Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,5 +1,4 @@
|
|
| 1 |
import os
|
| 2 |
-
import io
|
| 3 |
import json
|
| 4 |
import hashlib
|
| 5 |
import gradio as gr
|
|
@@ -12,55 +11,64 @@ from pipelines.openai_ingest import (
|
|
| 12 |
from pipelines.parsing import normalize_resume
|
| 13 |
from pipelines.merge import merge_normalized_records
|
| 14 |
from pipelines.skills import extract_skills
|
| 15 |
-
from pipelines.anonymize import anonymize_text, render_anonymized_pdf
|
| 16 |
from pipelines.scoring import compute_quality_score
|
| 17 |
-
from pipelines.storage import persist_to_hf
|
| 18 |
from pipelines.utils import detect_filetype, load_doc_text
|
| 19 |
|
| 20 |
-
APP_TITLE = "候補者インテーク &
|
| 21 |
|
| 22 |
|
| 23 |
-
def process_resumes(
|
| 24 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 25 |
raise gr.Error("少なくとも1ファイルをアップロードしてください。")
|
| 26 |
|
| 27 |
partial_records = []
|
| 28 |
raw_texts = []
|
| 29 |
|
| 30 |
-
for
|
| 31 |
-
|
| 32 |
-
|
|
|
|
|
|
|
|
|
|
| 33 |
|
| 34 |
# 1) テキスト抽出:画像/PDFはOpenAI Vision OCR、docx/txtは生文面+OpenAI整形
|
| 35 |
if filetype in {"pdf", "image"}:
|
| 36 |
-
text = extract_text_with_openai(raw_bytes, filename=
|
| 37 |
else:
|
| 38 |
base_text = load_doc_text(filetype, raw_bytes)
|
| 39 |
-
|
| 40 |
-
text = extract_text_with_openai(base_text.encode("utf-8"), filename=f.name, filetype="txt")
|
| 41 |
|
| 42 |
-
raw_texts.append({"filename":
|
| 43 |
|
| 44 |
-
# 2) OpenAIでセクション構造化
|
| 45 |
structured = structure_with_openai(text)
|
| 46 |
-
# 念のためルールベース正規化も適用(期間抽出など補助)
|
| 47 |
normalized = normalize_resume({
|
| 48 |
"work_experience": structured.get("work_experience_raw", ""),
|
| 49 |
"education": structured.get("education_raw", ""),
|
| 50 |
"certifications": structured.get("certifications_raw", ""),
|
| 51 |
"skills": ", ".join(structured.get("skills_list", [])),
|
| 52 |
})
|
|
|
|
| 53 |
partial_records.append({
|
| 54 |
-
"source":
|
| 55 |
"text": text,
|
| 56 |
"structured": structured,
|
| 57 |
"normalized": normalized,
|
| 58 |
})
|
| 59 |
|
| 60 |
-
#
|
| 61 |
merged = merge_normalized_records([r["normalized"] for r in partial_records])
|
| 62 |
|
| 63 |
-
#
|
| 64 |
merged_text = "\n\n".join([r["text"] for r in partial_records])
|
| 65 |
skills = extract_skills(merged_text, {
|
| 66 |
"work_experience": merged.get("raw_sections", {}).get("work_experience", ""),
|
|
@@ -69,65 +77,43 @@ def process_resumes(files, candidate_id: str, additional_notes: str = ""):
|
|
| 69 |
"skills": ", ".join(merged.get("skills", [])),
|
| 70 |
})
|
| 71 |
|
| 72 |
-
# 5) 匿名化
|
| 73 |
-
anonymized_text, anon_map = anonymize_text(merged_text)
|
| 74 |
-
anon_pdf_bytes = render_anonymized_pdf(anonymized_text)
|
| 75 |
-
|
| 76 |
# 6) 品質スコア
|
| 77 |
score = compute_quality_score(merged_text, merged)
|
| 78 |
|
| 79 |
# 7) 要約(300/100/1文)
|
| 80 |
