Update app.py

app.py

@@ -2,6 +2,7 @@ import os
 import io
 import json
 import hashlib
+import pathlib
 import gradio as gr
 
 from pipelines.openai_ingest import (
@@ -19,40 +20,18 @@ from pipelines.utils import detect_filetype, load_doc_text
 
 APP_TITLE = "候補者インテーク & レジュメ標準化（OpenAI版）"
 
-# Gradio v4 Filesの入力を安全にハンドリング
-def _iter_files(files):
-    """
-    Gradio v4:
-      - type="filepath": files は 文字列パスのリスト
-      - type="binary"  : files は {name: str, data: bytes} の辞書リスト
-    互換目的で file-like も許容。
+
+def _read_file_obj_or_path(f):
+    """gr.Files v4 用の堅牢リーダ。
+    - type="filepath" の場合: f は str/Path
+    - type="binary" の場合: f は UploadedFile ライク（.name/.read()）
+    戻り値: (filename, bytes)
     """
-    for f in files:
-        # filepath
-        if isinstance(f, str):
-            path = f
-            name = os.path.basename(path)
-            with open(path, "rb") as fh:
-                data = fh.read()
-            yield name, data
-            continue
-        # binary
-        if isinstance(f, dict) and "name" in f and "data" in f:
-            name = os.path.basename(f["name"] or "uploaded")
-            data = f["data"]
-            yield name, data
-            continue
-        # file-like
-        if hasattr(f, "read"):
-            try:
-                name = os.path.basename(getattr(f, "name", "uploaded"))
-            except Exception:
-                name = "uploaded"
-            data = f.read()
-            yield name, data
-            continue
-        # 不明形式はスキップ
-        continue
+    if isinstance(f, (str, pathlib.Path)):
+        p = pathlib.Path(f)
+        return p.name, p.read_bytes()
+    # UploadedFile 互換
+    return getattr(f, "name", "uploaded"), f.read()
 
 
 def process_resumes(files, candidate_id: str, additional_notes: str = ""):
@@ -62,20 +41,30 @@ def process_resumes(files, candidate_id: str, additional_notes: str = ""):
     partial_records = []
     raw_texts = []
 
-    for fname, raw_bytes in _iter_files(files):
+    for f in files:
+        fname, raw_bytes = _read_file_obj_or_path(f)
         filetype = detect_filetype(fname, raw_bytes)
 
         # 1) テキスト抽出：画像/PDFはOpenAI Vision OCR、docx/txtは生文面＋OpenAI整形
         if filetype in {"pdf", "image"}:
             text = extract_text_with_openai(raw_bytes, filename=fname, filetype=filetype)
-        else:
+        elif filetype in {"docx", "txt"}:
             base_text = load_doc_text(filetype, raw_bytes)
+            # 生テキストをOpenAIへ渡し、整形した全文を返す
+            text = extract_text_with_openai(base_text.encode("utf-8"), filename=fname, filetype="txt")
+        else:
+            # フォールバック：そのままテキスト化を試み、ダメならエラーを添えて続行
+            try:
+                base_text = raw_bytes.decode("utf-8", errors="ignore")
+            except Exception:
+                base_text = ""
             text = extract_text_with_openai(base_text.encode("utf-8"), filename=fname, filetype="txt")
 
         raw_texts.append({"filename": fname, "text": text})
 
-        # 2) OpenAIでセクション構造化 → ルール正規化
+        # 2) OpenAIでセクション構造化
         structured = structure_with_openai(text)
+        # 念のためルールベース正規化も適用（期間抽出など補助）
         normalized = normalize_resume({
             "work_experience": structured.get("work_experience_raw", ""),
             "education": structured.get("education_raw", ""),
@@ -92,7 +81,7 @@ def process_resumes(files, candidate_id: str, additional_notes: str = ""):
     # 3) 統合（複数ファイル→1候補者）
     merged = merge_normalized_records([r["normalized"] for r in partial_records])
 
-    # 4) スキル抽出
+    # 4) スキル抽出（辞書/正規表現）
     merged_text = "\n\n".join([r["text"] for r in partial_records])
     skills = extract_skills(merged_text, {
         "work_experience": merged.get("raw_sections", {}).get("work_experience", ""),
@@ -101,20 +90,20 @@ def process_resumes(files, candidate_id: str, additional_notes: str = ""):
         "skills": ", ".join(merged.get("skills", [])),
     })
 
-    # 5) 匿名化 → 匿名PDF生成
+    # 5) 匿名化
     anonymized_text, anon_map = anonymize_text(merged_text)
     anon_pdf_bytes = render_anonymized_pdf(anonymized_text)
 
