app.py

import streamlit as st
import fitz  # PyMuPDF
import tempfile
import os
import re
import json
import requests
from collections import Counter, defaultdict
from dotenv import load_dotenv
from paddleocr import PaddleOCR
from datetime import datetime

load_dotenv()
GROQ_API_KEY = os.getenv("GROQ_API_KEY")

# ------------------------------
# 設定: モデルごとのパラメータ
# ------------------------------
MODEL_CONFIG = {
    "qwen/qwen3-32b": {"temperature": 0.2, "max_tokens": 4000},
    "groq/compound": {"temperature": 0.3, "max_tokens": 4000},
    "meta-llama/llama-4-maverick-17b-128e-instruct": {"temperature": 0.2, "max_tokens": 5000},
    "moonshotai/kimi-k2-instruct-0905": {"temperature": 0.4, "max_tokens": 6000},
    "openai/gpt-oss-120b": {"temperature": 0.1, "max_tokens": 6000}
}

# ------------------------------
# 固定分野辞書（数学例。必要に応じて拡張）
# ------------------------------
FIELD_CATEGORIES = [
    "微分積分",
    "数列",
    "確率",
    "ベクトル",
    "整数",
    "図形",
    "複素数",
    "場合の数",
    "極限",
    "行列",
    "その他"
]

# ------------------------------
# 出題スコアの重み（デフォルト）
# 頻度を最重視（0.5）、未出年数0.3、LLM予測0.2
# ------------------------------
WEIGHTS = {
    "freq": 0.5,
    "years": 0.3,
    "llm": 0.2
}

# ------------------------------
# 難易度算出重み（典型性/誘導/発想）
# previous design: typicality 0.35, guidance 0.25, creativity 0.40
# ------------------------------
DIFFICULTY_WEIGHTS = {"typicality": 0.35, "guidance": 0.25, "creativity": 0.40}

# ------------------------------
# LLM呼び出し関数（Groq OpenAI互換エンドポイント想定）
# ------------------------------
def call_llm(prompt: str, model: str):
    """
    Groq chat completions (OpenAI-like) を呼ぶ簡易 wrapper。
    返り値は raw text を返す（JSONを返すようにプロンプトで依頼すること）。
    """
    if GROQ_API_KEY is None:
        raise RuntimeError("GROQ_API_KEY is not set in environment variables.")

    cfg = MODEL_CONFIG.get(model, {})
    payload = {
        "model": model,
        "messages": [{"role": "user", "content": prompt}],
        "temperature": cfg.get("temperature", 0.2),
        "max_tokens": cfg.get("max_tokens", 2000)
    }

    resp = requests.post(
        "https://api.groq.com/openai/v1/chat/completions",
        headers={
            "Authorization": f"Bearer {GROQ_API_KEY}",
            "Content-Type": "application/json"
        },
        json=payload,
        timeout=120
    )
    resp.raise_for_status()
    j = resp.json()
    # typical OpenAI-like response shape
    try:
        return j["choices"][0]["message"]["content"]
    except Exception:
        # フォールバックでraw json
        return json.dumps(j, ensure_ascii=False)

# ------------------------------
# OCR（PaddleOCR）初期化
# ------------------------------
# note: 実行環境によっては model_dir 指定や use_gpu などの調整が必要
ocr = PaddleOCR(use_angle_cls=True, lang="japan")  # 日本語OCR

def ocr_image_file(img_path: str) -> str:
    """PaddleOCRで画像ファイルを読み取り、テキストを返す（改行つき）。"""
    try:
        result = ocr.ocr(img_path, cls=True)
    except Exception as e:
        st.warning(f"OCRエラー: {e}")
        return ""
    text_lines = []
    # result の形式はページごとのリスト等環境によって差があるが通常は result[0]が行リスト
    if isinstance(result, list) and len(result) > 0 and isinstance(result[0], list):
        for line in result[0]:
            # line: [box, (text, confidence)]
            text_lines.append(line[1][0])
    else:
        # 互換形
        for page in result:
            for line in page:
                text_lines.append(line[1][0])
    return "\n".join(text_lines)

