Upload 10 files

app.py

@@ -11,33 +11,16 @@ from demo_text import DEMO_TEXT
 # ---------------------------------------------------------------------------
 # Handlers
 # ---------------------------------------------------------------------------
-import re as _re
-
-_ALREADY_ANON_RE = _re.compile(r"\[[A-Z_]+(?:_\d+)?\]|(\*{3,})")
-
-
 def _process(text: str, mode: str, min_score: float, use_regex: bool):
     if not text or not text.strip():
         e = '<div style="padding:40px; text-align:center; color:#6b7280; font-family:Arial,sans-serif;">Inserisci del testo per iniziare.</div>'
         return e, e, e
-
-    # Avviso se il testo sembra già offuscato
-    warning = ""
-    if _ALREADY_ANON_RE.search(text):
-        warning = (
-            '<div style="background:#fefce8; border-left:4px solid #ca8a04; '
-            'padding:10px 16px; margin-bottom:12px; border-radius:4px; '
-            'font-family:Arial,sans-serif; font-size:0.9em; color:#713f12;">'
-            '⚠️ Il testo sembra già parzialmente offuscato (contiene placeholder o asterischi). '
-            'I risultati potrebbero includere falsi positivi sulle parti già mascherate.'
-            '</div>'
-        )
-
     anon_text, entities = anonymize(text, mode, min_score, use_regex)
-    highlighted = warning + render_highlighted_text(text, entities)
-    anonymized  = render_anonymized_text(anon_text)
-    report      = warning + render_categorized_report(text, entities)
-    return highlighted, anonymized, report
+    return (
+        render_highlighted_text(text, entities),
+        render_anonymized_text(anon_text),
+        render_categorized_report(text, entities),
+    )
 
 
 def handle_text(text, mode, min_score, use_regex):
@@ -92,118 +75,6 @@ _CSS = """
     color: #6b7280 !important; margin-bottom: 6px !important;
 }
 button.primary { font-weight: 600 !important; }
-
-/* ── Loading overlay ─────────────────────────────────────────────────── */
-#anon-overlay {
-    display: none;
-    position: fixed;
-    inset: 0;
-    background: rgba(15, 23, 42, 0.72);
-    backdrop-filter: blur(4px);
-    -webkit-backdrop-filter: blur(4px);
-    z-index: 10000;
-    justify-content: center;
-    align-items: center;
-}
-#anon-overlay.anon-visible { display: flex; }
-#anon-box {
-    background: #ffffff;
-    border-radius: 16px;
-    padding: 48px 64px;
-    text-align: center;
-    box-shadow: 0 24px 64px rgba(0,0,0,0.35);
-    min-width: 300px;
-}
-#anon-ring {
-    width: 60px; height: 60px;
-    border: 5px solid #dbeafe;
-    border-top-color: #1d4ed8;
-    border-radius: 50%;
-    animation: anon-spin 0.85s linear infinite;
-    margin: 0 auto 24px;
-}
-#anon-title {
-    font-size: 1.15em; font-weight: 700;
-    color: #1e3a8a; margin: 0 0 8px;
-}
-#anon-sub {
-    font-size: 0.82em; color: #6b7280; margin: 0;
-}
-@keyframes anon-spin { to { transform: rotate(360deg); } }
-"""
-
-# ---------------------------------------------------------------------------
-# JS – overlay a schermo intero durante l'analisi
-# ---------------------------------------------------------------------------
-_JS = """
-() => {
-    /* ── Crea overlay ──────────────────────────────────────────────── */
-    const ov = document.createElement('div');
-    ov.id = 'anon-overlay';
-    ov.innerHTML = `
-        <div id="anon-box">
-            <div id="anon-ring"></div>
-            <p id="anon-title">Analisi in corso&hellip;</p>
-            <p id="anon-sub">Regex &rarr; NER &rarr; GLiNER</p>
-        </div>`;
-    document.body.appendChild(ov);
-
-    let processing = false;
-    let hideTimer  = null;
-
-    const show = () => {
-        processing = true;
-        clearTimeout(hideTimer);
-        ov.classList.add('anon-visible');
-        /* Sicurezza: auto-hide dopo 3 minuti */
-        hideTimer = setTimeout(hide, 180_000);
-    };
-
-    const hide = () => {
-        processing = false;
-        clearTimeout(hideTimer);
-        ov.classList.remove('anon-visible');
-    };
-
-    /* ── Osserva gli output HTML specifici per nascondere overlay ──── */
-    /* Quando Gradio aggiorna un gr.HTML(), cambia il suo innerHTML.    */
-    /* Usiamo elem_id per trovare i 3 output in modo affidabile.        */
-    const OUTPUT_IDS = ['out-highlighted', 'out-anonymized', 'out-report'];
-    const outputObs  = new MutationObserver(() => {
-        if (processing) {
-            clearTimeout(hideTimer);
-            /* Piccolo delay per consentire a Gradio di finire il render */
-            hideTimer = setTimeout(hide, 200);
-        }
-    });
-
-    function attachOutputObservers() {
-        OUTPUT_IDS.forEach(id => {
-            const el = document.getElementById(id);
-            if (el && !el.dataset.anonObs) {
-                el.dataset.anonObs = '1';
-                outputObs.observe(el, { childList: true, subtree: true, characterData: true });
-            }
-        });
-    }
-
-    /* ── Collega click ai bottoni "Anonimizza" ─────────────────────── */
-    function bindButtons() {
-        document.querySelectorAll('button').forEach(btn => {
-            if (btn.textContent.trim() === 'Anonimizza' && !btn.dataset.anonBound) {
-                btn.dataset.anonBound = '1';
-                btn.addEventListener('click', show);
-            }
-        });
-        attachOutputObservers();
-    }
-
-    /* MutationObserver globale solo per scoprire nuovi elementi Gradio */
-    const initObs = new MutationObserver(bindButtons);
-    initObs.observe(document.body, { childList: true, subtree: true });
-
-    [300, 1000, 2500].forEach(t => setTimeout(bindButtons, t));
-}
 """
 