     # 6) 品質スコア
     score = compute_quality_score(merged_text, merged)
 
-    # 7) 要約
+    # 7) 要約（300/100/1文）
     summaries = summarize_with_openai(merged_text)
 
     # 8) 構造化出力
-    cid = candidate_id or hashlib.sha256(merged_text.encode("utf-8")).hexdigest()[:16]
+    candidate_id_final = candidate_id or hashlib.sha256(merged_text.encode("utf-8")).hexdigest()[:16]
     result_json = {
-        "candidate_id": cid,
+        "candidate_id": candidate_id_final,
         "files": [r["source"] for r in partial_records],
         "merged": merged,
         "skills": skills,
@@ -124,28 +113,30 @@ def process_resumes(files, candidate_id: str, additional_notes: str = ""):
         "notes": additional_notes,
     }
 
-    # 9) HF Datasets 保存（任意）
+    # 9) HF Datasets 保存
     dataset_repo = os.environ.get("DATASET_REPO")
     commit_info = None
     if dataset_repo:
+        file_hash = result_json["candidate_id"]
         commit_info = persist_to_hf(
             dataset_repo=dataset_repo,
             record=result_json,
             anon_pdf_bytes=anon_pdf_bytes,
-            parquet_path=f"candidates/{cid}.parquet",
-            json_path=f"candidates/{cid}.json",
-            pdf_path=f"candidates/{cid}.anon.pdf",
+            parquet_path=f"candidates/{file_hash}.parquet",
+            json_path=f"candidates/{file_hash}.json",
+            pdf_path=f"candidates/{file_hash}.anon.pdf",
         )
 
-    anon_pdf = (f"{cid}.anon.pdf", anon_pdf_bytes)
+    anon_pdf = (candidate_id_final + ".anon.pdf", anon_pdf_bytes)
 
+    # 重要: gr.JSON を避けるため JSON文字列で返す
     return (
         json.dumps(result_json, ensure_ascii=False, indent=2),
-        json.dumps(skills, ensure_ascii=False, indent=2),   # ← out_skillsをCodeにしたのでJSON文字列を返す
+        json.dumps(skills, ensure_ascii=False, indent=2),
         json.dumps(score, ensure_ascii=False, indent=2),
-        summaries["300chars"],
-        summaries["100chars"],
-        summaries["onesent"],
+        summaries.get("300chars", ""),
+        summaries.get("100chars", ""),
+        summaries.get("onesent", ""),
         anon_pdf,
         json.dumps(commit_info or {"status": "skipped (DATASET_REPO not set)"}, ensure_ascii=False, indent=2),
     )
@@ -159,21 +150,22 @@ with gr.Blocks(title=APP_TITLE) as demo:
             label="レジュメ類 (PDF/画像/Word/テキスト) 複数可",
             file_count="multiple",
             file_types=[".pdf", ".png", ".jpg", ".jpeg", ".tiff", ".bmp", ".docx", ".txt"],
-            type="filepath"  # ← ここを修正
+            type="filepath",  # v4仕様: 'file' は無効
         )
         candidate_id = gr.Textbox(label="候補者ID（任意。未入力なら自動生成）")
     notes = gr.Textbox(label="補足メモ（任意）", lines=3)
 
-    run_btn = gr.Button("実行", variant="primary")
+    run_btn = gr.Button("実行")
 
     with gr.Tab("構造化JSON"):
         out_json = gr.Code(label="統合出力 (JSON)")
 
     with gr.Tab("抽出スキル"):
-        out_skills = gr.Code(label="スキル一覧 (JSON)")  # ← JSON Schema問題回避
+        # gr.JSON は Gradio v4 の API 情報生成で例外を起こすケースがあるため避ける
+        out_skills = gr.Code(label="スキル一覧(JSON)")
 
     with gr.Tab("品質スコア"):
-        out_score = gr.Code(label="品質評価 (JSON)")
+        out_score = gr.Code(label="品質評価(JSON)")
 
     with gr.Tab("要約 (300/100/1文)"):
         out_sum_300 = gr.Textbox(label="300字要約")
@@ -194,4 +186,9 @@ with gr.Blocks(title=APP_TITLE) as demo:
 
 
 if __name__ == "__main__":
-    demo.launch()
+    try:
+        # 通常はローカル公開で起動
+        demo.launch(server_name="0.0.0.0", server_port=7860)
+    except ValueError:
+        # 実行環境が localhost にアクセスできない場合のフォールバック
+        demo.launch(share=True)