def extract_text_from_pdf_bytes(file_bytes: bytes, max_pages: int = 20) -> str:
    """
    PDFバイト列を受け取り、各ページを画像化してOCRでテキストを抽出する（最大 max_pages）。
    """
    text_accum = []
    with fitz.open(stream=file_bytes, filetype="pdf") as doc:
        pages = min(len(doc), max_pages)
        for i in range(pages):
            page = doc[i]
            # 画像解像度を上げる - scale = 2.0 等で精度向上
            mat = fitz.Matrix(2.0, 2.0)
            pix = page.get_pixmap(matrix=mat, alpha=False)
            tmp = tempfile.NamedTemporaryFile(suffix=".png", delete=False)
            try:
                pix.save(tmp.name)
                txt = ocr_image_file(tmp.name)
                text_accum.append(txt)
            finally:
                try:
                    tmp.close()
                    os.unlink(tmp.name)
                except Exception:
                    pass
    return "\n\n".join(text_accum)

# ------------------------------
# 年度抽出（ファイル名から4桁の年を探す）
# ------------------------------
def extract_year_from_filename(filename: str):
    m = re.search(r"(20\d{2}|19\d{2})", filename)
    if m:
        try:
            return int(m.group(0))
        except:
            return None
    return None

# ------------------------------
# 問題分割: 正規表現で一次分割 + LLMに整形させる
# ------------------------------
def initial_split_by_regex(text: str):
    # よくある区切りにより一次分割
    # (第\d+問) や "問\d"、"(1)" などで分ける
    parts = re.split(r"\n{2,}", text)  # 空行で分割の一次案
    # さらに "第\d+問" を前処理で残す形にして細分化
    out = []
    for p in parts:
        sub = re.split(r"(第\s*\d+\s*問|問\s*\d+|^【問題】)", p)
        if len(sub) > 1:
            # join fragments that start with marker
            buf = ""
            for s in sub:
                if re.match(r"(第\s*\d+\s*問|問\s*\d+|^【問題】)", s):
                    if buf:
                        out.append(buf.strip())
                    buf = s
                else:
                    buf += s
            if buf:
                out.append(buf.strip())
        else:
            if p.strip():
                out.append(p.strip())
    # filter empty
    out = [o for o in out if o and len(o) > 10]
    return out

def split_questions_via_llm(raw_text: str, model_choice: str):
    """
    raw_text を LLM に投げ、JSONで問題リストを返すように促す。
    戻り: list of { "id": str, "text": str }
    """
    # まず一次分割（軽量）
    candidates = initial_split_by_regex(raw_text)
    # Join for context but limit length
    sample_for_prompt = "\n\n---\n\n".join(candidates[:80])
    prompt = f"""
あなたは日本語の大学入試問題を「独立した小問」ごとに分割する専門家です。
与えられたテキストから、独立した問題ごとに分割し、JSONで次の形式を返してください。

{{"questions": [{{"id": "Q1", "text": "..."} , ...]}}

注意:
- id は Q1, Q2 ... の形式で付与してください。
- 各 text は問題本文のみにしてください（選択肢や解答欄がある場合はそれらも含めてください）。
- 出力は **純粋なJSON**だけにしてください（説明文は入れないでください）。

テキスト:
\"\"\"{sample_for_prompt}\"\"\"
"""
    raw = call_llm(prompt, model_choice)
    # try parse JSON from response
    try:
        j = json.loads(raw)
        if isinstance(j, dict) and "questions" in j:
            return j["questions"]
    except Exception:
        # try to extract JSON substring
        m = re.search(r"(\{.*\}|\[.*\])", raw, flags=re.S)
        if m:
            try:
                j = json.loads(m.group(0))
                if isinstance(j, dict) and "questions" in j:
                    return j["questions"]
            except Exception:
                pass
    # fallback: convert candidates to simple dict list
    out = []
    for i, c in enumerate(candidates, start=1):
        out.append({"id": f"Q{i}", "text": c})
    return out

# ------------------------------
# 分野分類（複数可） — LLMにJSONで返させる
# ------------------------------
def classify_field_llm(problem_text: str, model_choice: str):
    prompt = f"""
次の問題文を、以下の選択肢から該当する分野を**複数選択可**で選んでください。必ずJSONで返してください。
選択肢:
{FIELD_CATEGORIES}

問題:
\"\"\"{problem_text}\"\"\"

