Update app.py
Browse files
app.py
CHANGED
|
@@ -1,36 +1,36 @@
|
|
| 1 |
import json
|
| 2 |
-
import ollama
|
| 3 |
import os
|
| 4 |
import re
|
| 5 |
import gradio as gr
|
| 6 |
-
from langchain_community.document_loaders import PyMuPDFLoader
|
| 7 |
import unicodedata
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 8 |
|
| 9 |
# Constantes
|
| 10 |
MAX_INPUT_SIZE = 4000
|
| 11 |
OVERLAP = 128
|
| 12 |
-
TEMPLATE_PATH =
|
| 13 |
-
MODEL_NAME = "nuextract" # Certifique-se de que o modelo está carregado no Ollama
|
| 14 |
|
| 15 |
-
# Função para carregar o template JSON
|
| 16 |
def load_template():
|
| 17 |
with open(TEMPLATE_PATH, "r", encoding="utf-8") as file:
|
| 18 |
return json.load(file)
|
| 19 |
|
| 20 |
-
# Função para extrair texto puro do PDF
|
| 21 |
def extract_text_from_pdf(pdf_file_path):
|
| 22 |
if not os.path.exists(pdf_file_path):
|
| 23 |
raise FileNotFoundError(f"Arquivo PDF não encontrado: {pdf_file_path}")
|
| 24 |
-
|
| 25 |
loader = PyMuPDFLoader(pdf_file_path)
|
| 26 |
data = loader.load()
|
| 27 |
return "\n".join([doc.page_content.strip() for doc in data])
|
| 28 |
|
| 29 |
-
# Função para dividir o documento em chunks de forma otimizada
|
| 30 |
def split_document(document, window_size=MAX_INPUT_SIZE, overlap=OVERLAP):
|
| 31 |
words = document.split()
|
| 32 |
chunks = []
|
| 33 |
-
|
| 34 |
if len(words) > window_size:
|
| 35 |
for i in range(0, len(words), window_size - overlap):
|
| 36 |
chunk = " ".join(words[i:i + window_size])
|
|
@@ -39,135 +39,65 @@ def split_document(document, window_size=MAX_INPUT_SIZE, overlap=OVERLAP):
|
|
| 39 |
break
|
| 40 |
else:
|
| 41 |
chunks.append(document)
|
| 42 |
-
|
| 43 |
return chunks
|
| 44 |
|
| 45 |
-
# Função para estruturar qualquer texto extraído
|
| 46 |
def structure_text(text):
|
| 47 |
lines = text.split("\n")
|
| 48 |
-
structured_data = {
|
| 49 |
-
"titulo": "",
|
| 50 |
-
"autor": "",
|
| 51 |
-
"conteudo": []
|
| 52 |
-
}
|
| 53 |
-
|
| 54 |
current_section = None
|
| 55 |
-
|
| 56 |
for line in lines:
|
| 57 |
line = line.strip()
|
| 58 |
-
|
| 59 |
-
# Detecta títulos prováveis (primeiras linhas, letras maiúsculas, tamanho grande)
|
| 60 |
if not structured_data["titulo"] and len(line) > 5 and line.istitle():
|
| 61 |
structured_data["titulo"] = line
|
| 62 |
continue
|
| 63 |
-
|
| 64 |
-
# Detecta possíveis autores (normalmente perto do título, contém nome e sobrenome)
|
| 65 |
if not structured_data["autor"] and re.search(r"\b[A-Z][a-z]+ [A-Z][a-z]+", line):
|
| 66 |
structured_data["autor"] = line
|
| 67 |
continue
|
| 68 |
-
|
| 69 |
-
# Detecta cabeçalhos de seção (geralmente curtos, sem pontuação, em negrito ou maiúsculas)
|
| 70 |
if len(line) < 60 and line.isupper():
|
| 71 |
-
current_section = {
|
| 72 |
-
"secao": line,
|
| 73 |
-
"conteudo": []
|
| 74 |
-
}
|
| 75 |
structured_data["conteudo"].append(current_section)
|
| 76 |
continue
|
| 77 |
-
|
| 78 |
-
# Adiciona conteúdo dentro da seção correta
|
| 79 |
if current_section:
|
| 80 |
current_section["conteudo"].append(line)
|
| 81 |
-
|
| 82 |
return structured_data
|
| 83 |
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
return
|
| 89 |
-
|
| 90 |
-
def normalize_text(text):
|
| 91 |
-
"""Normaliza caracteres acentuados malformados no texto."""
|
| 92 |
-
return unicodedata.normalize("NFC", text)
|
| 93 |
-
|
| 94 |
-
def fix_json(output_text):
|
| 95 |
-
"""Corrige erros no JSON gerado pelo modelo."""
