Update app.py
Browse files
app.py
CHANGED
|
@@ -5,7 +5,6 @@ from typing import Dict, Any, List, Tuple
|
|
| 5 |
from transformers import AutoTokenizer, AutoModelForQuestionAnswering, AutoModelForCausalLM, pipeline
|
| 6 |
|
| 7 |
# ================= CONFIGURAZIONE =================
|
| 8 |
-
|
| 9 |
MODEL_DEBERTA = "osiria/deberta-italian-question-answering"
|
| 10 |
MODEL_GEPPETTO = "LorenzoDeMattei/GePpeTto"
|
| 11 |
|
|
@@ -45,50 +44,25 @@ DOMANDE_IT: List[Tuple[str, str]] = [
|
|
| 45 |
("causale", "Qual è la causale della fattura? / Qual è la motivazione o descrizione del pagamento?")
|
| 46 |
]
|
| 47 |
|
| 48 |
-
# =================
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
if "deb" in LOADED: return LOADED["deb"]
|
| 53 |
-
tok = AutoTokenizer.from_pretrained(MODEL_DEBERTA)
|
| 54 |
-
mdl = AutoModelForQuestionAnswering.from_pretrained(MODEL_DEBERTA)
|
| 55 |
-
qa = pipeline("question-answering", model=mdl, tokenizer=tok, handle_impossible_answer=True, top_k=1, device=-1)
|
| 56 |
-
LOADED["deb"] = qa
|
| 57 |
-
return qa
|
| 58 |
-
|
| 59 |
-
def get_geppetto_pipeline():
|
| 60 |
-
if "gepp" in LOADED: return LOADED["gepp"]
|
| 61 |
-
tok = AutoTokenizer.from_pretrained(MODEL_GEPPETTO)
|
| 62 |
-
mdl = AutoModelForCausalLM.from_pretrained(MODEL_GEPPETTO)
|
| 63 |
-
gen = pipeline("text-generation", model=mdl, tokenizer=tok, device=-1)
|
| 64 |
-
LOADED["gepp"] = gen
|
| 65 |
-
return gen
|
| 66 |
|
| 67 |
-
|
|
|
|
|
|
|
| 68 |
|
|
|
|
| 69 |
def preprocess_markdown(text: str) -> str:
|
| 70 |
if not text: return ""
|
| 71 |
-
text = re.sub(r'\|[\s-]+\|', ' ', text)
|
| 72 |
text = text.replace('|', ' ')
|
| 73 |
text = text.replace('**', '').replace('##', '')
|
| 74 |
-
# mapping semantico leggero
|
| 75 |
-
text = text.replace('P.IVA', 'partita IVA').replace('PIVA', 'partita IVA')
|
| 76 |
text = re.sub(r'\s+', ' ', text).strip()
|
| 77 |
return text
|
| 78 |
|
| 79 |
-
|
| 80 |
-
if len(text) <= max_chars: return [text]
|
| 81 |
-
chunks = []
|
| 82 |
-
i = 0
|
| 83 |
-
while i < len(text):
|
| 84 |
-
end = min(i + max_chars, len(text))
|
| 85 |
-
chunks.append(text[i:end])
|
| 86 |
-
i = end - overlap
|
| 87 |
-
if i < 0: i = 0
|
| 88 |
-
return chunks
|
| 89 |
-
|
| 90 |
-
# ================= LOGICA PRINCIPALE =================
|
| 91 |
-
|
| 92 |
def analyze_invoice(md_text: str, custom_question_it: str):
|
| 93 |
logs: List[str] = []
|
| 94 |
final_output: Dict[str, Any] = {}
|
|
@@ -97,86 +71,72 @@ def analyze_invoice(md_text: str, custom_question_it: str):
|
|
| 97 |
return {"Error": "Testo troppo breve"}, "⚠️ Inserisci almeno 10 caratteri."