出力形式（例）:
{{"fields": ["微分積分", "ベクトル"]}}
"""
    raw = call_llm(prompt, model_choice)
    # parse JSON
    try:
        j = json.loads(raw)
        return j.get("fields", [])
    except Exception:
        m = re.search(r"\{.*\}", raw, flags=re.S)
        if m:
            try:
                j = json.loads(m.group(0))
                return j.get("fields", [])
            except:
                pass
    # fallback: simple keyword matching
    return heuristic_field_match(problem_text)

def heuristic_field_match(text: str):
    text_low = text
    matched = set()
    # simple keyword heuristics
    KEYWORDS = {
        "微分": "微分積分", "積分": "微分積分", "極限": "極限",
        "数列": "数列", "ベクトル": "ベクトル", "確率": "確率", "整数": "整数",
        "行列": "行列", "複素": "複素数", "図形": "図形", "場合の数": "場合の数"
    }
    for k, v in KEYWORDS.items():
        if k in text_low:
            matched.add(v)
    if not matched:
        return ["その他"]
    return list(matched)

# ------------------------------
# 指標評価（典型性・誘導度・発想）
# LLMに0-5で評価させJSONで受け取る
# ------------------------------
def evaluate_three_metrics(problem_text: str, model_choice: str):
    prompt = f"""
以下の問題を、次の3項目で0から5の整数（0が最も低く、5が最も高い）で評価し、
JSON形式で返してください。

1) 典型度 (0=非常に典型的, 5=強い応用)
2) 誘導度 (0=明確に誘導あり, 5=誘導なし)
3) 発想必要度 (0=発想不要, 5=高度な発想が必要)

出力例:
{{"typicality": 2, "guidance": 4, "creativity": 3}}

問題:
\"\"\"{problem_text}\"\"\"
"""
    raw = call_llm(prompt, model_choice)
    try:
        j = json.loads(raw)
        # Ensure integers and clamp 0-5
        for k in ["typicality", "guidance", "creativity"]:
            if k in j:
                try:
                    j[k] = int(round(float(j[k])))
                except:
                    j[k] = 0
                j[k] = max(0, min(5, j[k]))
            else:
                j[k] = 0
        return j
    except Exception:
        # extract numbers heuristically
        m = re.search(r"\{.*\}", raw, flags=re.S)
        if m:
            try:
                j = json.loads(m.group(0))
                for k in ["typicality", "guidance", "creativity"]:
                    if k in j:
                        try:
                            j[k] = int(round(float(j[k])))
                        except:
                            j[k] = 0
                        j[k] = max(0, min(5, j[k]))
                    else:
                        j[k] = 0
                return j
            except:
                pass
    # fallback defaults
    return {"typicality": 2, "guidance": 2, "creativity": 2}

# ------------------------------
# 難易度スコア計算 → 易/標準/難
# ------------------------------
def compute_difficulty_level(metrics: dict):
    typical = metrics.get("typicality", 2)
    guidance = metrics.get("guidance", 2)
    creativity = metrics.get("creativity", 2)
    score = (
        typical * DIFFICULTY_WEIGHTS["typicality"]
        + guidance * DIFFICULTY_WEIGHTS["guidance"]
        + creativity * DIFFICULTY_WEIGHTS["creativity"]
    )
    # 0..5 scale -> thresholds:
    if score < 2.0:
        level = "易"
    elif score < 3.5:
        level = "標準"
    else:
        level = "難"
    return {"score": round(score, 3), "level": level}

# ------------------------------
# LLMによる分野別「今後の出題可能性」スコア（0-10）
# 1回のファイル（ある大学の過去問集合）ごとに一括で聞く
# ------------------------------
def ask_llm_field_likelihood(problems_text: str, fields: list, model_choice: str):
    """
    fields: list of field strings to query (unique set)
    returns dict field->likelihood (0-10)
    """
    field_list_str = ", ".join(fields)
    prompt = f"""
あなたは大学入試の出題予測の専門家です。
以下は過去問（複数問題を連結したテキスト）です。与えられた分野ごとに、次年度に「その大学がその分野を出しそうな可能性」を 0 ～ 10 の整数で評価し、JSONで返してください。
0は「まず出ない」、10は「非常に出やすい」を表します。

対象分野: [{field_list_str}]

過去問全文:
\"\"\"{problems_text}\"\"\"

