Upload app.py
Browse files
app.py
ADDED
@@ -0,0 +1,410 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# -*- coding: utf-8 -*-
|
2 |
+
"""
|
3 |
+
NPS Assurance — Gradio (Paste-only)
|
4 |
+
- Entrée: verbatims collés (1 ligne = 1 verbatim, score NPS optionnel après un séparateur, ex: "|")
|
5 |
+
- Sorties: émotion (pos/neutre/neg), thématiques, occurrences, résumé Markdown, graphiques Plotly
|
6 |
+
- IA (facultatif): OpenAI pour sentiment/thèmes/synthèse. Sans clé, fallback HuggingFace (si installé) puis règles lexicales.
|
7 |
+
- Déployable tel quel sur Hugging Face Spaces (app_file = app.py)
|
8 |
+
"""
|
9 |
+
|
10 |
+
import os, re, json, collections, tempfile, zipfile
|
11 |
+
from typing import List, Dict, Optional
|
12 |
+
import pandas as pd
|
13 |
+
from unidecode import unidecode
|
14 |
+
import gradio as gr
|
15 |
+
import plotly.express as px
|
16 |
+
import plotly.graph_objects as go
|
17 |
+
|
18 |
+
# ---------------- Thésaurus ASSURANCE ----------------
|
19 |
+
THEMES = {
|
20 |
+
"Remboursements santé":[r"\bremboursement[s]?\b", r"\bt[eé]l[eé]transmission\b", r"\bno[eé]mie\b",
|
21 |
+
r"\bprise\s*en\s*charge[s]?\b", r"\btaux\s+de\s+remboursement[s]?\b", r"\b(ameli|cpam)\b",
|
22 |
+
r"\bcompl[eé]mentaire\s+sant[eé]\b", r"\bmutuelle\b", r"\battestation[s]?\b", r"\bcarte\s+(mutuelle|tiers\s*payant)\b"],
|
23 |
+
"Tiers payant / Réseau de soins":[r"\btiers\s*payant\b", r"\br[ée]seau[x]?\s+de\s+soins\b",
|
24 |
+
r"\b(optique|dentaire|hospitalisation|pharmacie)\b", r"\bitelis\b", r"\bsant[eé]clair\b", r"\bkalixia\b"],
|
25 |
+
"Sinistres / Indemnisation":[r"\bsinistre[s]?\b", r"\bindemni(sation|ser)\b", r"\bexpertis[ea]\b",
|
26 |
+
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"],
|
27 |
+
"Adhésion / Contrat":[r"\badh[eé]sion[s]?\b", r"\bsouscription[s]?\b", r"\baffiliation[s]?\b", r"\bcontrat[s]?\b",
|
28 |
+
r"\bavenant[s]?\b", r"\bcarence[s]?\b", r"\brenouvellement[s]?\b", r"\br[eé]siliation[s]?\b"],
|
29 |
+
"Garanties / Exclusions / Franchise":[r"\bgarantie[s]?\b", r"\bexclusion[s]?\b", r"\bplafond[s]?\b",
|
30 |
+
r"\bfranchise[s]?\b", r"\bconditions\s+g[eé]n[eé]rales\b", r"\bnotice\b"],
|
31 |
+
"Cotisations / Facturation":[r"\bcotisation[s]?\b", r"\bpr[eé]l[eè]vement[s]?\b", r"\bech[eé]ancier[s]?\b",
|
32 |
+
r"\bfacture[s]?\b", r"\berreur[s]?\s+de\s+facturation\b", r"\bremboursement[s]?\b", r"\bRIB\b", r"\bIBAN\b"],
|
33 |
+
"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"],
|
34 |
+
"Espace client / App / Connexion":[r"\bespace\s+client\b", r"\bapplication\b", r"\bapp\b", r"\bsite\b",
|
35 |
+
r"\bconnexion\b", r"\bidentifiant[s]?\b", r"\bmot\s+de\s+passe\b", r"\bpaiement\s+en\s+ligne\b",
|
36 |
+
r"\bbogue[s]?\b", r"\bbug[s]?\b", r"\bnavigation\b", r"\binterface\b", r"\bUX\b"],
|
37 |
+
"Support / Conseiller":[r"\bSAV\b", r"\bservice[s]?\s+client[s]?\b", r"\bconseiller[s]?\b",