|
| 96 |
-
# Remover qualquer coisa após <|end-output|>
|
| 97 |
-
output_text = re.split(r"<\|end-output\|>", output_text)[0].strip()
|
| 98 |
-
|
| 99 |
-
# Normalizar caracteres acentuados
|
| 100 |
-
output_text = normalize_text(output_text)
|
| 101 |
|
|
|
|
|
|
|
|
|
|
| 102 |
try:
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
return json.dumps(parsed_json, indent=2, ensure_ascii=False)
|
| 106 |
-
except json.JSONDecodeError as e:
|
| 107 |
-
print("⚠️ Erro ao corrigir JSON:", e)
|
| 108 |
return output_text
|
| 109 |
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
input_llm = (
|
| 115 |
-
f"<|input|>\n### Template:\n{template}\n### Current:\n{current}\n### Text:\n{text}\n\n<|output|>" + "{"
|
| 116 |
-
)
|
| 117 |
-
|
| 118 |
-
response = ollama.chat(
|
| 119 |
-
model=MODEL_NAME,
|
| 120 |
-
messages=[{"role": "user", "content": input_llm}],
|
| 121 |
-
options={"num_ctx": 4000}
|
| 122 |
-
)
|
| 123 |
-
|
| 124 |
-
output_text = response["message"]["content"]
|
| 125 |
-
|
| 126 |
-
# 🔍 Debug: Mostrar saída bruta do modelo
|
| 127 |
-
print("======= RAW OUTPUT FROM OLLAMA =======")
|
| 128 |
-
print(output_text)
|
| 129 |
-
print("======================================")
|
| 130 |
-
|
| 131 |
-
# Remove o marcador "<|end-output|>" se ele existir
|
| 132 |
-
output_text_cleaned = output_text.replace("<|end-output|>", "").strip()
|
| 133 |
-
|
| 134 |
-
try:
|
| 135 |
-
# Tenta carregar diretamente como JSON
|
| 136 |
-
return json.dumps(json.loads(output_text_cleaned), indent=2, ensure_ascii=False)
|
| 137 |
-
except json.JSONDecodeError:
|
| 138 |
-
print("⚠️ WARNING: Invalid JSON output. Returning raw text.")
|
| 139 |
-
return clean_json_text(output_text_cleaned)
|
| 140 |
|
| 141 |
-
# Função principal para processar PDF e enviar ao modelo
|
| 142 |
def process_and_generate(pdf_file):
|
| 143 |
pdf_path = pdf_file.name
|
| 144 |
extracted_text = extract_text_from_pdf(pdf_path)
|
| 145 |
-
|
| 146 |
if not extracted_text:
|
| 147 |
return "Falha ao extrair texto do PDF."
|
| 148 |
-
|
| 149 |
-
# Aplica a estruturação antes de enviar para o modelo
|
| 150 |
structured_data = structure_text(extracted_text)
|
| 151 |
-
|
| 152 |
-
|
| 153 |
-
template = json.dumps(load_template(), ensure_ascii=False) # Template inicial
|
| 154 |
-
current = json.dumps(structured_data, ensure_ascii=False) # Estado inicial estruturado
|
| 155 |
-
|
| 156 |
chunks = split_document(extracted_text)
|
| 157 |
-
|
| 158 |
-
|
| 159 |
-
print(f"Processando chunk {i}...")
|
| 160 |
-
current = predict_chunk(chunk, template, current)
|
| 161 |
-
|
| 162 |
return json.dumps(json.loads(current), indent=2, ensure_ascii=False)
|
| 163 |
|
| 164 |
-
# Interface Gradio
|
| 165 |
interface = gr.Interface(
|
| 166 |
fn=process_and_generate,
|
| 167 |
inputs=gr.File(label="Upload PDF"),
|
| 168 |
outputs="text",
|
| 169 |
-
title="Extração de Dados com
|
| 170 |
-
description="Extrai texto de PDFs e
|
| 171 |
)
|
| 172 |
-
|
| 173 |
interface.launch()
|
|
|
|
| 1 |
import json
|
|
|
|
| 2 |
import os
|
| 3 |
import re
|
| 4 |
import gradio as gr
|
|
|
|
| 5 |
import unicodedata
|
| 6 |
+
import torch
|
| 7 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 8 |
+
from langchain_community.document_loaders import PyMuPDFLoader
|
| 9 |
+
|
| 10 |
+
# Configurações do modelo
|
| 11 |
+
MODEL_PATH = "numind/NuExtract-v1.5" # Substitua pelo caminho do modelo local
|
| 12 |
+
tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)
|
| 13 |
+
model = AutoModelForCausalLM.from_pretrained(MODEL_PATH, torch_dtype=torch.float32)
|
| 14 |
|
| 15 |
# Constantes
|
| 16 |
MAX_INPUT_SIZE = 4000
|
| 17 |
OVERLAP = 128
|
| 18 |
+
TEMPLATE_PATH = # colocar template
|
|
|
|
| 19 |
|
|
|
|
| 20 |
def load_template():
|
| 21 |
with open(TEMPLATE_PATH, "r", encoding="utf-8") as file:
|
| 22 |
return json.load(file)
|
| 23 |
|
|
|
|
| 24 |
