app.py

# -*- coding: utf-8 -*-
"""
Verbatify — Analyse sémantique NPS (Paste-only, NPS inféré)
Interface simplifiée : toutes les options sont appliquées automatiquement.
"""

import os, re, json, collections, tempfile, zipfile
from typing import List, Dict, Optional
import pandas as pd
import gradio as gr
import plotly.express as px
import plotly.graph_objects as go
import plotly.io as pio

# ====================== CSS (externe si présent, sinon fallback) ======================

BASE_DIR = os.path.dirname(os.path.abspath(__file__))
CSS_FILE = os.path.join(BASE_DIR, "verbatim.css")

VB_CSS_FALLBACK = r"""
@import url('https://fonts.googleapis.com/css2?family=Manrope:wght@400;500;700;800&display=swap');
:root{--vb-bg:#F8FAFC;--vb-text:#0F172A;--vb-primary:#7C3AED;--vb-primary-2:#06B6D4;--vb-border:#E2E8F0;}
*{color-scheme:light !important}
html,body,.gradio-container{background:var(--vb-bg)!important;color:var(--vb-text)!important;
  font-family:Manrope,system-ui,-apple-system,'Segoe UI',Roboto,Arial,sans-serif!important}
.gradio-container{max-width:1120px!important;margin:0 auto!important}
.vb-hero{display:flex;align-items:center;gap:16px;padding:20px 22px;margin:10px 0 20px;
  background:linear-gradient(90deg,rgba(124,58,237,.18),rgba(6,182,212,.18));border:1px solid var(--vb-border);
  border-radius:14px;box-shadow:0 10px 26px rgba(2,6,23,.08)}
.vb-hero .vb-title{font-size:22px;color:#0F172A;font-weight:500}
.vb-hero .vb-sub{font-size:13px;color:#0F172A}
.gradio-container .vb-cta{background:linear-gradient(90deg,var(--vb-primary),var(--vb-primary-2))!important;color:#fff!important;
  border:0!important;font-weight:700!important;font-size:16px!important;padding:14px 18px!important;border-radius:14px!important;
  box-shadow:0 10px 24px rgba(124,58,237,.28)}
.gradio-container .vb-cta:hover{transform:translateY(-2px);filter:brightness(1.05)}
/* Patch encarts vides & texte noir partout */
.gradio-container .empty,
.gradio-container [class*="unpadded_box"],
.gradio-container [class*="unpadded-box"],
.gradio-container .empty[class*="box"]{background:#FFFFFF!important;background-image:none!important;border:1px solid transparent!important;box-shadow:none!important}
.gradio-container .empty *, .gradio-container [class*="unpadded_box"] *{color:#0F172A!important;fill:#0F172A!important}
"""

VB_CSS = None
try:
    if os.path.exists(CSS_FILE):
        with open(CSS_FILE, "r", encoding="utf-8") as f:
            VB_CSS = f.read()
except Exception:
    VB_CSS = None
if not VB_CSS:
    VB_CSS = VB_CSS_FALLBACK

# ====================== Plotly theme ======================

def apply_plotly_theme():
    pio.templates["verbatify"] = go.layout.Template(
        layout=go.Layout(
            font=dict(family="Manrope, system-ui, -apple-system, Segoe UI, Roboto, Arial, sans-serif",
                      size=13, color="#0F172A"),
            paper_bgcolor="white", plot_bgcolor="white",
            colorway=["#7C3AED","#06B6D4","#2563EB","#10B981","#A855F7","#22D3EE","#1D4ED8","#0EA5E9"],
            xaxis=dict(gridcolor="#E2E8F0", zerolinecolor="#E2E8F0"),
            yaxis=dict(gridcolor="#E2E8F0", zerolinecolor="#E2E8F0"),
            legend=dict(borderwidth=0, bgcolor="rgba(255,255,255,0)")
        )
    )
    pio.templates.default = "verbatify"

LOGO_SVG = """<svg xmlns='http://www.w3.org/2000/svg' width='224' height='38' viewBox='0 0 224 38'>
  <defs><linearGradient id='g' x1='0%' y1='0%' x2='100%'><stop offset='0%' stop-color='#7C3AED'/><stop offset='100%' stop-color='#06B6D4'/></linearGradient></defs>
  <g fill='none' fill-rule='evenodd'>
    <rect x='0' y='7' width='38' height='24' rx='12' fill='url(#g)'/>
    <circle cx='13' cy='19' r='5' fill='#fff' opacity='0.95'/><circle cx='25' cy='19' r='5' fill='#fff' opacity='0.72'/>
    <text x='46' y='25' font-family='Manrope, system-ui, -apple-system, Segoe UI, Roboto, Arial, sans-serif' font-size='20' font-weight='800' fill='#0F172A' letter-spacing='0.2'>Verbatify</text>
  </g>
</svg>"""