 # ---------------------------------------------------------------------------
@@ -214,7 +85,6 @@ with gr.Blocks(
     theme=gr.themes.Soft(primary_hue="blue", neutral_hue="slate",
                          font=["Arial", "Helvetica", "sans-serif"]),
     css=_CSS,
-    js=_JS,
 ) as demo:
 
     with gr.Row(elem_classes=["header-block"]):
@@ -262,11 +132,11 @@ with gr.Blocks(
     gr.Markdown("### Risultato", elem_classes=["section-label"])
     with gr.Tabs():
         with gr.TabItem("🎨 Evidenziato"):
-            out_highlighted = gr.HTML(elem_id="out-highlighted")
+            out_highlighted = gr.HTML()
         with gr.TabItem("🔒 Anonimizzato"):
-            out_anonymized  = gr.HTML(elem_id="out-anonymized")
+            out_anonymized  = gr.HTML()
         with gr.TabItem("📊 Report"):
-            out_report      = gr.HTML(elem_id="out-report")
+            out_report      = gr.HTML()
 
     outputs       = [out_highlighted, out_anonymized, out_report]
     inputs_common = [mode_radio, score_slider, use_regex_chk]

gliner_recognizer.py

@@ -44,12 +44,7 @@ SUPPORTED_ENTITIES: list[str] = sorted(set(_LABEL_TO_ENTITY.values()))
 
 
 class GlinerRecognizer(EntityRecognizer):
-    """
-    Wrapper Presidio per GLiNER zero-shot.
-
-    IMPORTANTE: self._model viene inizializzato PRIMA di super().__init__()
-    perché Presidio chiama self.load() durante l'inizializzazione del parent.
-    """
+    """Wrapper Presidio per GLiNER zero-shot. Lazy loading, gestione errori robusta."""
 
     def __init__(
         self,
@@ -57,22 +52,17 @@ class GlinerRecognizer(EntityRecognizer):
         labels: Optional[list[str]] = None,
         model_name: str = "DeepMount00/GLiNER_PII_ITA",
     ):
-        # ── DEVE essere prima di super().__init__() ────────────────────────
-        self._model     = None   # None = non caricato, False = caricamento fallito
-        self.threshold  = threshold
-        self.labels     = labels or PROCUREMENT_LABELS
-        self.model_name = model_name
-        # ──────────────────────────────────────────────────────────────────
         super().__init__(
             supported_entities=SUPPORTED_ENTITIES,
             supported_language="it",
             name="GlinerRecognizer",
         )
+        self.threshold  = threshold
+        self.labels     = labels or PROCUREMENT_LABELS
+        self.model_name = model_name
+        self._model     = None  # None = non caricato, False = fallito
 