|
38 |
+
r"\b[rR][eé]ponse[s]?\b", r"\bjoignable[s]?\b", r"\brapp?el\b"],
|
39 |
+
"Communication / Transparence":[r"\binformation[s]?\b", r"\bcommunication\b", r"\btransparence\b",
|
40 |
+
r"\bclart[eé]\b", r"\bcourrier[s]?\b", r"\bmail[s]?\b", r"\bnotification[s]?\b"],
|
41 |
+
"Prix":[r"\bprix\b", r"\bcher[s]?\b", r"\bco[uû]t[s]?\b", r"\btarif[s]?\b",
|
42 |
+
r"\bcomp[eé]titif[s]?\b", r"\babusif[s]?\b", r"\bbon\s+rapport\s+qualit[eé]\s*prix\b"],
|
43 |
+
"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"],
|
44 |
+
"Produit/Qualité":[r"\bqualit[eé]s?\b", r"\bfiable[s]?\b", r"\bconforme[s]?\b", r"\bnon\s+conforme[s]?\b",
|
45 |
+
r"\bd[eé]fectueux?[es]?\b", r"\bperformant[e]?[s]?\b"],
|
46 |
+
"Agence / Accueil":[r"\bagence[s]?\b", r"\bboutique[s]?\b", r"\baccueil\b", r"\bconseil[s]?\b", r"\battente\b", r"\bcaisse[s]?\b"],
|
47 |
+
}
|
48 |
+
|
49 |
+
# --------------- Sentiment (fallback règles) ---------------
|
50 |
+
POS_WORDS = {"bien":1.0,"super":1.2,"parfait":1.4,"excellent":1.5,"ravi":1.2,"satisfait":1.0,
|
51 |
+
"rapide":0.8,"efficace":1.0,"fiable":1.0,"simple":0.8,"facile":0.8,"clair":0.8,"conforme":0.8,
|
52 |
+
"sympa":0.8,"professionnel":1.0,"réactif":1.0,"reactif":1.0,"compétent":1.0,"competent":1.0,
|
53 |
+
"top":1.2,"recommande":1.2,"recommandé":1.2,"bon":0.8}
|
54 |
+
NEG_WORDS = {"mauvais":-1.2,"horrible":-1.5,"nul":-1.2,"lent":-0.8,"cher":-0.9,"arnaque":-1.5,
|
55 |
+
"déçu":-1.2,"decu":-1.2,"incompétent":-1.3,"bug":-0.9,"bogue":-0.9,"problème":-1.0,
|
56 |
+
"probleme":-1.0,"attente":-0.6,"retard":-0.9,"erreur":-1.0,"compliqué":-0.8,"complique":-0.8,
|
57 |
+
"défectueux":-1.3,"defectueux":-1.3,"non conforme":-1.2,"impossible":-1.0,"difficile":-0.7}
|
58 |
+
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"]
|
59 |
+
INTENSIFIERS = [r"\btr[eè]s\b", r"\bvraiment\b", r"\bextr[eê]mement\b", r"\bhyper\b"]
|
60 |
+
DIMINISHERS = [r"\bun[e]?\s+peu\b", r"\bassez\b", r"\bplut[oô]t\b", r"\bl[eé]g[eè]rement\b"]
|
61 |
+
INTENSIFIER_W, DIMINISHER_W = 1.5, 0.7
|
62 |
+
|
63 |
+
# --------------- OpenAI (optionnel) ---------------
|
64 |
+
OPENAI_AVAILABLE = False
|
65 |
+
try:
|
66 |
+
from openai import OpenAI
|
67 |
+
_client = OpenAI() # clé via OPENAI_API_KEY (en secret HF)
|
68 |
+
OPENAI_AVAILABLE = True
|
69 |
+
except Exception:
|
70 |
+
OPENAI_AVAILABLE = False
|
71 |
+
|
72 |
+
# ---------------- Utils ----------------
|
73 |
+
def normalize(t:str)->str:
|
74 |
+
if not isinstance(t,str): return ""
|
75 |
+
return re.sub(r"\s+"," ",t.strip())
|
76 |
+
|
77 |
+
def to_analyzable(t:str)->str:
|
78 |
+
return unidecode(normalize(t.lower()))
|
79 |
+
|
80 |
+
def window_has(patterns:List[str], toks:List[str], i:int, w:int=3)->bool:
|
81 |
+
s=max(0,i-w); e=min(len(toks),i+w+1); win=" ".join(toks[s:e])
|
82 |
+
return any(re.search(p,win) for p in patterns)
|
83 |
+
|
84 |
+