summaries = summarize_with_openai(merged_text)
|
| 81 |
|
| 82 |
-
# 8)
|
| 83 |
result_json = {
|
| 84 |
"candidate_id": candidate_id or hashlib.sha256(merged_text.encode("utf-8")).hexdigest()[:16],
|
| 85 |
-
"files": [
|
| 86 |
"merged": merged,
|
| 87 |
"skills": skills,
|
| 88 |
"quality_score": score,
|
| 89 |
"summaries": summaries,
|
| 90 |
-
"anonymization_map": anon_map,
|
| 91 |
"notes": additional_notes,
|
| 92 |
}
|
| 93 |
|
| 94 |
-
# 9) HF Datasets 保存
|
| 95 |
-
dataset_repo = os.environ.get("DATASET_REPO")
|
| 96 |
-
commit_info = None
|
| 97 |
-
if dataset_repo:
|
| 98 |
-
file_hash = result_json["candidate_id"]
|
| 99 |
-
commit_info = persist_to_hf(
|
| 100 |
-
dataset_repo=dataset_repo,
|
| 101 |
-
record=result_json,
|
| 102 |
-
anon_pdf_bytes=anon_pdf_bytes,
|
| 103 |
-
parquet_path=f"candidates/{file_hash}.parquet",
|
| 104 |
-
json_path=f"candidates/{file_hash}.json",
|
| 105 |
-
pdf_path=f"candidates/{file_hash}.anon.pdf",
|
| 106 |
-
)
|
| 107 |
-
|
| 108 |
-
anon_pdf = (result_json["candidate_id"] + ".anon.pdf", anon_pdf_bytes)
|
| 109 |
-
|
| 110 |
return (
|
| 111 |
-
json.dumps(result_json, ensure_ascii=False, indent=2),
|
| 112 |
-
json.dumps(skills, ensure_ascii=False, indent=2),
|
| 113 |
-
json.dumps(score, ensure_ascii=False, indent=2),
|
| 114 |
-
summaries
|
| 115 |
-
summaries
|
| 116 |
-
summaries
|
| 117 |
-
anon_pdf,
|
| 118 |
-
json.dumps(commit_info or {"status": "skipped (DATASET_REPO not set)"}, ensure_ascii=False, indent=2),
|
| 119 |
)
|
| 120 |
|
| 121 |
|
| 122 |
with gr.Blocks(title=APP_TITLE) as demo:
|
| 123 |
-
gr.Markdown(f"# {APP_TITLE}\
|
| 124 |
|
| 125 |
with gr.Row():
|
|
|
|
| 126 |
in_files = gr.Files(
|
| 127 |
label="レジュメ類 (PDF/画像/Word/テキスト) 複数可",
|
| 128 |
file_count="multiple",
|
| 129 |
file_types=[".pdf", ".png", ".jpg", ".jpeg", ".tiff", ".bmp", ".docx", ".txt"],
|
| 130 |
-
type="
|
| 131 |
)
|
| 132 |
candidate_id = gr.Textbox(label="候補者ID(任意。未入力なら自動生成)")
|
| 133 |
notes = gr.Textbox(label="補足メモ(任意)", lines=3)
|
|
@@ -138,30 +124,24 @@ with gr.Blocks(title=APP_TITLE) as demo:
|
|
| 138 |
out_json = gr.Code(label="統合出力 (JSON)")
|
| 139 |
|
| 140 |
with gr.Tab("抽出スキル"):
|
| 141 |
-
# ★
|
| 142 |
out_skills = gr.Code(label="スキル一覧 (JSON)")
|
| 143 |
|
| 144 |
with gr.Tab("品質スコア"):
|
| 145 |
-
out_score = gr.Code(label="品質評価")
|
| 146 |
|
| 147 |
with gr.Tab("要約 (300/100/1文)"):
|
| 148 |
out_sum_300 = gr.Textbox(label="300字要約")
|
| 149 |
out_sum_100 = gr.Textbox(label="100字要約")
|
| 150 |
out_sum_1 = gr.Textbox(label="1文要約")
|
| 151 |
|
| 152 |
-
with gr.Tab("匿名PDF"):
|
| 153 |
-
out_pdf = gr.File(label="匿名PDFダウンロード")
|
| 154 |
-
|
| 155 |
-
with gr.Tab("Datasets 保存ログ"):
|
| 156 |
-
out_commit = gr.Code(label="コミット情報")
|
| 157 |
-
|
| 158 |
run_btn.click(
|
| 159 |
process_resumes,
|
| 160 |
inputs=[in_files, candidate_id, notes],
|
