Update ui/ui_app.py

ui/ui_app.py

@@ -1,5 +1,6 @@
+# ui/ui_app.py
 from __future__ import annotations
-import json, traceback, base64
+import json, traceback, base64, math
 from typing import Any, Dict, List
 import gradio as gr
 import pandas as pd
@@ -8,9 +9,12 @@ import plotly.graph_objects as go
 from core.pdf_io import pdf_to_images, pdf_to_text
 from core.extract import extract_financials
 from core.scoring import score_company
-from core.external_score import get_external_template_df, fill_missing_with_external, score_external_from_df
+from core.external_score import (
+    get_external_template_df, fill_missing_with_external, score_external_from_df, apply_llm_signals_to_df
+)
 from core.units import detect_unit, unit_factor, scale_financials_yen
 from core.openai_client import VISION_MODEL, TEXT_MODEL, get_client
+from core.llm_quant import extract_market_product_signals
 
 # ===== helpers =====
 def fin_to_df(fin: Dict[str, Any]) -> pd.DataFrame:
@@ -35,6 +39,29 @@ def df_to_fin(df: pd.DataFrame) -> Dict[str, Any]:
             out[cat][item] = parsed
     return out
 
+def _fmt_yen(n: float) -> str:
+    if n is None: return "—"
+    try:
+        n = float(n)
+    except Exception:
+        return "—"
+    # 視認性のため 兆/億/万円 に自動スケール
+    absn = abs(n)
+    if absn >= 1e12:
+        return f"{n/1e12:.2f} 兆円"
+    if absn >= 1e8:
+        return f"{n/1e8:.2f} 億円"
+    if absn >= 1e4:
+        return f"{n/1e4:.1f} 万円"
+    return f"{int(n):,} 円"
+
+def _fmt_pct(r: float) -> str:
+    if r is None: return "—"
+    try:
+        return f"{r*100:.2f}%"
+    except Exception:
+        return "—"
+
 def radar(score: Dict[str, Any]) -> go.Figure:
     labels = [d["metric"] for d in score["details"]]
     values = [d["score"] for d in score["details"]]
@@ -59,18 +86,63 @@ def health() -> str:
     msgs.append(f"ℹ️ Vision={VISION_MODEL} / Text={TEXT_MODEL}")
     return "<br>".join(msgs)
 