|
| 98 |
|
| 99 |
clean_text = preprocess_markdown(md_text)
|
| 100 |
-
chunks = chunk_text(clean_text, max_chars=3000, overlap=200)
|
| 101 |
-
logs.append(f"📄 Testo originale: {len(md_text)} chars | Pulito: {len(clean_text)} chars | Chunks: {len(chunks)}")
|
| 102 |
-
|
| 103 |
-
qa_deb = get_deberta_pipeline()
|
| 104 |
-
gen_gepp = get_geppetto_pipeline()
|
| 105 |
|
| 106 |
# 1) DeBERTa: QA estrattivo su tutte le domande + opzionale
|
| 107 |
t_start_deb = time.time()
|
| 108 |
deb_res: Dict[str, Any] = {}
|
| 109 |
success_count = 0
|
| 110 |
|
| 111 |
-
def ask_all_chunks(question: str) -> Tuple[str, float]:
|
| 112 |
-
best_answer, best_score = "", 0.0
|
| 113 |
-
for c in chunks:
|
| 114 |
-
try:
|
| 115 |
-
r = qa_deb(question=question, context=c)
|
| 116 |
-
ans = r.get("answer", "").strip()
|
| 117 |
-
score = float(r.get("score", 0.0))
|
| 118 |
-
if score > best_score and ans:
|
| 119 |
-
best_answer, best_score = ans, score
|
| 120 |
-
except Exception as e:
|
| 121 |
-
logs.append(f"❌ Errore QA chunk: {str(e)}")
|
| 122 |
-
return best_answer, best_score
|
| 123 |
-
|
| 124 |
for key, question_text in DOMANDE_IT:
|
| 125 |
-
|
| 126 |
-
|
| 127 |
-
|
| 128 |
-
|
| 129 |
-
"
|
| 130 |
-
"
|
| 131 |
-
|
| 132 |
-
|
| 133 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 134 |
|
| 135 |
custom_q = custom_question_it.strip()
|
| 136 |
if custom_q:
|
| 137 |
-
|
| 138 |
-
|
| 139 |
-
|
| 140 |
-
|
| 141 |
-
"
|
| 142 |
-
"
|
| 143 |
-
"
|
| 144 |
-
|
| 145 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 146 |
|
| 147 |
t_elapsed_deb = round(time.time() - t_start_deb, 2)
|
| 148 |
final_output["DeBERTa (estrattivo)"] = deb_res
|
| 149 |
logs.append(f"✅ DeBERTa completato in {t_elapsed_deb}s | Successi: {success_count}/{len(DOMANDE_IT) + (1 if custom_q else 0)}")
|
| 150 |
|
| 151 |
-
# 2) GePpeTto:
|
| 152 |
-
|
| 153 |
-
|
| 154 |
-
|
| 155 |
-
|
| 156 |
-
|
| 157 |
-
|
| 158 |
-
|
| 159 |
-
|
| 160 |
-
|
| 161 |
-
|
| 162 |
-
|
| 163 |
-
|
| 164 |
-
|
| 165 |
-
|
| 166 |
-
|
| 167 |
-
final_output["GePpeTto (generativo)"] = {"risposte": generative_text}
|
| 168 |
-
t_elapsed_gepp = round(time.time() - t_start_gepp, 2)
|
| 169 |
-
logs.append(f"✅ GePpeTto completato in {t_elapsed_gepp}s")
|
| 170 |
-
except Exception as e:
|
| 171 |
-
final_output["GePpeTto (generativo)"] = {"errore": str(e)}
|
| 172 |
-
logs.append(f"❌ Errore GePpeTto: {e}")
|
| 173 |
|
| 174 |
return final_output, "\n".join(logs)
|
| 175 |
|
| 176 |
# ================== UI GRADIO ==================
|
| 177 |
with gr.Blocks(theme=gr.themes.Base()) as demo:
|
| 178 |
-
gr.Markdown("# 🧾 Invoice QA: Domande standard + opzionale
|
| 179 |
-
gr.Markdown("Risposte estrattive
|
| 180 |
|
| 181 |
with gr.Row():
|
| 182 |
with gr.Column(scale=1):
|
|
@@ -194,11 +154,11 @@ with gr.Blocks(theme=gr.themes.Base()) as demo:
|
|
| 194 |
btn = gr.Button("🔍 Analizza documento", variant="primary")
|
| 195 |
|
| 196 |
with gr.Column(scale=1):
|
| 197 |
-
out_json = gr.JSON(label="Risultati
|
| 198 |
with gr.Accordion("📝 Log di Sistema (Tempi e Debug)", open=False):
|
| 199 |
out_log = gr.Textbox(label="Process Log", lines=12)
|
| 200 |
|
| 201 |
btn.click(fn=analyze_invoice, inputs=[md_input, custom_q_input], outputs=[out_json, out_log])
|
| 202 |
|
| 203 |
if __name__ == "__main__":
|
| 204 |
-
demo.launch()
|
|
|
|
| 5 |
from transformers import AutoTokenizer, AutoModelForQuestionAnswering, AutoModelForCausalLM, pipeline
|
| 6 |
|
| 7 |
# ================= CONFIGURAZIONE =================
|
|
|
|
| 8 |
MODEL_DEBERTA = "osiria/deberta-italian-question-answering"
|
| 9 |
MODEL_GEPPETTO = "LorenzoDeMattei/GePpeTto"
|
| 10 |
|
|
|
|
| 44 |
("causale", "Qual è la causale della fattura? / Qual è la motivazione o descrizione del pagamento?")