# ====================== unidecode fallback ======================

try:
    from unidecode import unidecode
except Exception:
    import unicodedata
    def unidecode(x):
        try:
            return unicodedata.normalize('NFKD', str(x)).encode('ascii','ignore').decode('ascii')
        except Exception:
            return str(x)

# ====================== Thésaurus ASSURANCE ======================

THEMES = {
    "Remboursements santé":[r"\bremboursement[s]?\b", r"\bt[eé]l[eé]transmission\b", r"\bno[eé]mie\b",
        r"\bprise\s*en\s*charge[s]?\b", r"\btaux\s+de\s+remboursement[s]?\b", r"\b(ameli|cpam)\b",
        r"\bcompl[eé]mentaire\s+sant[eé]\b", r"\bmutuelle\b", r"\battestation[s]?\b", r"\bcarte\s+(mutuelle|tiers\s*payant)\b"],
    "Tiers payant / Réseau de soins":[r"\btiers\s*payant\b", r"\br[ée]seau[x]?\s+de\s+soins\b",
        r"\b(optique|dentaire|hospitalisation|pharmacie)\b", r"\bitelis\b", r"\bsant[eé]clair\b", r"\bkalixia\b"],
    "Sinistres / Indemnisation":[r"\bsinistre[s]?\b", r"\bindemni(sation|ser)\b", r"\bexpertis[ea]\b",
        r"\bd[eé]claration\s+de\s+sinistre\b", r"\bconstat\b", r"\bbris\s+de\s+glace\b", r"\bassistance\b", r"\bd[ée]pannage\b"],
    "Adhésion / Contrat":[r"\badh[eé]sion[s]?\b", r"\bsouscription[s]?\b", r"\baffiliation[s]?\b", r"\bcontrat[s]?\b",
        r"\bavenant[s]?\b", r"\bcarence[s]?\b", r"\brenouvellement[s]?\b", r"\br[eé]siliation[s]?\b"],
    "Garanties / Exclusions / Franchise":[r"\bgarantie[s]?\b", r"\bexclusion[s]?\b", r"\bplafond[s]?\b",
        r"\bfranchise[s]?\b", r"\bconditions\s+g[eé]n[eé]rales\b", r"\bnotice\b"],
    "Cotisations / Facturation":[r"\bcotisation[s]?\b", r"\bpr[eé]l[eè]vement[s]?\b", r"\bech[eé]ancier[s]?\b",
        r"\bfacture[s]?\b", r"\berreur[s]?\s+de\s+facturation\b", r"\bremboursement[s]?\b", r"\bRIB\b", r"\bIBAN\b"],
    "Délais & Suivi dossier":[r"\bd[eé]lai[s]?\b", r"\btraitement[s]?\b", r"\bsuivi[s]?\b", r"\brelance[s]?\b", r"\bretard[s]?\b"],
    "Espace client / App / Connexion":[r"\bespace\s+client\b", r"\bapplication\b", r"\bapp\b", r"\bsite\b",
        r"\bconnexion\b", r"\bidentifiant[s]?\b", r"\bmot\s+de\s+passe\b", r"\bpaiement\s+en\s+ligne\b",
        r"\bbogue[s]?\b", r"\bbug[s]?\b", r"\bnavigation\b", r"\binterface\b", r"\bUX\b"],
    "Support / Conseiller":[r"\bSAV\b", r"\bservice[s]?\s+client[s]?\b", r"\bconseiller[s]?\b",
        r"\b[rR][eé]ponse[s]?\b", r"\bjoignable[s]?\b", r"\brapp?el\b"],
    "Communication / Transparence":[r"\binformation[s]?\b", r"\bcommunication\b", r"\btransparence\b",
        r"\bclart[eé]\b", r"\bcourrier[s]?\b", r"\bmail[s]?\b", r"\bnotification[s]?\b"],
    "Prix":[r"\bprix\b", r"\bcher[s]?\b", r"\bco[uû]t[s]?\b", r"\btarif[s]?\b",
        r"\bcomp[eé]titif[s]?\b", r"\babusif[s]?\b", r"\bbon\s+rapport\s+qualit[eé]\s*prix\b"],
    "Offre / Gamme":[r"\boffre[s]?\b", r"\bgamme[s]?\b", r"\bdisponibilit[eé][s]?\b", r"\bdevis\b", r"\bchoix\b", r"\bcatalogue[s]?\b"],
    "Produit/Qualité":[r"\bqualit[eé]s?\b", r"\bfiable[s]?\b", r"\bconforme[s]?\b", r"\bnon\s+conforme[s]?\b",
        r"\bd[eé]fectueux?[es]?\b", r"\bperformant[e]?[s]?\b"],
    "Agence / Accueil":[r"\bagence[s]?\b", r"\bboutique[s]?\b", r"\baccueil\b", r"\bconseil[s]?\b", r"\battente\b", r"\bcaisse[s]?\b"],
}