+def kpi_cards_html(fin: Dict[str, Any], score: Dict[str, Any]) -> str:
+    bs = fin.get("balance_sheet", {}) or {}
+    pl = fin.get("income_statement", {}) or {}
+    assets = bs.get("total_assets")
+    equity = bs.get("total_equity")
+    curA = bs.get("current_assets"); curL = bs.get("current_liabilities")
+    sales = pl.get("sales"); op = pl.get("operating_income"); net = pl.get("net_income")
+
+    equity_ratio = (equity or 0) / (assets or 1) if assets else None
+    current_ratio = (curA or 0) / (curL or 1) if curL else None
+    opm = (op or 0) / (sales or 1) if sales else None
+    npm = (net or 0) / (sales or 1) if sales else None
+    roa = (net or 0) / (assets or 1) if assets else None
+
+    def badge(val, kind="pct"):
+        if val is None: return '<span class="px-2 py-1 rounded bg-gray-200 text-gray-700">—</span>'
+        v = val*100 if kind=="pct" else val
+        col = "#22c55e" if v >= (20 if kind=="pct" else 0) else "#f59e0b" if v >= (10 if kind=="pct" else 0) else "#ef4444"
+        return f'<span class="px-2 py-1 rounded" style="background:{col}20;color:{col}">{v:.2f}{"%" if kind=="pct" else ""}</span>'
+
+    html = f"""
+<div style="display:grid;grid-template-columns:repeat(3,minmax(0,1fr));gap:12px">
+  <div class="card" style="padding:12px;border:1px solid #eee;border-radius:12px">
+    <div>総資産</div><div style="font-size:1.2rem;font-weight:700">{_fmt_yen(assets)}</div>
+  </div>
+  <div class="card" style="padding:12px;border:1px solid #eee;border-radius:12px">
+    <div>売上高</div><div style="font-size:1.2rem;font-weight:700">{_fmt_yen(sales)}</div>
+  </div>
+  <div class="card" style="padding:12px;border:1px solid #eee;border-radius:12px">
+    <div>営業利益</div><div style="font-size:1.2rem;font-weight:700">{_fmt_yen(op)}</div>
+  </div>
+  <div class="card" style="padding:12px;border:1px solid #eee;border-radius:12px">
+    <div>自己資本比率</div><div style="font-size:1.1rem;font-weight:700">{_fmt_pct(equity_ratio)} {badge(equity_ratio)}</div>
+  </div>
+  <div class="card" style="padding:12px;border:1px solid #eee;border-radius:12px">
+    <div>流動比率</div><div style="font-size:1.1rem;font-weight:700">{_fmt_pct(current_ratio)} {badge(current_ratio)}</div>
+  </div>
+  <div class="card" style="padding:12px;border:1px solid #eee;border-radius:12px">
+    <div>ROA</div><div style="font-size:1.1rem;font-weight:700">{_fmt_pct(roa)} {badge(roa)}</div>
+  </div>
+</div>
+<div style="margin-top:8px;color:#64748b">社内スコア: <b>{score['total_score']}</b> （グレード {score['grade']}）</div>
+"""
+    return html
+
 # ===== main actions =====
-def run_analyze(company: str, use_vision: bool, unit_sel: str, files: list[str]):
+def run_analyze(company: str, use_vision: bool, unit_sel: str, use_pdf_for_ext: bool, aux_text: str, files: list[str]):
     if not files:
         raise gr.Error("PDF をアップロードしてください。")
 
-    # 1) 単位の自動推定（先頭数ページのテキスト）
+    # 1) 単位の自動推定
     first_text = pdf_to_text(files[0], pages=2)
     detected = detect_unit(first_text) or "円"
     unit_label = unit_sel if unit_sel != "自動推定" else detected
     factor = unit_factor(unit_label)
 
-    # 2) Vision（失敗時はテキストモデル）で抽出
+    # 2) 抽出（Vision→失敗時Text）
     try:
         imgs: List[bytes] = []
         for p in files:
@@ -85,23 +157,48 @@ def run_analyze(company: str, use_vision: bool, unit_sel: str, files: list[str])
     # 3) 円換算
     fin_yen = scale_financials_yen(fin_raw, factor=factor)
 
-    # 4) 内部スコア
+    # 4) 社内スコア
     df = fin_to_df(fin_yen)
     score = score_company(fin_yen)
     fig = radar(score)
+    kpi_html = kpi_cards_html(fin_yen, score)
 
-    # 5) AI所見（中立性を強化）
+    # 5) 外部評価に使う LLM 定量シグナル（市場/製品）
+    signals = None
+    try:
+        ext_text = ""
+        if use_pdf_for_ext:
+            # 追加でテキストを多めに取得（セクション説明文など）
+            for p in files:
+                ext_text += pdf_to_text(p, pages=8) + "\n\n"
+        ext_text = (aux_text or "") + "\n\n" + ext_text
+        if len(ext_text.strip()) > 0:
+            signals = extract_market_product_signals(ext_text[:18000], company_hint=company or "")
+    except Exception:
+        signals = None
+
+    # 6) 外部評価テンプレを生成 → LLMで可能な項目を自動補完
+    df_ext = get_external_template_df()
+    df_ext = fill_missing_with_external(df_ext)
+    if signals:
+        df_ext = apply_llm_signals_to_df(df_ext, signals)
+
+    # 7) AI所見（中立・根拠明示）
     try:
         client = get_client()
-        prompt = f"""あなたは独立の財務アナリストです。主観や推測を避け、事実と比率を根拠に短くコメントしてください。
-- 書き方: 箇条書き、断定的表現・煽り表現は禁止。将来予測はしない。
-- 必ず根拠（指標名と値）を各行に併記する。
+        prompt = f"""あなたは独立の財務アナリストです。主観や推測を避け、事実と“比率”を根拠に簡潔な箇条書きを作成。
+- 断定・煽り禁止、将来予測禁止
+- 各行の末尾に根拠(指標名=値)を括弧で併記
+- 外部評価(市場/製品)は参考情報として最後に別見出しで一言添える
 