def extract_text_from_pdf(pdf_file_path):
|
| 25 |
if not os.path.exists(pdf_file_path):
|
| 26 |
raise FileNotFoundError(f"Arquivo PDF não encontrado: {pdf_file_path}")
|
|
|
|
| 27 |
loader = PyMuPDFLoader(pdf_file_path)
|
| 28 |
data = loader.load()
|
| 29 |
return "\n".join([doc.page_content.strip() for doc in data])
|
| 30 |
|
|
|
|
| 31 |
def split_document(document, window_size=MAX_INPUT_SIZE, overlap=OVERLAP):
|
| 32 |
words = document.split()
|
| 33 |
chunks = []
|
|
|
|
| 34 |
if len(words) > window_size:
|
| 35 |
for i in range(0, len(words), window_size - overlap):
|
| 36 |
chunk = " ".join(words[i:i + window_size])
|
|
|
|
| 39 |
break
|
| 40 |
else:
|
| 41 |
chunks.append(document)
|
|
|
|
| 42 |
return chunks
|
| 43 |
|
|
|
|
| 44 |
def structure_text(text):
|
| 45 |
lines = text.split("\n")
|
| 46 |
+
structured_data = {"titulo": "", "autor": "", "conteudo": []}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 47 |
current_section = None
|
|
|
|
| 48 |
for line in lines:
|
| 49 |
line = line.strip()
|
|
|
|
|
|
|
| 50 |
if not structured_data["titulo"] and len(line) > 5 and line.istitle():
|
| 51 |
structured_data["titulo"] = line
|
| 52 |
continue
|
|
|
|
|
|
|
| 53 |
if not structured_data["autor"] and re.search(r"\b[A-Z][a-z]+ [A-Z][a-z]+", line):
|
| 54 |
structured_data["autor"] = line
|
| 55 |
continue
|
|
|
|
|
|
|
| 56 |
if len(line) < 60 and line.isupper():
|
| 57 |
+
current_section = {"secao": line, "conteudo": []}
|
|
|
|
|
|
|
|
|
|
| 58 |
structured_data["conteudo"].append(current_section)
|
| 59 |
continue
|
|
|
|
|
|
|
| 60 |
if current_section:
|
| 61 |
current_section["conteudo"].append(line)
|
|
|
|
| 62 |
return structured_data
|
| 63 |
|
| 64 |
+
def generate_text(prompt):
|
| 65 |
+
inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=MAX_INPUT_SIZE)
|
| 66 |
+
with torch.no_grad():
|
| 67 |
+
output = model.generate(**inputs, max_new_tokens=512)
|
| 68 |
+
return tokenizer.decode(output[0], skip_special_tokens=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 69 |
|
| 70 |
+
def process_chunk(text, template, current):
|
| 71 |
+
input_text = f"### Template:\n{template}\n### Current:\n{current}\n### Text:\n{text}\n"
|
| 72 |
+
output_text = generate_text(input_text)
|
| 73 |
try:
|
| 74 |
+
return json.dumps(json.loads(output_text), indent=2, ensure_ascii=False)
|
| 75 |
+
except json.JSONDecodeError:
|
|
|
|
|
|
|
|
|
|
| 76 |
return output_text
|
| 77 |
|
| 78 |
+
def send_chunk_to_model(text, template, current):
|
| 79 |
+
"""Envia um chunk de texto para o modelo local e processa a saída."""
|
| 80 |
+
input_text = f"<|input|>\n### Template:\n{template}\n### Current:\n{current}\n### Text:\n{text}\n\n<|output|>"
|
| 81 |
+
return process_chunk(input_text, template, current)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 82 |
|
|
|
|
| 83 |
def process_and_generate(pdf_file):
|
| 84 |
pdf_path = pdf_file.name
|
| 85 |
extracted_text = extract_text_from_pdf(pdf_path)
|
|
|
|
| 86 |
if not extracted_text:
|
| 87 |
return "Falha ao extrair texto do PDF."
|
|
|
|
|
|
|
| 88 |
structured_data = structure_text(extracted_text)
|
| 89 |
+
template = json.dumps(load_template(), ensure_ascii=False)
|
| 90 |
+
current = json.dumps(structured_data, ensure_ascii=False)
|
|
|
|
|
|
|
|
|
|
| 91 |
chunks = split_document(extracted_text)
|
| 92 |
+
for chunk in chunks:
|
| 93 |
+
current = send_chunk_to_model(chunk, template, current)
|
|
|
|
|
|
|
|
|
|
| 94 |
return json.dumps(json.loads(current), indent=2, ensure_ascii=False)
|
| 95 |
|
|
|
|
| 96 |
interface = gr.Interface(
|
| 97 |
fn=process_and_generate,
|
| 98 |
inputs=gr.File(label="Upload PDF"),
|
| 99 |
outputs="text",
|
| 100 |
+
title="Extração de Dados com Modelo Local",
|
| 101 |
+
description="Extrai texto de PDFs e processa utilizando um modelo local de linguagem."
|
| 102 |
)
|
|
|
|
| 103 |
interface.launch()
|