# ====================== Sentiment (règles) ======================

POS_WORDS = {
    "bien":1.0,"super":1.2,"parfait":1.4,"excellent":1.5,"ravi":1.2,"satisfait":1.0,
    "rapide":0.8,"efficace":1.0,"fiable":1.0,"simple":0.8,"facile":0.8,"clair":0.8,"conforme":0.8,
    "sympa":0.8,"professionnel":1.0,"réactif":1.0,"reactif":1.0,"compétent":1.0,"competent":1.0,
    "top":1.2,"recommande":1.2,"recommandé":1.2,"bon":0.8
}
NEG_WORDS = {
    "mauvais":-1.2,"horrible":-1.5,"nul":-1.2,"lent":-0.8,"cher":-0.9,"arnaque":-1.5,
    "déçu":-1.2,"decu":-1.2,"incompétent":-1.3,"bug":-0.9,"bogue":-0.9,"problème":-1.0,
    "probleme":-1.0,"attente":-0.6,"retard":-0.9,"erreur":-1.0,"compliqué":-0.8,"complique":-0.8,
    "défectueux":-1.3,"defectueux":-1.3,"non conforme":-1.2,"impossible":-1.0,"difficile":-0.7
}
NEGATIONS = [r"\bpas\b", r"\bjamais\b", r"\bplus\b", r"\baucun[e]?\b", r"\brien\b", r"\bni\b", r"\bgu[eè]re\b"]
INTENSIFIERS = [r"\btr[eè]s\b", r"\bvraiment\b", r"\bextr[eê]mement\b", r"\bhyper\b"]
DIMINISHERS  = [r"\bun[e]?\s+peu\b", r"\bassez\b", r"\bplut[oô]t\b", r"\bl[eé]g[eè]rement\b"]
INTENSIFIER_W, DIMINISHER_W = 1.5, 0.7

# ====================== OpenAI (auto) ======================

OPENAI_AVAILABLE = False
try:
    from openai import OpenAI
    if os.getenv("OPENAI_API_KEY"):
        _client = OpenAI()
        OPENAI_AVAILABLE = True
except Exception:
    OPENAI_AVAILABLE = False

OA_MODEL = os.getenv("OPENAI_MODEL", "gpt-4o-mini")
OA_TEMP  = float(os.getenv("OPENAI_TEMP", "0.1"))
TOP_K    = int(os.getenv("VERBATIFY_TOPK", "10"))

# ====================== Utils ======================

def normalize(t:str)->str:
    if not isinstance(t,str): return ""
    return re.sub(r"\s+"," ",t.strip())

def to_analyzable(t:str)->str:
    return unidecode(normalize(t.lower()))

def window_has(patterns:List[str], toks:List[str], i:int, w:int=3)->bool:
    s=max(0,i-w); e=min(len(toks),i+w+1); win=" ".join(toks[s:e])
    return any(re.search(p,win) for p in patterns)

def lexical_sentiment_score(text:str)->float:
    toks = to_analyzable(text).split(); score=0.0
    for i,t in enumerate(toks):
        base = POS_WORDS.get(t,0.0) or NEG_WORDS.get(t,0.0)
        if not base and i<len(toks)-1:
            bi=f"{t} {toks[i+1]}"; base = NEG_WORDS.get(bi,0.0)
        if base:
            w=1.0
            if window_has(INTENSIFIERS,toks,i): w*=INTENSIFIER_W
            if window_has(DIMINISHERS,toks,i):  w*=DIMINISHER_W
            if window_has(NEGATIONS,toks,i):    base*=-1
            score+=base*w
    return max(min(score,4.0),-4.0)