def lexical_sentiment_score(text:str)->float:
|
85 |
+
toks = to_analyzable(text).split(); score=0.0
|
86 |
+
for i,t in enumerate(toks):
|
87 |
+
base = POS_WORDS.get(t,0.0) or NEG_WORDS.get(t,0.0)
|
88 |
+
if not base and i<len(toks)-1:
|
89 |
+
bi=f"{t} {toks[i+1]}"; base = NEG_WORDS.get(bi,0.0)
|
90 |
+
if base:
|
91 |
+
w=1.0
|
92 |
+
if window_has(INTENSIFIERS,toks,i): w*=INTENSIFIER_W
|
93 |
+
if window_has(DIMINISHERS,toks,i): w*=DIMINISHER_W
|
94 |
+
if window_has(NEGATIONS,toks,i): base*=-1
|
95 |
+
score+=base*w
|
96 |
+
return max(min(score,4.0),-4.0)
|
97 |
+
|
98 |
+
def lexical_sentiment_label(s:float)->str:
|
99 |
+
return "positive" if s>=0.3 else ("negatif" if s<=-0.3 else "neutre")
|
100 |
+
|
101 |
+
def detect_themes_regex(text:str):
|
102 |
+
t=to_analyzable(text); counts={}
|
103 |
+
for th,pats in THEMES.items():
|
104 |
+
c=sum(len(re.findall(p,t)) for p in pats)
|
105 |
+
if c>0: counts[th]=c
|
106 |
+
return list(counts.keys()), counts
|
107 |
+
|
108 |
+
def nps_bucket(s):
|
109 |
+
try:
|
110 |
+
v=int(s)
|
111 |
+
except:
|
112 |
+
return "inconnu"
|
113 |
+
return "promoter" if v>=9 else ("passive" if v>=7 else ("detractor" if v>=0 else "inconnu"))
|
114 |
+
|
115 |
+
def compute_nps(series):
|
116 |
+
vals=[]
|
117 |
+
for x in series.dropna().tolist():
|
118 |
+
try:
|
119 |
+
v=int(x)
|
120 |
+
if 0<=v<=10: vals.append(v)
|
121 |
+
except: pass
|
122 |
+
if not vals: return None
|
123 |
+
tot=len(vals); pro=sum(1 for v in vals if v>=9); det=sum(1 for v in vals if v<=6)
|
124 |
+
return 100.0*(pro/tot - det/tot)
|
125 |
+
|
126 |
+
def anonymize(t:str)->str:
|
127 |
+
if not isinstance(t,str): return ""
|
128 |
+
t=re.sub(r"[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Za-z]{2,}","[email]",t)
|
129 |
+
t=re.sub(r"\b(?:\+?\d[\s.-]?){7,}\b","[tel]",t)
|
130 |
+
return t
|
131 |
+
|
132 |
+
# --------- Coller du texte → DataFrame ----------
|
133 |
+
def df_from_pasted(text:str, sep="|", has_score=False) -> pd.DataFrame:
|
134 |
+
lines = [l.strip() for l in (text or "").splitlines() if l.strip()]
|
135 |
+
rows = []
|
136 |
+
for i, line in enumerate(lines, 1):
|
137 |
+
if has_score and sep in line:
|
138 |
+
verb, score = line.split(sep, 1)
|
139 |
+
rows.append({"id": i, "comment": verb.strip(), "nps_score": pd.to_numeric(score.strip(), errors="coerce")})
|
140 |
+
else:
|
141 |
+
rows.append({"id": i, "comment": line.strip(), "nps_score": None})
|
142 |
+
return pd.DataFrame(rows)
|
143 |
+
|
144 |
+
# --------- OpenAI helpers (optionnels) ----------
|
145 |
+
def openai_json(model:str, system:str, user:str, temperature:float=0.0) -> Optional[dict]:
|
146 |
+
if not OPENAI_AVAILABLE: return None
|
147 |
+
try:
|
148 |
+
resp = _client.chat.completions.create(
|
149 |
+
model=model, temperature=temperature,
|
150 |
+
messages=[{"role":"system","content":system},{"role":"user","content":user}],
|
151 |
+
)