出力例:
{{"微分積分": 6, "確率": 3}}
"""
    raw = call_llm(prompt, model_choice)
    try:
        j = json.loads(raw)
        # ensure integer 0-10
        for k in list(j.keys()):
            try:
                j[k] = int(round(float(j[k])))
            except:
                j[k] = 0
            j[k] = max(0, min(10, j[k]))
        return j
    except Exception:
        # fallback: empty scores
        fallback = {f: 5 for f in fields}
        return fallback

# ------------------------------
# 類題 + 解答 生成（分野 × 難易度を与えて1問ずつ生成）
# ------------------------------
def generate_similar_problem_and_solution(field: str, difficulty_level: str, context_examples: str, model_choice: str):
    """
    difficulty_level in ("易","標準","難")
    returns dict with "problem", "solution", "explanation"
    """
    prompt = f"""
あなたは大学入試の問題を作る教師です。
分野: {field}
難易度: {difficulty_level}
文脈（過去問の例／作風）:
\"\"\"{context_examples}\"\"\"

この条件で**大学入試らしい類題を1問**作成し、続けて**模範解答**と**丁寧な解説**を作ってください。
出力は**JSONのみ**で、キーは "problem", "solution", "explanation" としてください。

例:
{{"problem": "...", "solution": "...", "explanation": "..."}}
"""
    raw = call_llm(prompt, model_choice)
    try:
        j = json.loads(raw)
        return {
            "problem": j.get("problem", "").strip(),
            "solution": j.get("solution", "").strip(),
            "explanation": j.get("explanation", "").strip()
        }
    except Exception:
        # try to extract parts heuristically
        # fallback: put entire raw into problem and blank solution (not ideal)
        return {"problem": raw.strip(), "solution": "", "explanation": ""}

# ------------------------------
# 出題スコア算出
# freq_norm: normalized frequency 0..1
# years_since_last: integer (0 if unknown)
# llm_likelihood: 0..10 -> normalize to 0..1
# ------------------------------
def compute_field_score(freq_norm: float, years_since_last: int, llm_likelihood: int):
    # normalize years: assume recent = 0 years -> transform to a 0..1 where larger is increases score
    # We'll cap years_since_last to 5 for scaling
    years_factor = min(max(years_since_last, 0), 5) / 5.0  # 0..1
    llm_norm = min(max(llm_likelihood, 0), 10) / 10.0  # 0..1
    score = (
        WEIGHTS["freq"] * freq_norm +
        WEIGHTS["years"] * years_factor +
        WEIGHTS["llm"] * llm_norm
    )
    return round(score, 4)

# ------------------------------
# ユーティリティ: 年度ごとに各分野最終出現年を推定（ファイル名に年があれば）
# uploaded_files_meta: list of dict {"name":..., "year": int or None, "problems": [...]}
# problems have 'fields' and maybe 'year' info
# ------------------------------
def estimate_last_appearance_years(uploaded_files_meta):
    """
    Returns dict field -> latest_year it appears in uploaded_files_meta; None if unknown.
    """
    last_year = {}
    for f in uploaded_files_meta:
        year = f.get("year")
        problems = f.get("problems", [])
        for p in problems:
            for fld in p.get("fields", []):
                if fld not in last_year:
                    last_year[fld] = None
                if year:
                    if last_year[fld] is None or year > last_year[fld]:
                        last_year[fld] = year
    return last_year

# ------------------------------
# Streamlit UI & main workflow
# ------------------------------
st.set_page_config(page_title="過去問解析・出題予測アプリ", layout="wide")
st.title("過去問解析・出題予測（PaddleOCR + LLM）")

with st.sidebar:
    st.header("設定")
    model_choice = st.selectbox("モデル選択", list(MODEL_CONFIG.keys()))
    max_pages = st.number_input("PDF 最大ページ数（1ファイルあたり）", min_value=1, max_value=20, value=20)
    gen_count_per_field = st.number_input("生成する予想問/分野あたり（件）", min_value=1, max_value=5, value=2)
    show_raw_llm = st.checkbox("LLMの生出力を表示（デバッグ）", value=False)

uploaded_files = st.file_uploader("過去問PDF（スキャン）をアップロード（最大20ファイル）", type=["pdf"], accept_multiple_files=True)

if uploaded_files:
    if len(uploaded_files) > 20:
        st.error("最大20ファイルまでです。")
    else:
        if st.button("解析開始"):
            overall_field_counter = Counter()
            uploaded_meta = []  # list of {name, year, problems: [{id,text,fields,metrics,difficulty_score,level}]}
            st.info("OCRと解析を開始します。ファイル数とページ数によって時間がかかります。")