def lexical_sentiment_label(s:float)->str:
    return "positive" if s>=0.3 else ("negatif" if s<=-0.3 else "neutre")

def detect_themes_regex(text:str):
    t=to_analyzable(text); counts={}
    for th,pats in THEMES.items():
        c=sum(len(re.findall(p,t)) for p in pats)
        if c>0: counts[th]=c
    return list(counts.keys()), counts

def nps_bucket(s):
    try:
        v=int(s)
    except:
        return "inconnu"
    return "promoter" if v>=9 else ("passive" if v>=7 else ("detractor" if v>=0 else "inconnu"))

def compute_nps(series):
    vals=[]
    for x in series.dropna().tolist():
        try:
            v=int(x)
            if 0<=v<=10: vals.append(v)
        except: pass
    if not vals: return None
    tot=len(vals); pro=sum(1 for v in vals if v>=9); det=sum(1 for v in vals if v<=6)
    return 100.0*(pro/tot - det/tot)

def anonymize(t:str)->str:
    if not isinstance(t,str): return ""
    t=re.sub(r"[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Za-z]{2,}","[email]",t)
    t=re.sub(r"\b(?:\+?\d[\s.-]?){7,}\b","[tel]",t)
    return t

# --------- Coller du texte → DataFrame (AUTO : détecte "| note" en fin de ligne) ----------
def df_from_pasted_auto(text:str) -> pd.DataFrame:
    lines = [l.strip() for l in (text or "").splitlines() if l.strip()]
    rows = []
    pat = re.compile(r"\|\s*(-?\d{1,2})\s*$")
    for i, line in enumerate(lines, 1):
        m = pat.search(line)
        if m:
            verb = line[:m.start()].strip()
            score = m.group(1)
            rows.append({"id": i, "comment": verb, "nps_score": pd.to_numeric(score, errors="coerce")})
        else:
            rows.append({"id": i, "comment": line, "nps_score": None})
    return pd.DataFrame(rows)

# --------- OpenAI helpers (auto) ----------
def openai_json(model:str, system:str, user:str, temperature:float=0.0) -> Optional[dict]:
    if not OPENAI_AVAILABLE: return None
    try:
        resp = _client.chat.completions.create(
            model=model, temperature=temperature,
            messages=[{"role":"system","content":system},{"role":"user","content":user}],
        )
        txt = resp.choices[0].message.content.strip()
        m = re.search(r"\{.*\}", txt, re.S)
        return json.loads(m.group(0) if m else txt)
    except Exception:
        return None

def oa_sentiment(comment:str) -> Optional[dict]:
    system = "Tu es un classifieur FR. Réponds strictement en JSON."
    user = f'Texte: {comment}\nDonne "label" parmi ["positive","neutre","negatif"] et "score" entre -4 et 4. JSON.'
    return openai_json(OA_MODEL, system, user, OA_TEMP)

def oa_themes(comment:str) -> Optional[dict]:
    system = "Tu maps le texte client vers un thésaurus assurance. Réponds strictement en JSON."
    user = f"Texte: {comment}\nThésaurus: {json.dumps(list(THEMES.keys()), ensure_ascii=False)}\nRetourne {{'themes': [...], 'counts': {{...}}}}"
    return openai_json(OA_MODEL, system, user, OA_TEMP)

def oa_summary(nps:Optional[float], dist:Dict[str,int], themes_df:pd.DataFrame) -> Optional[str]:
    system = "Tu es un analyste CX FR. Donne une synthèse courte et actionnable en Markdown."
    top = [] if themes_df is None else themes_df.head(6).to_dict(orient="records")
    user = f"Données: NPS={None if nps is None else round(nps,1)}, Répartition={dist}, Thèmes={json.dumps(top, ensure_ascii=False)}"
    j = openai_json(OA_MODEL, system, user, 0.2)
    if isinstance(j, dict) and "text" in j: return j["text"]
    if isinstance(j, dict): return ' '.join(str(v) for v in j.values())
    return None

# --------- HF sentiment (optionnel)
def make_hf_pipe():
    try:
        from transformers import pipeline
        return pipeline("text-classification",
                        model="cmarkea/distilcamembert-base-sentiment",
                        tokenizer="cmarkea/distilcamembert-base-sentiment")
    except Exception:
        return None