|
152 |
+
txt = resp.choices[0].message.content.strip()
|
153 |
+
m = re.search(r"\{.*\}", txt, re.S)
|
154 |
+
return json.loads(m.group(0) if m else txt)
|
155 |
+
except Exception:
|
156 |
+
return None
|
157 |
+
|
158 |
+
def oa_sentiment(comment:str, model:str, temperature:float=0.0) -> Optional[dict]:
|
159 |
+
system = "Tu es un classifieur FR. Réponds strictement en JSON."
|
160 |
+
user = f'Texte: {comment}\nDonne "label" parmi ["positive","neutre","negatif"] et "score" entre -4 et 4. JSON.'
|
161 |
+
return openai_json(model, system, user, temperature)
|
162 |
+
|
163 |
+
def oa_themes(comment:str, model:str, temperature:float=0.0) -> Optional[dict]:
|
164 |
+
system = "Tu maps le texte client vers un thésaurus assurance. Réponds strictement en JSON."
|
165 |
+
user = f"Texte: {comment}\nThésaurus: {json.dumps(list(THEMES.keys()), ensure_ascii=False)}\nRetourne {{'themes': [...], 'counts': {{...}}}}"
|
166 |
+
return openai_json(model, system, user, temperature)
|
167 |
+
|
168 |
+
def oa_summary(nps:Optional[float], dist:Dict[str,int], themes_df:pd.DataFrame, model:str, temperature:float=0.2) -> Optional[str]:
|
169 |
+
system = "Tu es un analyste CX FR. Donne une synthèse courte et actionnable en Markdown."
|
170 |
+
top = [] if themes_df is None else themes_df.head(6).to_dict(orient="records")
|
171 |
+
user = f"Données: NPS={None if nps is None else round(nps,1)}, Répartition={dist}, Thèmes={json.dumps(top, ensure_ascii=False)}"
|
172 |
+
j = openai_json(model, system, user, temperature)
|
173 |
+
if isinstance(j, dict) and "text" in j: return j["text"]
|
174 |
+
if isinstance(j, dict): return ' '.join(str(v) for v in j.values())
|
175 |
+
return None
|
176 |
+
|
177 |
+
# --------- Graphiques ----------
|
178 |
+
def fig_nps_gauge(nps: Optional[float]) -> go.Figure:
|
179 |
+
v = 0.0 if nps is None else float(nps)
|
180 |
+
return go.Figure(go.Indicator(mode="gauge+number", value=v,
|
181 |
+
gauge={"axis":{"range":[-100,100]}, "bar":{"thickness":0.3}},
|
182 |
+
title={"text":"NPS (−100 à +100)"}))
|
183 |
+
|
184 |
+
def fig_sentiment_bar(dist: Dict[str,int]) -> go.Figure:
|
185 |
+
order = ["negatif","neutre","positive"]
|
186 |
+
x = [o for o in order if o in dist]; y = [dist.get(o,0) for o in x]
|
187 |
+
return px.bar(x=x, y=y, labels={"x":"Sentiment","y":"Nombre"}, title="Répartition des émotions")
|
188 |
+
|
189 |
+
def fig_top_themes(themes_df: pd.DataFrame, k: int) -> go.Figure:
|
190 |
+
if themes_df is None or themes_df.empty: return go.Figure()
|
191 |
+
d = themes_df.head(k); fig = px.bar(d, x="theme", y="total_mentions", title=f"Top {k} thèmes — occurrences")
|
192 |
+
fig.update_layout(xaxis_tickangle=-30); return fig
|
193 |
+
|
194 |
+
def fig_theme_balance(themes_df: pd.DataFrame, k: int) -> go.Figure:
|
195 |
+
if themes_df is None or themes_df.empty: return go.Figure()
|
196 |
+
d = themes_df.head(k)