 [財務データ（円換算）]
 {json.dumps(fin_yen, ensure_ascii=False)}
 
-[社内スコア]
+[社内スコア（財務比率基準）]
 {json.dumps(score, ensure_ascii=False)}
+
+[外部シグナル（市場/製品；参考）]
+{json.dumps(signals or {}, ensure_ascii=False)}
 """
         resp = client.chat.completions.create(
             model=TEXT_MODEL,
@@ -114,13 +211,26 @@ def run_analyze(company: str, use_vision: bool, unit_sel: str, files: list[str])
         insight = f"AI所見の生成に失敗: {e}"
 
     unit_info = f"PDF表記の単位: <b>{detected}</b> / 適用単位: <b>{unit_label}</b>（円換算係数={factor:g}）"
-    return (unit_info,
-            json.dumps(fin_raw, ensure_ascii=False, indent=2),   # 参照用（換算前）
-            json.dumps(fin_yen, ensure_ascii=False, indent=2),   # 実計算に使用
-            df,
-            json.dumps(score, ensure_ascii=False, indent=2),
-            fig,
-            insight)
+    signals_md = "—"
+    if signals:
+        ev = signals.get("market", {}).get("evidence") or []
+        cagr = signals.get("market", {}).get("cagr_pct")
+        pc = signals.get("products", {}).get("count")
+        pg = signals.get("products", {}).get("growing_count")
+        signals_md = f"- 市場CAGR: {cagr if cagr is not None else '—'}%\n- 主力商品数: {pc or '—'} / 成長中: {pg or '—'}\n- エビデンス: {(' / '.join(ev[:2])) if ev else '—'}"
+
+    return (
+        unit_info,                 # HTML
+        kpi_html,                  # HTML
+        json.dumps(fin_raw, ensure_ascii=False, indent=2),   # 折りたたみ表示用
+        json.dumps(fin_yen, ensure_ascii=False, indent=2),   # 折りたたみ表示用
+        df,                        # 編集表（円換算）
+        json.dumps(score, ensure_ascii=False, indent=2),
+        fig,
+        insight,
+        df_ext,                    # 外部テンプレ（LLM補完済み）
+        signals_md                 # LLM抽出の要点
+    )
 
 def run_recalc(df: pd.DataFrame):
     try:
@@ -134,12 +244,6 @@ def run_recalc(df: pd.DataFrame):
         tb = traceback.format_exc(limit=6)
         raise gr.Error(f"再計算に失敗しました: {e}\n\n<pre style='white-space:pre-wrap'>{tb}</pre>")
 