# --------- Inférence NPS depuis le sentiment ----------
def infer_nps_from_sentiment(label: str, score: float) -> int:
    scaled = int(round((float(score) + 4.0) * 1.25))  # -4 -> 0, 0 -> 5, +4 -> 10
    scaled = max(0, min(10, scaled))
    if label == "positive":
        return max(9, scaled)
    if label == "negatif":
        return min(6, scaled)
    return 8 if score >= 0 else 7

# ====================== Graphiques ======================

def fig_nps_gauge(nps: Optional[float]) -> go.Figure:
    v = 0.0 if nps is None else float(nps)
    return go.Figure(go.Indicator(mode="gauge+number", value=v,
                                  gauge={"axis":{"range":[-100,100]}, "bar":{"thickness":0.3}},
                                  title={"text":"NPS (−100 à +100)"}))

def fig_sentiment_bar(dist: Dict[str,int]) -> go.Figure:
    order = ["negatif","neutre","positive"]
    x = [o for o in order if o in dist]; y = [dist.get(o,0) for o in x]
    return px.bar(x=x, y=y, labels={"x":"Sentiment","y":"Nombre"}, title="Répartition des émotions")

def fig_top_themes(themes_df: pd.DataFrame, k: int) -> go.Figure:
    if themes_df is None or themes_df.empty: return go.Figure()
    d = themes_df.head(k); fig = px.bar(d, x="theme", y="total_mentions", title=f"Top {k} thèmes — occurrences")
    fig.update_layout(xaxis_tickangle=-30); return fig

def fig_theme_balance(themes_df: pd.DataFrame, k: int) -> go.Figure:
    if themes_df is None or themes_df.empty: return go.Figure()
    d = themes_df.head(k)
    d2 = d.melt(id_vars=["theme"], value_vars=["verbatims_pos","verbatims_neg"], var_name="type", value_name="count")
    d2["type"] = d2["type"].map({"verbatims_pos":"Positifs","verbatims_neg":"Négatifs"})
    fig = px.bar(d2, x="theme", y="count", color="type", barmode="stack", title=f"Top {k} thèmes — balance Pos/Neg")
    fig.update_layout(xaxis_tickangle=-30); return fig

# ====================== Analyse principale (AUTO) ======================

def analyze_text(pasted_txt: str):
    # 1) Parse auto + anonymisation
    df = df_from_pasted_auto(pasted_txt or "")
    if df.empty:
        raise gr.Error("Colle au moins un verbatim (une ligne).")
    df["comment"] = df["comment"].apply(anonymize)

    # 2) Pipes
    use_oa_sent = use_oa_themes = use_oa_summary = True
    if not OPENAI_AVAILABLE:
        use_oa_sent = use_oa_themes = use_oa_summary = False
    hf_pipe = make_hf_pipe()

    # 3) Boucle verbatims
    rows=[]; theme_agg=collections.defaultdict(lambda:{"mentions":0,"pos":0,"neg":0})
    used_hf=False; used_oa=False; any_inferred=False

    for idx, r in df.iterrows():
        cid=r.get("id", idx+1); comment=normalize(str(r["comment"]))

        # Sentiment: OpenAI -> HF -> règles
        sent=None
        if use_oa_sent:
            sent=oa_sentiment(comment); used_oa = used_oa or bool(sent)
        if not sent and hf_pipe is not None and comment.strip():
            try:
                res=hf_pipe(comment); lab=str(res[0]["label"]).lower(); p=float(res[0].get("score",0.5))
                if "1" in lab or "2" in lab: sent = {"label":"negatif","score":-4*p}
                elif "3" in lab:           sent = {"label":"neutre","score":0.0}
                else:                      sent = {"label":"positive","score":4*p}
                used_hf=True
            except Exception:
                sent=None
        if not sent:
            s=float(lexical_sentiment_score(comment))
            sent={"label":lexical_sentiment_label(s),"score":s}

        # Thèmes: regex (+ fusion OpenAI)
        themes, counts = detect_themes_regex(comment)
        if use_oa_themes:
            tjson=oa_themes(comment)
            if isinstance(tjson, dict):
                used_oa=True
                for th, c in (tjson.get("counts",{}) or {}).items():
                    if th in THEMES and int(c) > 0:
                        counts[th] = max(counts.get(th, 0), int(c))
                themes = [th for th, c in counts.items() if c > 0]