            # iterate files
            for file in uploaded_files:
                st.write(f"処理中: {file.name}")
                raw_bytes = file.read()
                year = extract_year_from_filename(file.name)
                # OCR
                try:
                    raw_text = extract_text_from_pdf_bytes(raw_bytes, max_pages=max_pages)
                except Exception as e:
                    st.error(f"OCR失敗: {e}")
                    raw_text = ""
                if not raw_text.strip():
                    st.warning(f"{file.name} からテキストが抽出できませんでした。")
                    continue

                # split into questions
                try:
                    questions = split_questions_via_llm(raw_text, model_choice)
                except Exception as e:
                    st.warning(f"問題分割でエラー: {e}")
                    # fallback to simple split
                    parts = initial_split_by_regex(raw_text)
                    questions = [{"id": f"Q{i+1}", "text": parts[i]} for i in range(len(parts))]

                st.write(f"{file.name} -> 検出問題数: {len(questions)}")

                # For per-file LLM field-likelihood, prepare sample context
                sample_context_for_llm = "\n\n".join([q["text"] for q in questions[:60]])

                # classify + metrics for each question
                problems_meta = []
                for q in questions:
                    qtext = q.get("text", "")
                    qid = q.get("id", f"Q??")
                    # classify fields (multiple)
                    try:
                        fields = classify_field_llm(qtext, model_choice)
                        if not fields:
                            fields = heuristic_field_match(qtext)
                    except Exception as e:
                        fields = heuristic_field_match(qtext)

                    # evaluate metrics
                    try:
                        metrics = evaluate_three_metrics(qtext, model_choice)
                    except Exception as e:
                        metrics = {"typicality": 2, "guidance": 2, "creativity": 2}

                    # compute difficulty
                    diff = compute_difficulty_level(metrics)

                    problems_meta.append({
                        "id": qid,
                        "text": qtext,
                        "fields": fields,
                        "metrics": metrics,
                        "difficulty_score": diff["score"],
                        "difficulty_level": diff["level"]
                    })

                    # update overall field counter (複合は両方1カウント)
                    for f in fields:
                        overall_field_counter[f] += 1

                # ask LLM for per-field likelihoods (for this file)
                unique_fields = sorted(list({f for p in problems_meta for f in p["fields"]}))
                try:
                    llm_field_likelihoods = ask_llm_field_likelihood(sample_context_for_llm, unique_fields, model_choice)
                except Exception:
                    llm_field_likelihoods = {f: 5 for f in unique_fields}

                uploaded_meta.append({
                    "name": file.name,
                    "year": year,
                    "problems": problems_meta,
                    "llm_field_likelihoods": llm_field_likelihoods
                })

            # ===== 集計フェーズ =====
            st.success("個別ファイル解析完了。集計を行います。")
            total_problems = sum(len(m["problems"]) for m in uploaded_meta)
            st.write(f"合計検出問題数: {total_problems}")

            # field -> total count across all files
            field_counts = dict(overall_field_counter)
            # compute normalized frequency
            field_freq_norm = {}
            for fld, cnt in field_counts.items():
                field_freq_norm[fld] = cnt / total_problems if total_problems > 0 else 0.0

            # estimate last appearance year per field
            last_years = estimate_last_appearance_years(uploaded_meta)
            # compute years_since_last relative to max year seen (if any)
            years_seen = [m["year"] for m in uploaded_meta if m["year"]]
            current_year_ref = max(years_seen) if years_seen else datetime.now().year
            years_since_last = {}
            for fld in field_freq_norm.keys():
                last = last_years.get(fld)
                if last:
                    diff = max(0, current_year_ref - last)
                    years_since_last[fld] = diff
                else:
                    years_since_last[fld] = 0  # unknown -> 0

            # combine llm field likelihoods across files by averaging (normalize 0..10)
            llm_scores_accum = defaultdict(list)
            for m in uploaded_meta:
                for f, v in m.get("llm_field_likelihoods", {}).items():
                    llm_scores_accum[f].append(v)
            llm_avg = {}
            for f, arr in llm_scores_accum.items():
                llm_avg[f] = int(round(sum(arr) / max(1, len(arr))))