| 161 |
-
outputs=[out_json, out_skills, out_score, out_sum_300, out_sum_100, out_sum_1
|
| 162 |
)
|
| 163 |
|
| 164 |
|
| 165 |
if __name__ == "__main__":
|
| 166 |
-
#
|
| 167 |
demo.launch(share=True, server_name="0.0.0.0", server_port=7860)
|
|
|
|
| 1 |
import os
|
|
|
|
| 2 |
import json
|
| 3 |
import hashlib
|
| 4 |
import gradio as gr
|
|
|
|
| 11 |
from pipelines.parsing import normalize_resume
|
| 12 |
from pipelines.merge import merge_normalized_records
|
| 13 |
from pipelines.skills import extract_skills
|
|
|
|
| 14 |
from pipelines.scoring import compute_quality_score
|
|
|
|
| 15 |
from pipelines.utils import detect_filetype, load_doc_text
|
| 16 |
|
| 17 |
+
APP_TITLE = "候補者インテーク & レジュメ標準化(安定版・最小機能)"
|
| 18 |
|
| 19 |
|
| 20 |
+
def process_resumes(filepaths, candidate_id: str, additional_notes: str = ""):
|
| 21 |
+
"""
|
| 22 |
+
安定版(最小機能):
|
| 23 |
+
- OCR/抽出 → 構造化 → 正規化 → マージ → スキル抽出 → 要約 → 品質スコア
|
| 24 |
+
- 匿名化PDF生成・HF Datasets保存は一旦オフ
|
| 25 |
+
入力:
|
| 26 |
+
filepaths: gr.Files(type="filepath") からのパス配列
|
| 27 |
+
出力:
|
| 28 |
+
JSON文字列 / スキルJSON文字列 / スコアJSON文字列 / 要約(3種)
|
| 29 |
+
"""
|
| 30 |
+
if not filepaths:
|
| 31 |
raise gr.Error("少なくとも1ファイルをアップロードしてください。")
|
| 32 |
|
| 33 |
partial_records = []
|
| 34 |
raw_texts = []
|
| 35 |
|
| 36 |
+
for path in filepaths:
|
| 37 |
+
# パスからバイトとファイル名を取得
|
| 38 |
+
with open(path, "rb") as rf:
|
| 39 |
+
raw_bytes = rf.read()
|
| 40 |
+
fname = os.path.basename(path)
|
| 41 |
+
filetype = detect_filetype(fname, raw_bytes)
|
| 42 |
|
| 43 |
# 1) テキスト抽出:画像/PDFはOpenAI Vision OCR、docx/txtは生文面+OpenAI整形
|
| 44 |
if filetype in {"pdf", "image"}:
|
| 45 |
+
text = extract_text_with_openai(raw_bytes, filename=fname, filetype=filetype)
|
| 46 |
else:
|
| 47 |
base_text = load_doc_text(filetype, raw_bytes)
|
| 48 |
+
text = extract_text_with_openai(base_text.encode("utf-8"), filename=fname, filetype="txt")
|
|
|
|
| 49 |
|
| 50 |
+
raw_texts.append({"filename": fname, "text": text})
|
| 51 |
|
| 52 |
+
# 2) OpenAIでセクション構造化 → 3) 正規化
|
| 53 |
structured = structure_with_openai(text)
|
|
|
|
| 54 |
normalized = normalize_resume({
|
| 55 |
"work_experience": structured.get("work_experience_raw", ""),
|
| 56 |
"education": structured.get("education_raw", ""),
|
| 57 |
"certifications": structured.get("certifications_raw", ""),
|
| 58 |
"skills": ", ".join(structured.get("skills_list", [])),
|
| 59 |
})
|
| 60 |
+
|
| 61 |
partial_records.append({
|
| 62 |
+
"source": fname,
|
| 63 |
"text": text,
|
| 64 |
"structured": structured,
|
| 65 |
"normalized": normalized,
|
| 66 |
})
|
| 67 |
|
| 68 |
+
# 4) 複数ファイルの統合
|
| 69 |
merged = merge_normalized_records([r["normalized"] for r in partial_records])
|
| 70 |
|
| 71 |
+
# 5) スキル抽出(辞書/正規表現)
|
| 72 |
merged_text = "\n\n".join([r["text"] for r in partial_records])
|
| 73 |
skills = extract_skills(merged_text, {