-# ===== 外部評価 =====
-def open_external_template():
-    df = get_external_template_df()
-    df = fill_missing_with_external(df)
-    return df
-
 def calc_external(df_ext: pd.DataFrame):
     res = score_external_from_df(df_ext)
     return json.dumps(res, ensure_ascii=False, indent=2)
@@ -155,6 +259,9 @@ def build_ui() -> gr.Blocks:
                 use_vision = gr.Checkbox(value=True, label="OpenAIでPDFをAI解析（Vision）")
                 unit_sel = gr.Dropdown(choices=["自動推定","円","千円","百万円","千万円","億円"],
                                        value="自動推定", label="金額単位（PDF記載）")
+                use_pdf_for_ext = gr.Checkbox(value=True, label="PDFから市場/製品情報も抽出（外部評価用）")
+                aux_text = gr.Textbox(label="補助テキスト・Web抜粋（任意）", lines=4,
+                                      placeholder="業界説明/IRの抜粋などを貼ると、市場CAGRや主力商品数を自動推定します。")
                 files = gr.File(label="決算書PDF（複数可）", file_count="multiple", type="filepath")
                 run_btn = gr.Button("📄 解析して反映", variant="primary")
                 recalc_btn = gr.Button("🔁 表の値で再計算")
@@ -163,33 +270,38 @@ def build_ui() -> gr.Blocks:
 
             with gr.Column(scale=1):
                 unit_html = gr.HTML(label="単位情報")
-                fin_json_raw = gr.Code(label="抽出JSON（換算前）", language="json", interactive=False)
+                kpi_html = gr.HTML(label="主要KPI")
 
         with gr.Tabs():
-            with gr.Tab("抽出結果（円換算・表で編集可）"):
-                fin_json_yen = gr.Code(label="抽出JSON（円換算）", language="json")
+            with gr.Tab("概要"):
+                chart = gr.Plot(label="スコアレーダー（社内ルール）")
+            with gr.Tab("抽出結果（編集可）"):
+                with gr.Accordion("抽出JSON（換算前/円換算）", open=False):
+                    fin_json_raw = gr.Code(label="抽出JSON（換算前）", language="json", interactive=False)
+                    fin_json_yen = gr.Code(label="抽出JSON（円換算）", language="json")
                 df_out = gr.Dataframe(headers=["category", "item", "value"], interactive=True, wrap=True)
-            with gr.Tab("スコアリング（社内ルール）"):
+            with gr.Tab("スコア（社内ルール）"):
                 score_json = gr.Code(label="スコア（JSON）", language="json")
-                chart = gr.Plot(label="スコアレーダー")
             with gr.Tab("AI所見（中立）"):
                 insight_md = gr.Markdown()
-            with gr.Tab("外部評価（テンプレ&採点）"):
-                with gr.Row():
-                    open_t = gr.Button("📋 テンプレを開く")
-                    calc_t = gr.Button("🧮 外部スコア計算")
+            with gr.Tab("外部評価（定量化）"):
+                signals_md = gr.Markdown(label="抽出シグナル（市場/製品）")
                 df_ext = gr.Dataframe(headers=["カテゴリー","入力項目","値"], interactive=True, wrap=True)
+                calc_t = gr.Button("🧮 外部スコア計算")
                 ext_json = gr.Code(label="外部評価（JSON）", language="json")
 
-        run_btn.click(run_analyze,
-                      inputs=[company, use_vision, unit_sel, files],
-                      outputs=[unit_html, fin_json_raw, fin_json_yen, df_out, score_json, chart, insight_md],
-                      concurrency_limit=1)
+        run_btn.click(
+            run_analyze,
+            inputs=[company, use_vision, unit_sel, use_pdf_for_ext, aux_text, files],
+            outputs=[unit_html, kpi_html, fin_json_raw, fin_json_yen, df_out, score_json, chart, insight_md, df_ext, signals_md],
+            concurrency_limit=1
+        )
 
-        recalc_btn.click(run_recalc, inputs=[df_out], outputs=[score_json, chart, fin_json_yen],
-                         concurrency_limit=1)
+        recalc_btn.click(
+            run_recalc, inputs=[df_out], outputs=[score_json, chart, fin_json_yen],
+            concurrency_limit=1
+        )
 
-        open_t.click(open_external_template, outputs=[df_ext], concurrency_limit=1)
         calc_t.click(calc_external, inputs=[df_ext], outputs=[ext_json], concurrency_limit=1)
 
         health_btn.click(health, outputs=health_out, concurrency_limit=1)