|
197 |
+
d2 = d.melt(id_vars=["theme"], value_vars=["verbatims_pos","verbatims_neg"], var_name="type", value_name="count")
|
198 |
+
d2["type"] = d2["type"].map({"verbatims_pos":"Positifs","verbatims_neg":"Négatifs"})
|
199 |
+
fig = px.bar(d2, x="theme", y="count", color="type", barmode="stack", title=f"Top {k} thèmes — balance Pos/Neg")
|
200 |
+
fig.update_layout(xaxis_tickangle=-30); return fig
|
201 |
+
|
202 |
+
# --------- Analyse principale ----------
|
203 |
+
def analyze_text(pasted_txt, has_sc, sep_chr,
|
204 |
+
do_anonymize, use_oa_sent, use_oa_themes, use_oa_summary,
|
205 |
+
oa_model, oa_temp, top_k):
|
206 |
+
|
207 |
+
df = df_from_pasted(pasted_txt or "", sep=sep_chr or "|", has_score=bool(has_sc))
|
208 |
+
if df.empty:
|
209 |
+
raise gr.Error("Colle au moins un verbatim (une ligne).")
|
210 |
+
|
211 |
+
if do_anonymize:
|
212 |
+
df["comment"]=df["comment"].apply(anonymize)
|
213 |
+
|
214 |
+
if (use_oa_sent or use_oa_themes or use_oa_summary) and not OPENAI_AVAILABLE:
|
215 |
+
raise gr.Error("OpenAI non dispo : installe `openai` et définis OPENAI_API_KEY, ou décoche les options OpenAI.")
|
216 |
+
|
217 |
+
# HF sentiment (optionnel)
|
218 |
+
HF_AVAILABLE=False
|
219 |
+
try:
|
220 |
+
from transformers import pipeline
|
221 |
+
hf_pipe = pipeline("text-classification",
|
222 |
+
model="cmarkea/distilcamembert-base-sentiment",
|
223 |
+
tokenizer="cmarkea/distilcamembert-base-sentiment")
|
224 |
+
HF_AVAILABLE=True
|
225 |
+
except Exception:
|
226 |
+
HF_AVAILABLE=False
|
227 |
+
def hf_sent(text:str):
|
228 |
+
if not HF_AVAILABLE or not text.strip(): return None
|
229 |
+
try:
|
230 |
+
res=hf_pipe(text); lab=str(res[0]["label"]).lower(); p=float(res[0].get("score",0.5))
|
231 |
+
if "1" in lab or "2" in lab: return {"label":"negatif","score":-4*p}
|
232 |
+
if "3" in lab: return {"label":"neutre","score":0.0}
|
233 |
+
return {"label":"positive","score":4*p}
|
234 |
+
except Exception:
|
235 |
+
return None
|
236 |
+
|
237 |
+
rows=[]
|
238 |
+
theme_agg=collections.defaultdict(lambda:{"mentions":0,"pos":0,"neg":0})
|
239 |
+
used_hf=False; used_oa=False
|
240 |
+
|
241 |
+
for _, r in df.iterrows():
|
242 |
+
cid=r["id"]; comment=normalize(str(r["comment"]))
|
243 |
+
|
244 |
+
# Sentiment: OpenAI -> HF -> règles
|
245 |
+
sent=None
|
246 |
+
if use_oa_sent:
|
247 |
+
sent=oa_sentiment(comment, oa_model, oa_temp); used_oa = used_oa or bool(sent)
|
248 |
+
if not sent:
|
249 |
+
hf=hf_sent(comment)
|
250 |
+
if hf: sent=hf; used_hf=True
|
251 |
+
if not sent:
|
252 |
+
s=float(lexical_sentiment_score(comment))
|
253 |
+
sent={"label":lexical_sentiment_label(s),"score":s}
|
254 |
+
|
255 |
+
# Thèmes: regex (+ fusion OpenAI)
|
256 |
+
themes, counts = detect_themes_regex(comment)
|
257 |
+
if use_oa_themes:
|
258 |
+
tjson=oa_themes(comment, oa_model, oa_temp)
|
259 |
+
if isinstance(tjson, dict):