|
| 74 |
"work_experience": merged.get("raw_sections", {}).get("work_experience", ""),
|
|
|
|
| 77 |
"skills": ", ".join(merged.get("skills", [])),
|
| 78 |
})
|
| 79 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 80 |
# 6) 品質スコア
|
| 81 |
score = compute_quality_score(merged_text, merged)
|
| 82 |
|
| 83 |
# 7) 要約(300/100/1文)
|
| 84 |
summaries = summarize_with_openai(merged_text)
|
| 85 |
|
| 86 |
+
# 8) 構造化出力(文字列にして Code に安全に渡す)
|
| 87 |
result_json = {
|
| 88 |
"candidate_id": candidate_id or hashlib.sha256(merged_text.encode("utf-8")).hexdigest()[:16],
|
| 89 |
+
"files": [os.path.basename(p) for p in filepaths],
|
| 90 |
"merged": merged,
|
| 91 |
"skills": skills,
|
| 92 |
"quality_score": score,
|
| 93 |
"summaries": summaries,
|
|
|
|
| 94 |
"notes": additional_notes,
|
| 95 |
}
|
| 96 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 97 |
return (
|
| 98 |
+
json.dumps(result_json, ensure_ascii=False, indent=2), # out_json -> Code
|
| 99 |
+
json.dumps(skills, ensure_ascii=False, indent=2), # out_skills -> Code
|
| 100 |
+
json.dumps(score, ensure_ascii=False, indent=2), # out_score -> Code
|
| 101 |
+
summaries.get("300chars", ""),
|
| 102 |
+
summaries.get("100chars", ""),
|
| 103 |
+
summaries.get("onesent", ""),
|
|
|
|
|
|
|
| 104 |
)
|
| 105 |
|
| 106 |
|
| 107 |
with gr.Blocks(title=APP_TITLE) as demo:
|
| 108 |
+
gr.Markdown(f"# {APP_TITLE}\nOpenAIでOCR/構造化/要約→統合→スコア(匿名化・HF保存なしの安定版)")
|
| 109 |
|
| 110 |
with gr.Row():
|
| 111 |
+
# ★ Gradio v4仕様: Filesは type='filepath' or 'binary' のみ
|
| 112 |
in_files = gr.Files(
|
| 113 |
label="レジュメ類 (PDF/画像/Word/テキスト) 複数可",
|
| 114 |
file_count="multiple",
|
| 115 |
file_types=[".pdf", ".png", ".jpg", ".jpeg", ".tiff", ".bmp", ".docx", ".txt"],
|
| 116 |
+
type="filepath"
|
| 117 |
)
|
| 118 |
candidate_id = gr.Textbox(label="候補者ID(任意。未入力なら自動生成)")
|
| 119 |
notes = gr.Textbox(label="補足メモ(任意)", lines=3)
|
|
|
|
| 124 |
out_json = gr.Code(label="統合出力 (JSON)")
|
| 125 |
|
| 126 |
with gr.Tab("抽出スキル"):
|
| 127 |
+
# ★ JSONスキーマ推論の例外回避のため Code を使用
|
| 128 |
out_skills = gr.Code(label="スキル一覧 (JSON)")
|
| 129 |
|
| 130 |
with gr.Tab("品質スコア"):
|
| 131 |
+
out_score = gr.Code(label="品質評価 (JSON)")
|
| 132 |
|
| 133 |
with gr.Tab("要約 (300/100/1文)"):
|
| 134 |
out_sum_300 = gr.Textbox(label="300字要約")
|
| 135 |
out_sum_100 = gr.Textbox(label="100字要約")
|
| 136 |
out_sum_1 = gr.Textbox(label="1文要約")
|
| 137 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 138 |
run_btn.click(
|
| 139 |
process_resumes,
|
| 140 |
inputs=[in_files, candidate_id, notes],
|
| 141 |
+
outputs=[out_json, out_skills, out_score, out_sum_300, out_sum_100, out_sum_1],
|
| 142 |
)
|
| 143 |
|
| 144 |
|
| 145 |
if __name__ == "__main__":
|
| 146 |
+
# localhost 到達不可環境でも動くように share=True を明示
|
| 147 |
demo.launch(share=True, server_name="0.0.0.0", server_port=7860)
|