|
| 45 |
]
|
| 46 |
|
| 47 |
+
# ================= PIPELINES =================
|
| 48 |
+
tok_deb = AutoTokenizer.from_pretrained(MODEL_DEBERTA)
|
| 49 |
+
mdl_deb = AutoModelForQuestionAnswering.from_pretrained(MODEL_DEBERTA)
|
| 50 |
+
qa_deb = pipeline("question-answering", model=mdl_deb, tokenizer=tok_deb, device=-1)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 51 |
|
| 52 |
+
tok_gepp = AutoTokenizer.from_pretrained(MODEL_GEPPETTO)
|
| 53 |
+
mdl_gepp = AutoModelForCausalLM.from_pretrained(MODEL_GEPPETTO)
|
| 54 |
+
qa_gepp = pipeline("text-generation", model=mdl_gepp, tokenizer=tok_gepp, device=-1)
|
| 55 |
|
| 56 |
+
# ================= UTILITY =================
|
| 57 |
def preprocess_markdown(text: str) -> str:
|
| 58 |
if not text: return ""
|
| 59 |
+
text = re.sub(r'\|[\s-]+\|', ' ', text)
|
| 60 |
text = text.replace('|', ' ')
|
| 61 |
text = text.replace('**', '').replace('##', '')
|
|
|
|
|
|
|
| 62 |
text = re.sub(r'\s+', ' ', text).strip()
|
| 63 |
return text
|
| 64 |
|
| 65 |
+
# ================= FUNZIONE PRINCIPALE =================
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 66 |
def analyze_invoice(md_text: str, custom_question_it: str):
|
| 67 |
logs: List[str] = []
|
| 68 |
final_output: Dict[str, Any] = {}
|
|
|
|
| 71 |
return {"Error": "Testo troppo breve"}, "⚠️ Inserisci almeno 10 caratteri."