            # compute final field score
            final_field_scores = {}
            for fld in field_freq_norm.keys():
                freq_norm = field_freq_norm.get(fld, 0.0)
                yrs = years_since_last.get(fld, 0)
                llm_v = llm_avg.get(fld, 5)
                final_score = compute_field_score(freq_norm, yrs, llm_v)
                final_field_scores[fld] = {
                    "freq_count": field_counts.get(fld, 0),
                    "freq_norm": round(freq_norm, 4),
                    "years_since_last": yrs,
                    "llm_avg": llm_v,
                    "final_score": final_score
                }

            # sort fields by final_score desc
            sorted_fields = sorted(final_field_scores.items(), key=lambda x: x[1]["final_score"], reverse=True)

            # Display summary table
            st.subheader("分野ランキング（出題スコア順）")
            st.table([{ "分野": k,
                        "出題回数": v["freq_count"],
                        "頻度(norm)": v["freq_norm"],
                        "直近未出年数": v["years_since_last"],
                        "LLM予測(0-10)": v["llm_avg"],
                        "出題スコア": v["final_score"] } for k, v in sorted_fields])

            # For each top field, generate similar problems + solutions
            st.subheader("予想問題（分野ごと）")
            downloadable_data = {}  # field -> text content
            # Prepare context examples (concatenate some examples per field)
            field_examples = defaultdict(list)
            for meta in uploaded_meta:
                for p in meta["problems"]:
                    for f in p["fields"]:
                        if len(field_examples[f]) < 40:
                            # add up to limited examples
                            field_examples[f].append(p["text"])

            # For each field selected (sorted by final score), generate requested count
            for fld, info in sorted_fields:
                if info["freq_count"] == 0:
                    continue
                st.markdown(f"### 分野: {fld} （スコア {info['final_score']}）")
                context_examples_text = "\n\n---\n\n".join(field_examples.get(fld, [])[:30])
                gen_list = []
                for i in range(int(gen_count_per_field)):
                    # choose difficulty attempt: pick distribution from existing problems: we will pick a difficulty target based on freq of difficulties in that field
                    # compute distribution of difficulties for this field
                    diffs = []
                    for meta in uploaded_meta:
                        for p in meta["problems"]:
                            if fld in p["fields"]:
                                diffs.append(p["difficulty_level"])
                    # derive a likely difficulty: if no data, default to 標準
                    target_difficulty = "標準"
                    if diffs:
                        # simple heuristic: sample most common
                        c = Counter(diffs)
                        target_difficulty = c.most_common(1)[0][0]
                    # generate
                    try:
                        gen = generate_similar_problem_and_solution(fld, target_difficulty, context_examples_text, model_choice)
                    except Exception as e:
                        gen = {"problem": f"生成に失敗しました: {e}", "solution": "", "explanation": ""}
                    gen_list.append(gen)
                    # show one by one
                    st.write(f"**予想問 {i+1}（難易度推定: {target_difficulty}）**")
                    st.write(gen.get("problem", ""))
                    st.write("**模範解答**")
                    st.write(gen.get("solution", ""))
                    st.write("**解説**")
                    st.write(gen.get("explanation", ""))
                    if show_raw_llm:
                        st.text_area(f"LLM raw for {fld} #{i+1}", value=json.dumps(gen, ensure_ascii=False, indent=2), height=150)

                # prepare downloadable text
                out_lines = []
                out_lines.append(f"=== 分野: {fld} ===\n")
                for idx, g in enumerate(gen_list, start=1):
                    out_lines.append(f"【予想問 {idx}】\n{g.get('problem','')}\n\n【模範解答】\n{g.get('solution','')}\n\n【解説】\n{g.get('explanation','')}\n\n\n")
                downloadable_data[fld] = "\n".join(out_lines)

            # Offer download for each field
            st.subheader("ダウンロード")
            for fld, content in downloadable_data.items():
                filename = f"predicted_{fld}.txt"
                st.download_button(label=f"{fld} の予想問題と解答をダウンロード", data=content, file_name=filename, mime="text/plain")

            # Option: download a single aggregate file
            if downloadable_data:
                aggregate = []
                for fld, cont in downloadable_data.items():
                    aggregate.append(cont)
                aggregate_text = "\n\n====\n\n".join(aggregate)
                st.download_button(label="全分野まとめてダウンロード", data=aggregate_text, file_name="predicted_all_fields.txt", mime="text/plain")

            st.success("完了しました。")