        # Note NPS existante ou inférée
        given = r.get("nps_score", None)
        try:
            given = int(given) if given is not None and str(given).strip() != "" else None
        except Exception:
            given = None

        if given is None:
            inferred = infer_nps_from_sentiment(sent["label"], float(sent["score"]))
            nps_final, nps_source, any_inferred = inferred, "inferred", True
        else:
            nps_final, nps_source = given, "given"

        bucket = nps_bucket(nps_final)

        for th, c in counts.items():
            theme_agg[th]["mentions"] += c
            if sent["label"] == "positive": theme_agg[th]["pos"] += 1
            elif sent["label"] == "negatif": theme_agg[th]["neg"] += 1

        rows.append({
            "id": cid, "comment": comment,
            "nps_score_given": given, "nps_score_inferred": nps_final if given is None else None,
            "nps_score_final": nps_final, "nps_source": nps_source, "nps_bucket": bucket,
            "sentiment_score": round(float(sent["score"]), 3), "sentiment_label": sent["label"],
            "sentiment_source": "openai" if (use_oa_sent and used_oa) else ("huggingface" if used_hf else "rules"),
            "themes": ", ".join(themes) if themes else "", "theme_counts_json": json.dumps(counts, ensure_ascii=False)
        })

    out_df=pd.DataFrame(rows)
    nps=compute_nps(out_df["nps_score_final"])
    dist=out_df["sentiment_label"].value_counts().to_dict()

    # Stats par thème
    trs=[]
    for th, d in theme_agg.items():
        trs.append({"theme":th,"total_mentions":int(d["mentions"]),
                    "verbatims_pos":int(d["pos"]),"verbatims_neg":int(d["neg"]),
                    "net_sentiment":int(d["pos"]-d["neg"])})
    themes_df=pd.DataFrame(trs).sort_values(["total_mentions","net_sentiment"],ascending=[False,False])

    # Synthèse
    method = "OpenAI + HF + règles" if (OPENAI_AVAILABLE and used_hf) else ("OpenAI + règles" if OPENAI_AVAILABLE else ("HF + règles" if used_hf else "Règles"))
    nps_label = "NPS global (inféré)" if any_inferred else "NPS global"
    lines=[ "# Synthèse NPS & ressentis clients",
            f"- **Méthode** : {method}",
            f"- **{nps_label}** : {nps:.1f}" if nps is not None else f"- **{nps_label}** : n/a" ]
    if dist:
        tot=sum(dist.values()); pos=dist.get("positive",0); neg=dist.get("negatif",0); neu=dist.get("neutre",0)
        lines.append(f"- **Répartition émotions** : positive {pos}/{tot}, neutre {neu}/{tot}, négative {neg}/{tot}")
    if not themes_df.empty:
        lines.append("\n## Thèmes les plus cités")
        for th,m in themes_df.head(5)[["theme","total_mentions"]].values.tolist():
            lines.append(f"- **{th}** : {m} occurrence(s)")
    summary_md="\n".join(lines)

    if OPENAI_AVAILABLE:
        md = oa_summary(nps, dist, themes_df)
        if md: summary_md = md + "\n\n---\n" + summary_md

    # Exports
    tmpdir=tempfile.mkdtemp(prefix="nps_gradio_")
    enriched=os.path.join(tmpdir,"enriched_comments.csv")
    themes=os.path.join(tmpdir,"themes_stats.csv")
    summ=os.path.join(tmpdir,"summary.md")
    out_df.to_csv(enriched,index=False,encoding="utf-8-sig")
    themes_df.to_csv(themes,index=False,encoding="utf-8-sig")
    with open(summ,"w",encoding="utf-8") as f: f.write(summary_md)
    zip_path=os.path.join(tmpdir,"nps_outputs.zip")
    with zipfile.ZipFile(zip_path,"w",zipfile.ZIP_DEFLATED) as z:
        z.write(enriched,arcname="enriched_comments.csv")
        z.write(themes,arcname="themes_stats.csv")
        z.write(summ,arcname="summary.md")