     def load(self) -> None:
-        """Carica il modello GLiNER (idempotente)."""
-        if not hasattr(self, "_model"):
-            self._model = None
         if self._model is not None:
             return
         try:
@@ -86,16 +76,12 @@ class GlinerRecognizer(EntityRecognizer):
     def analyze(self, text: str, entities: List[str], nlp_artifacts=None) -> List[RecognizerResult]:
         if not text or not text.strip():
             return []
-
-        if not hasattr(self, "_model") or self._model is None:
+        if self._model is None:
             self.load()
         if self._model is False:
             return []
-
         try:
-            predictions = self._model.predict_entities(
-                text, self.labels, threshold=self.threshold,
-            )
+            predictions = self._model.predict_entities(text, self.labels, threshold=self.threshold)
         except Exception as exc:
             logger.error("[GLiNER] Predizione fallita: %s", exc)
             return []

pipeline.py

@@ -1,4 +1,4 @@
-"""Pipeline a 3 livelli: Regex (opz.) → NER → GLiNER, chunking, agreement boost."""
+"""Pipeline a 3 livelli: Regex (opz.) → NER → GLiNER, con chunking e agreement boost."""
 import hashlib
 import logging
 import re
@@ -60,7 +60,7 @@ _gliner: Optional[GlinerRecognizer] = None
 
 
 # ---------------------------------------------------------------------------
-# Helpers – singleton, utils
+# Helpers
 # ---------------------------------------------------------------------------
 def _get_gliner(threshold: float = 0.65) -> GlinerRecognizer:
     global _gliner
@@ -88,7 +88,7 @@ def _reclassify(results: list, text: str) -> list:
 
 
 def _scale_amount(original: str, scale_pct: float = 0.20) -> str:
-    """Scala importo ±scale_pct deterministicamente (hash MD5)."""
+    """Scala importo numerico ±scale_pct deterministicamente (hash MD5)."""
     match = re.search(r"[\d\.,]+", original)
     if not match:
         return original
@@ -121,7 +121,7 @@ def _apply_mode(s: str, mode: str, label: str) -> str:
 
 
 # ---------------------------------------------------------------------------
-# Session – placeholder numerati coerenti per documento
+# Session – placeholder numerati coerenti
 # ---------------------------------------------------------------------------
 class Session:
     def __init__(self):
@@ -256,50 +256,11 @@ def _post_boost_check(entities: list, text: str) -> list:
     return entities
 
 
-# ---------------------------------------------------------------------------
-# Filtro falsi positivi da re-parsing
-# ---------------------------------------------------------------------------
-def _filter_noise(entities: list, text: str) -> list:
-    """
-    Rimuove falsi positivi prodotti da re-parsing di testo già offuscato.
-
-    Casi gestiti:
-     • Span con >50% asterischi → già mascherato, salta
-     • VALUTA/IMPORTO senza cifre → simbolo isolato (€, EUR), salta
-     • FREQUENZA con score < 0.80 → falso positivo comune, salta
-     • Span che inizia/finisce con parentesi aperta → boundary NER rotto, salta
-    """
-    clean: list = []
-    for r in entities:
-        span = text[r.start:r.end]
-
-        # Span quasi interamente asterischi (testo già mascherato)
-        if len(span) > 1 and span.count("*") / len(span) > 0.5:
-            continue
-
-        # VALUTA/IMPORTO senza nessuna cifra → simbolo isolato
-        if r.entity_type in ("VALUTA", "IMPORTO_GARA"):
-            if not re.search(r"\d", span):
-                continue
-
-        # FREQUENZA generica senza context sufficiente
-        if r.entity_type == "FREQUENZA" and r.score < 0.80:
-            continue
-
-        # Boundary NER rotto: span inizia o finisce con parentesi aperta
-        stripped = span.strip()
-        if stripped.startswith("(") or stripped.endswith("("):
-            continue
-
-        clean.append(r)
-    return clean
-
-
 # ---------------------------------------------------------------------------
 # Pipeline pubblica
 # ---------------------------------------------------------------------------
 def detect(text: str, min_score: float = 0.65, use_regex: bool = True) -> list[RecognizerResult]:
-    """Rilevazione a 3 livelli con chunking + cross-layer boost + post-boost + noise filter."""
+    """Rilevazione a 3 livelli con chunking + cross-layer boost + post-boost."""
     if not text or not text.strip():
         return []
 
@@ -334,9 +295,7 @@ def detect(text: str, min_score: float = 0.65, use_regex: bool = True) -> list[R
     )
 
     final = resolve_overlapping_spans([regex_results, ner_results, gliner_results])
-    final = _post_boost_check(final, text)
-    final = _filter_noise(final, text)
-    return final
+    return _post_boost_check(final, text)
 
 
 def anonymize_from_entities(text: str, entities: list[RecognizerResult], mode_label: str) -> str: