of introductory text. It explains the purpose of the webpage and provides some context.
+
+    *   This is the first bullet point.
+    *   This is the second bullet point.
+        *   This is a nested bullet point.
+
+    | Feature      | Description                | Price   |
+    |--------------|----------------------------|---------|
+    | Feature A    | Description of Feature A  | $10     |
+    | Feature B    | Description of Feature B  | $20     |
+
+    ```
+
+    **Do NOT include text like "File Edit View" (from a menu bar) or "[X] Minimize Maximize Close" (from a window title bar).**
     """
 
-    # PDFをGeminiに送信
     try:
         response = model.generate_content(
             [
                 prompt,
                 {"mime_type": "application/pdf", "data": pdf_base64}
             ],
-            # generation_configで不要な応答を抑制 (モデルによっては効かない場合あり)
             generation_config=genai.types.GenerationConfig(
-                candidate_count=1,  # 候補を1つに絞る
-                max_output_tokens=8192, # 必要に応じてトークン数を調整
-                # stop_sequences=["."],  # 句点で生成を停止 (モデルによっては無効)
+                candidate_count=1,
+                max_output_tokens=8192,
             )
-
         )
-        # 結果を返す
         return response.text
     except Exception as e:
         logging.error(f"Error during Gemini API call: {e}")
@@ -108,18 +134,15 @@ def process_pdf(pdf_file, progress=gr.Progress()):
     """PDFファイルを処理するメイン関数"""
     logging.info(f"Received file: {pdf_file.name if hasattr(pdf_file, 'name') else pdf_file}")
 
-    # 一時ディレクトリを作成
     with tempfile.TemporaryDirectory() as temp_dir:
         temp_pdf_path = pdf_file.name
         logging.info(f"Temporary PDF path: {temp_pdf_path}")
 
-        # PDFを分割
-        split_pdf_paths = split_pdf(temp_pdf_path, temp_dir)  # [(start_page, path), ...]
+        split_pdf_paths = split_pdf(temp_pdf_path, temp_dir)
         logging.info(f"Split PDF paths: {split_pdf_paths}")
         progress(0.2, desc="PDFを分割中...")
 
-        # 並列処理でOCR変換
-        markdown_results = {}  # {start_page: markdown_text, ...}  辞書に変更
+        markdown_results = {}
         with ThreadPoolExecutor() as executor:
             futures = {executor.submit(ocr_pdf_with_gemini, path, MODEL_NAME): start_page for start_page, path in split_pdf_paths}
             completed_count = 0
@@ -137,7 +160,6 @@ def process_pdf(pdf_file, progress=gr.Progress()):
         logging.info(f"Markdown results length: {len(markdown_results)}")
         progress(0.8, desc="結果を結合中...")
 
-        # 結果を結合 (開始ページ番号でソート)
         combined_markdown = "\n\n".join(markdown_results[page] for page in sorted(markdown_results.keys()))
         progress(1.0, desc="完了")
         time.sleep(0.5)
@@ -145,7 +167,6 @@ def process_pdf(pdf_file, progress=gr.Progress()):
         return combined_markdown
 
 
-# Gradioインターフェースの作成
 def create_interface():
     with gr.Blocks() as demo:
         gr.Markdown("# PDF OCR & マークダウン変換ツール")
@@ -166,7 +187,6 @@ def create_interface():
             copy_btn = gr.Button("クリップボードにコピー")
             download_btn = gr.Button("ダウンロード")
 
-        # JavaScriptコードを埋め込むためのHTMLコンポーネント
         js_code = gr.HTML(
             """
             <script>
@@ -178,17 +198,14 @@ def create_interface():
             visible=False,
         )
 
-        # js_code が読み込まれた後に styleButton 関数を実行
         pdf_input.upload(None, [], [], js="styleButton")
 
-
         convert_btn.click(
             fn=process_pdf,
             inputs=pdf_input,
             outputs=markdown_output
         )
 
-        # クリップボードにコピー
         copy_btn.click(
             None,
             markdown_output,
@@ -196,7 +213,6 @@ def create_interface():
             js=f"(x) => {{ navigator.clipboard.writeText(x); }}",
         )
 
-        # ダウンロード
         download_btn.click(
             None,
             markdown_output,
@@ -216,7 +232,6 @@ def create_interface():
     return demo
 
 
-# アプリの起動
 if __name__ == "__main__":
     demo = create_interface()
     demo.launch()