|
260 |
+
used_oa=True
|
261 |
+
for th, c in (tjson.get("counts",{}) or {}).items():
|
262 |
+
if th in THEMES and int(c) > 0:
|
263 |
+
counts[th] = max(counts.get(th, 0), int(c))
|
264 |
+
themes = [th for th, c in counts.items() if c > 0]
|
265 |
+
|
266 |
+
bucket = nps_bucket(r.get("nps_score", None))
|
267 |
+
|
268 |
+
for th, c in counts.items():
|
269 |
+
theme_agg[th]["mentions"] += c
|
270 |
+
if sent["label"] == "positive":
|
271 |
+
theme_agg[th]["pos"] += 1
|
272 |
+
elif sent["label"] == "negatif":
|
273 |
+
theme_agg[th]["neg"] += 1
|
274 |
+
|
275 |
+
rows.append({
|
276 |
+
"id": cid, "nps_score": r.get("nps_score", None), "nps_bucket": bucket,
|
277 |
+
"comment": comment,
|
278 |
+
"sentiment_score": round(float(sent["score"]), 3),
|
279 |
+
"sentiment_label": sent["label"],
|
280 |
+
"sentiment_source": "openai" if (use_oa_sent and used_oa) else ("huggingface" if used_hf else "rules"),
|
281 |
+
"themes": ", ".join(themes) if themes else "",
|
282 |
+
"theme_counts_json": json.dumps(counts, ensure_ascii=False)
|
283 |
+
})
|
284 |
+
|
285 |
+
out_df=pd.DataFrame(rows)
|
286 |
+
nps=compute_nps(df["nps_score"]) # peut être None si pas de scores
|
287 |
+
dist=out_df["sentiment_label"].value_counts().to_dict()
|
288 |
+
|
289 |
+
# Stats par thème
|
290 |
+
trs=[]
|
291 |
+
for th, d in theme_agg.items():
|
292 |
+
trs.append({"theme":th,"total_mentions":int(d["mentions"]),
|
293 |
+
"verbatims_pos":int(d["pos"]),"verbatims_neg":int(d["neg"]),
|
294 |
+
"net_sentiment":int(d["pos"]-d["neg"])})
|
295 |
+
themes_df=pd.DataFrame(trs).sort_values(["total_mentions","net_sentiment"],ascending=[False,False])
|
296 |
+
|
297 |
+
# Synthèse texte
|
298 |
+
method = "OpenAI + HF + règles" if (use_oa_sent and used_hf) else ("OpenAI + règles" if use_oa_sent else ("HF + règles" if used_hf else "Règles"))
|
299 |
+
lines=[ "# Synthèse NPS & ressentis clients",
|
300 |
+
f"- **Méthode** : {method}",
|
301 |
+
f"- **NPS global** : {nps:.1f}" if nps is not None else "- **NPS global** : n/a" ]
|
302 |
+
if dist:
|
303 |
+
tot=sum(dist.values()); pos=dist.get("positive",0); neg=dist.get("negatif",0); neu=dist.get("neutre",0)
|
304 |
+
lines.append(f"- **Répartition émotions** : positive {pos}/{tot}, neutre {neu}/{tot}, négative {neg}/{tot}")
|
305 |
+
if not themes_df.empty:
|
306 |
+
lines.append("\n## Thèmes les plus cités")
|
307 |
+
for th,m in themes_df.head(5)[["theme","total_mentions"]].values.tolist():
|
308 |
+
lines.append(f"- **{th}** : {m} occurrence(s)")
|
309 |
+
summary_md="\n".join(lines)
|
310 |
+
|
311 |
+
if use_oa_summary:
|
312 |
+
md = oa_summary(nps, dist, themes_df, oa_model, oa_temp)
|
313 |
+
if md: summary_md = md + "\n\n---\n" + summary_md
|
314 |
+
|
315 |
+
# Fichiers export
|
316 |
+
tmpdir=tempfile.mkdtemp(prefix="nps_gradio_")
|
317 |
+
enriched=os.path.join(tmpdir,"enriched_comments.csv")
|
318 |
+
themes=os.path.join(tmpdir,"themes_stats.csv")
|
319 |
+
summ=os.path.join(tmpdir,"summary.md")
|
320 |
+
out_df.to_csv(enriched,index=False,encoding="utf-8-sig")
|
321 |
+
themes_df.to_csv(themes,index=False,encoding="utf-8-sig")
|
322 |
+
with open(summ,"w",encoding="utf-8") as f: f.write(summary_md)
|
323 |
+
zip_path=os.path.join(tmpdir,"nps_outputs.zip")
|
324 |
+
with zipfile.ZipFile(zip_path,"w",zipfile.ZIP_DEFLATED) as z:
|
325 |
+
z.write(enriched,arcname="enriched_comments.csv")
|
326 |
+
z.write(themes,arcname="themes_stats.csv")
|
327 |
+
z.write(summ,arcname="summary.md")
|
328 |
+
|
329 |
+
# Graphiques
|
330 |
+
fig_gauge = fig_nps_gauge(nps)
|
331 |
+
fig_emots = fig_sentiment_bar(dist)
|
332 |
+
k = max(1, int(top_k or 10))
|
333 |
+
fig_top = fig_top_themes(themes_df, k)
|
334 |
+
fig_bal = fig_theme_balance(themes_df, k)
|
335 |
+
|
336 |
+
# Panneaux (rapide)
|
337 |
+
def make_panels(dfT: pd.DataFrame):
|
338 |
+
if dfT is None or dfT.empty: return "—","—","—"
|
339 |
+
pos_top = dfT.sort_values(["verbatims_pos","total_mentions"], ascending=[False,False]).head(4)
|
340 |
+
neg_top = dfT.sort_values(["verbatims_neg","total_mentions"], ascending=[False,False]).head(4)
|
341 |
+
def bullets(df, col, label):
|
342 |
+
lines=[f"**{label}**"]
|
343 |
+
for _, r in df.iterrows(): lines.append(f"- **{r['theme']}** — {int(r[col])} verbatims")
|
344 |
+
return "\n".join(lines)
|
345 |
+
ench_md = bullets(pos_top, "verbatims_pos", "Points d’enchantement")
|
346 |
+
irr_md = bullets(neg_top, "verbatims_neg", "Irritants")
|
347 |
+
RECO_RULES = {
|
348 |
+
"Délais & Suivi dossier": "Réduire les délais (SLA), suivi proactif.",
|
349 |
+
"Cotisations / Facturation": "Clarifier factures, alerter anomalies.",
|
350 |
+
"Espace client / App / Connexion": "Corriger login/MDP, QA navigateurs.",
|
351 |
+
"Support / Conseiller": "Améliorer joignabilité, scripts, rappel auto.",
|
352 |
+
"Communication / Transparence": "Notifications étapes clés, messages clairs.",
|
353 |
+
"Sinistres / Indemnisation": "Transparence délais + suivi dossier.",
|
354 |
+
}
|
355 |
+
rec_lines=["**Recommandations**"]
|
356 |
+
for _, r in neg_top.iterrows():
|
357 |
+
rec_lines.append(f"- **{r['theme']}** — {RECO_RULES.get(r['theme'],'Plan d’action dédié')}")
|
358 |
+
return ench_md, irr_md, "\n".join(rec_lines)
|
359 |
+
|
360 |
+
ench_md, irr_md, reco_md = make_panels(themes_df)
|
361 |
+
|
362 |
+
return (summary_md, themes_df.head(100), out_df.head(200), [enriched, themes, summ, zip_path],
|
363 |
+
ench_md, irr_md, reco_md, fig_gauge, fig_emots, fig_top, fig_bal)
|
364 |
+
|
365 |
+
# ---------------- UI ----------------
|
366 |
+
with gr.Blocks(title="NPS — Analyse (Assurance)") as demo:
|
367 |
+
gr.Markdown("## 🔎 NPS — Analyse sémantique (Assurance)\nColle tes verbatims (1 par ligne). Option: score NPS après un `|`.")
|
368 |
+
|
369 |
+
with gr.Column():
|
370 |
+
pasted = gr.Textbox(label="Verbatims (un par ligne)", lines=10,
|
371 |
+
placeholder="Exemple :\nRemboursement rapide, télétransmission OK | 10\nConnexion impossible à l’app | 3\nDélais corrects | 7")
|
372 |
+
with gr.Row():
|
373 |
+
has_score = gr.Checkbox(label="J’ai un score NPS par ligne", value=False)
|
374 |
+
sep = gr.Textbox(label="Séparateur score", value="|", scale=1)
|
375 |
+
|
376 |
+
with gr.Row():
|
377 |
+
anon=gr.Checkbox(label="Anonymiser emails / téléphones", value=True)
|
378 |
+
use_oa_sent=gr.Checkbox(label="OpenAI pour le sentiment", value=False)
|
379 |
+
use_oa_themes=gr.Checkbox(label="OpenAI pour les thèmes", value=False)
|
380 |
+
use_oa_summary=gr.Checkbox(label="OpenAI pour la synthèse", value=False)
|
381 |
+
|
382 |
+
with gr.Row():
|
383 |
+
oa_model=gr.Textbox(label="Modèle OpenAI", value="gpt-4o-mini")
|
384 |
+
oa_temp=gr.Slider(label="Température", minimum=0.0, maximum=1.0, value=0.1, step=0.1)
|
385 |
+
top_k=gr.Slider(label="Top thèmes (K) pour les graphes", minimum=5, maximum=20, value=10, step=1)
|
386 |
+
run=gr.Button("Lancer l'analyse", variant="primary")
|
387 |
+
|
388 |
+
with gr.Row():
|
389 |
+
ench_panel=gr.Markdown(); irr_panel=gr.Markdown(); reco_panel=gr.Markdown()
|
390 |
+
|
391 |
+
summary=gr.Markdown(label="Synthèse")
|
392 |
+
themes_table=gr.Dataframe(label="Thèmes — statistiques")
|
393 |
+
enriched_table=gr.Dataframe(label="Verbatims enrichis (aperçu)")
|
394 |
+
files_out=gr.Files(label="Téléchargements (CSV & ZIP)")
|
395 |
+
|
396 |
+
with gr.Row():
|
397 |
+
plot_nps = gr.Plot(label="NPS — Jauge")
|
398 |
+
plot_sent= gr.Plot(label="Répartition des émotions")
|
399 |
+
with gr.Row():
|
400 |
+
plot_top = gr.Plot(label="Top thèmes — occurrences")
|
401 |
+
plot_bal = gr.Plot(label="Top thèmes — balance Pos/Neg")
|
402 |
+
|
403 |
+
run.click(
|
404 |
+
analyze_text,
|
405 |
+
inputs=[pasted, has_score, sep, anon, use_oa_sent, use_oa_themes, use_oa_summary, oa_model, oa_temp, top_k],
|
406 |
+
outputs=[summary, themes_table, enriched_table, files_out, ench_panel, irr_panel, reco_panel, plot_nps, plot_sent, plot_top, plot_bal]
|
407 |
+
)
|
408 |
+
|
409 |
+
if __name__=="__main__":
|
410 |
+
demo.launch(share=False, show_api=False)
|