|
| 72 |
|
| 73 |
clean_text = preprocess_markdown(md_text)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 74 |
|
| 75 |
# 1) DeBERTa: QA estrattivo su tutte le domande + opzionale
|
| 76 |
t_start_deb = time.time()
|
| 77 |
deb_res: Dict[str, Any] = {}
|
| 78 |
success_count = 0
|
| 79 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 80 |
for key, question_text in DOMANDE_IT:
|
| 81 |
+
try:
|
| 82 |
+
res = qa_deb(question=question_text, context=clean_text)
|
| 83 |
+
answer = res["answer"].strip()
|
| 84 |
+
score = round(res.get("score", 0.0), 3)
|
| 85 |
+
status = "Successo" if score > 0.05 and answer else "Non Trovato"
|
| 86 |
+
if status == "Successo": success_count += 1
|
| 87 |
+
deb_res[key] = {
|
| 88 |
+
"domanda": question_text,
|
| 89 |
+
"risposta": answer,
|
| 90 |
+
"confidenza": score,
|
| 91 |
+
"status": status
|
| 92 |
+
}
|
| 93 |
+
except Exception as e:
|
| 94 |
+
deb_res[key] = {"status": f"Errore inferenza: {str(e)}"}
|
| 95 |
|
| 96 |
custom_q = custom_question_it.strip()
|
| 97 |
if custom_q:
|
| 98 |
+
try:
|
| 99 |
+
res = qa_deb(question=custom_q, context=clean_text)
|
| 100 |
+
answer = res["answer"].strip()
|
| 101 |
+
score = round(res.get("score", 0.0), 3)
|
| 102 |
+
status = "Successo" if score > 0.05 and answer else "Non Trovato"
|
| 103 |
+
if status == "Successo": success_count += 1
|
| 104 |
+
deb_res["domanda_opzionale"] = {
|
| 105 |
+
"domanda": custom_q,
|
| 106 |
+
"risposta": answer,
|
| 107 |
+
"confidenza": score,
|
| 108 |
+
"status": status
|
| 109 |
+
}
|
| 110 |
+
except Exception as e:
|
| 111 |
+
deb_res["domanda_opzionale"] = {"status": f"Errore inferenza: {str(e)}"}
|
| 112 |
|
| 113 |
t_elapsed_deb = round(time.time() - t_start_deb, 2)
|
| 114 |
final_output["DeBERTa (estrattivo)"] = deb_res
|
| 115 |
logs.append(f"✅ DeBERTa completato in {t_elapsed_deb}s | Successi: {success_count}/{len(DOMANDE_IT) + (1 if custom_q else 0)}")
|
| 116 |
|
| 117 |
+
# 2) GePpeTto: SOLO domanda opzionale
|
| 118 |
+
if custom_q:
|
| 119 |
+
try:
|
| 120 |
+
t_start_gepp = time.time()
|
| 121 |
+
short_context = clean_text[:800] # taglio prudenziale
|
| 122 |
+
prompt = f"Domanda: {custom_q}\nContesto: {short_context}\nRisposta:"
|
| 123 |
+
res_gepp = qa_gepp(prompt, max_new_tokens=64, do_sample=False)
|
| 124 |
+
generative_text = res_gepp[0]["generated_text"].replace(prompt, "").strip()
|
| 125 |
+
final_output["GePpeTto (generativo)"] = {"risposta_opzionale": generative_text}
|
| 126 |
+
t_elapsed_gepp = round(time.time() - t_start_gepp, 2)
|
| 127 |
+
logs.append(f"✅ GePpeTto completato in {t_elapsed_gepp}s")
|
| 128 |
+
except Exception as e:
|
| 129 |
+
final_output["GePpeTto (generativo)"] = {"errore": str(e)}
|
| 130 |
+
logs.append(f"❌ Errore GePpeTto: {e}")
|
| 131 |
+
else:
|
| 132 |
+
final_output["GePpeTto (generativo)"] = {"info": "Nessuna domanda opzionale fornita"}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 133 |
|
| 134 |
return final_output, "\n".join(logs)
|
| 135 |
|
| 136 |
# ================== UI GRADIO ==================
|
| 137 |
with gr.Blocks(theme=gr.themes.Base()) as demo:
|
| 138 |
+
gr.Markdown("# 🧾 Invoice QA: Domande standard + opzionale")
|
| 139 |
+
gr.Markdown("Risposte estrattive (DeBERTa) su tutte le domande e generative (GePpeTto) solo sulla domanda opzionale.")
|
| 140 |
|
| 141 |
with gr.Row():
|
| 142 |
with gr.Column(scale=1):
|
|
|
|
| 154 |
btn = gr.Button("🔍 Analizza documento", variant="primary")
|
| 155 |
|
| 156 |
with gr.Column(scale=1):
|
| 157 |
+
out_json = gr.JSON(label="Risultati (DeBERTa estrattivo + GePpeTto opzionale)")
|
| 158 |
with gr.Accordion("📝 Log di Sistema (Tempi e Debug)", open=False):
|
| 159 |
out_log = gr.Textbox(label="Process Log", lines=12)
|
| 160 |
|
| 161 |
btn.click(fn=analyze_invoice, inputs=[md_input, custom_q_input], outputs=[out_json, out_log])
|
| 162 |
|
| 163 |
if __name__ == "__main__":
|
| 164 |
+
demo.launch()
|