    # Panneaux & Graphes
    def make_panels(dfT: pd.DataFrame):
        if dfT is None or dfT.empty: return "—","—","—"
        pos_top = dfT.sort_values(["verbatims_pos","total_mentions"], ascending=[False,False]).head(4)
        neg_top = dfT.sort_values(["verbatims_neg","total_mentions"], ascending=[False,False]).head(4)
        def bullets(df, col, label):
            lines=[f"**{label}**"]
            for _, r in df.iterrows(): lines.append(f"- **{r['theme']}** — {int(r[col])} verbatims")
            return "\n".join(lines)
        ench_md = bullets(pos_top, "verbatims_pos", "Points d’enchantement")
        irr_md  = bullets(neg_top, "verbatims_neg", "Irritants")
        RECO_RULES = {
            "Délais & Suivi dossier": "Réduire les délais (SLA), suivi proactif.",
            "Cotisations / Facturation": "Clarifier factures, alerter anomalies.",
            "Espace client / App / Connexion": "Corriger login/MDP, QA navigateurs.",
            "Support / Conseiller": "Améliorer joignabilité, scripts, rappel auto.",
            "Communication / Transparence": "Notifications étapes clés, messages clairs.",
            "Sinistres / Indemnisation": "Transparence délais + suivi dossier.",
        }
        rec_lines=["**Recommandations**"]
        for _, r in neg_top.iterrows():
            rec_lines.append(f"- **{r['theme']}** — {RECO_RULES.get(r['theme'],'Plan d’action dédié')}")
        return ench_md, irr_md, "\n".join(rec_lines)

    ench_md, irr_md, reco_md = make_panels(themes_df)
    fig_gauge = fig_nps_gauge(nps)
    fig_emots = fig_sentiment_bar(dist)
    k = max(1, int(TOP_K))
    fig_top   = fig_top_themes(themes_df, k)
    fig_bal   = fig_theme_balance(themes_df, k)

    return (summary_md, themes_df.head(100), out_df.head(200), [enriched, themes, summ, zip_path],
            ench_md, irr_md, reco_md, fig_gauge, fig_emots, fig_top, fig_bal)

# ====================== UI ======================

apply_plotly_theme()

with gr.Blocks(title="Verbatify, révélez la voix de vos assurés, simplement...", css=VB_CSS) as demo:
    # Header
    gr.HTML(
        "<div class='vb-hero'>"
        f"{LOGO_SVG}"
        "<div><div class='vb-title'>Verbatify, révélez la voix de vos assurés, simplement...</div>"
        "<div class='vb-sub'>Émotions • Thématiques • Occurrences • Synthèse • NPS</div></div>"
        "</div>"
    )

    # Entrée minimale + bouton
    with gr.Row():
        pasted = gr.Textbox(
            label="Verbatims (un par ligne)", lines=10,
            placeholder="Exemple :\nRemboursement rapide, télétransmission OK | 10\nImpossible de joindre un conseiller | 3\nEspace client : bug à la connexion | 4",
            scale=4
        )
           # cellule dédiée au bouton, qu'on va centrer avec le CSS
    with gr.Column(elem_id="vb-cta-cell", scale=1):
        run = gr.Button("Lancer l'analyse", elem_classes=["vb-cta"])

    # Panneaux
    with gr.Row():
        ench_panel=gr.Markdown()
        irr_panel=gr.Markdown()
        reco_panel=gr.Markdown()

    # Résultats + téléchargements
    summary=gr.Markdown(label="Synthèse NPS & ressentis clients")
    themes_table=gr.Dataframe(label="Thèmes — statistiques")
    enriched_table=gr.Dataframe(label="Verbatims enrichis (aperçu)")
    files_out=gr.Files(label="Téléchargements (CSV & ZIP)")

    # Graphes
    with gr.Row():
        plot_nps = gr.Plot(label="NPS — Jauge")
        plot_sent= gr.Plot(label="Répartition des émotions")
    with gr.Row():
        plot_top  = gr.Plot(label="Top thèmes — occurrences")
        plot_bal  = gr.Plot(label="Top thèmes — balance Pos/Neg")

    # Lancer
    run.click(
        analyze_text,
        inputs=[pasted],
        outputs=[summary, themes_table, enriched_table, files_out,
                 ench_panel, irr_panel, reco_panel,
                 plot_nps, plot_sent, plot_top, plot_bal]
    )

    gr.HTML(
        '<div class="vb-footer">© Verbatify.com — Construit par '
        '<a href="https://jeremy-lagache.fr/" target="_blank" rel="dofollow">Jérémy Lagache</a></div>'
    )

if __name__ == "__main__":
    demo.launch(share=False, show_api=False)