Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -15,7 +15,7 @@ def get_blocks_from_docx():
|
|
| 15 |
blocks = []
|
| 16 |
for p in doc.paragraphs:
|
| 17 |
txt = p.text.strip()
|
| 18 |
-
#
|
| 19 |
if (
|
| 20 |
txt
|
| 21 |
and not (len(txt) <= 3 and txt.isdigit())
|
|
@@ -24,14 +24,13 @@ def get_blocks_from_docx():
|
|
| 24 |
and txt == txt.upper()
|
| 25 |
and txt.endswith(('.', ':', '?', '!')) is False
|
| 26 |
)
|
| 27 |
-
and len(txt.split()) > 3
|
| 28 |
):
|
| 29 |
blocks.append(txt)
|
| 30 |
-
#
|
| 31 |
for table in doc.tables:
|
| 32 |
for row in table.rows:
|
| 33 |
row_text = " | ".join(cell.text.strip() for cell in row.cells if cell.text.strip())
|
| 34 |
-
# Аналогично — игнорируем сверхкороткие строки/возможные заголовки из таблиц:
|
| 35 |
if row_text and len(row_text) > 35 and len(row_text.split()) > 3:
|
| 36 |
blocks.append(row_text)
|
| 37 |
seen = set()
|
|
@@ -55,7 +54,12 @@ def rut5_answer(question, context):
|
|
| 55 |
prompt = f"question: {question} context: {context}"
|
| 56 |
input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)
|
| 57 |
with torch.no_grad():
|
| 58 |
-
output_ids = model.generate(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 59 |
return tokenizer.decode(output_ids[0], skip_special_tokens=True)
|
| 60 |
|
| 61 |
def ask_chatbot(question):
|
|
@@ -65,15 +69,15 @@ def ask_chatbot(question):
|
|
| 65 |
return "Ошибка: база знаний пуста или слишком мала. Проверьте .docx."
|
| 66 |
user_vec = vectorizer.transform([question])
|
| 67 |
sims = cosine_similarity(user_vec, matrix)
|
| 68 |
-
|
| 69 |
-
top_idxs = sims.argsort()[-
|
| 70 |
-
# Используем только НЕКОРОТКИЕ блоки как контекст
|
| 71 |
context_blocks = [
|
| 72 |
blocks[i] for i in top_idxs if sims[i] > 0.08 and len(blocks[i].split()) > 3 and len(blocks[i]) > 35
|
| 73 |
]
|
| 74 |
context = " ".join(context_blocks)
|
|
|
|
|
|
|
| 75 |
answer = rut5_answer(question, context)
|
| 76 |
-
# Подстраховка — если ответ ТОЛЬКО заголовок, просто версифицируем и дополняем контекстом:
|
| 77 |
if len(answer.strip().split()) < 8 or len(answer.split('.')) < 2:
|
| 78 |
answer += "\n\n" + context
|
| 79 |
return answer
|
|
@@ -92,17 +96,22 @@ with gr.Blocks() as demo:
|
|
| 92 |
"""
|
| 93 |
# Русскоязычный FAQ-чат-бот на базе вашей методички и нейросетевой модели
|
| 94 |
|
| 95 |
-
Задайте вопрос — получите развернутый AI-ответ
|
| 96 |
"""
|
| 97 |
)
|
| 98 |
question = gr.Textbox(label="Ваш вопрос", lines=2)
|
| 99 |
ask_btn = gr.Button("Получить ответ")
|
| 100 |
answer = gr.Markdown(label="Ответ", visible=True)
|
| 101 |
-
ask_btn.click(ask_chatbot, question, answer)
|
| 102 |
-
question.submit(ask_chatbot, question, answer)
|
| 103 |
-
gr.Markdown("#### Примеры вопросов:")
|
| 104 |
|
| 105 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 106 |
gr.Examples(EXAMPLES, inputs=question)
|
| 107 |
|
| 108 |
gr.Markdown("""
|
|
|
|
| 15 |
blocks = []
|
| 16 |
for p in doc.paragraphs:
|
| 17 |
txt = p.text.strip()
|
| 18 |
+
# Исключаем очень короткие и похожие на заголовки блоки
|
| 19 |
if (
|
| 20 |
txt
|
| 21 |
and not (len(txt) <= 3 and txt.isdigit())
|
|
|
|
| 24 |
and txt == txt.upper()
|
| 25 |
and txt.endswith(('.', ':', '?', '!')) is False
|
| 26 |
)
|
| 27 |
+
and len(txt.split()) > 3
|
| 28 |
):
|
| 29 |
blocks.append(txt)
|
| 30 |
+
# Таблицы
|
| 31 |
for table in doc.tables:
|
| 32 |
for row in table.rows:
|
| 33 |
row_text = " | ".join(cell.text.strip() for cell in row.cells if cell.text.strip())
|
|
|
|
| 34 |
if row_text and len(row_text) > 35 and len(row_text.split()) > 3:
|
| 35 |
blocks.append(row_text)
|
| 36 |
seen = set()
|
|
|
|
| 54 |
prompt = f"question: {question} context: {context}"
|
| 55 |
input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)
|
| 56 |
with torch.no_grad():
|
| 57 |
+
output_ids = model.generate(
|
| 58 |
+
input_ids,
|
| 59 |
+
max_length=250, num_beams=4, min_length=40,
|
| 60 |
+
no_repeat_ngram_size=3,
|
| 61 |
+
do_sample=False
|
| 62 |
+
)
|
| 63 |
return tokenizer.decode(output_ids[0], skip_special_tokens=True)
|
| 64 |
|
| 65 |
def ask_chatbot(question):
|
|
|
|
| 69 |
return "Ошибка: база знаний пуста или слишком мала. Проверьте .docx."
|
| 70 |
user_vec = vectorizer.transform([question])
|
| 71 |
sims = cosine_similarity(user_vec, matrix)
|
| 72 |
+
n_blocks = min(3, len(blocks))
|
| 73 |
+
top_idxs = sims.argsort()[-n_blocks:][::-1] if n_blocks > 0 else []
|
|
|
|
| 74 |
context_blocks = [
|
| 75 |
blocks[i] for i in top_idxs if sims[i] > 0.08 and len(blocks[i].split()) > 3 and len(blocks[i]) > 35
|
| 76 |
]
|
| 77 |
context = " ".join(context_blocks)
|
| 78 |
+
if not context:
|
| 79 |
+
return "Не найден релевантный фрагмент в документе. Попробуйте иначе сформулировать вопрос."
|
| 80 |
answer = rut5_answer(question, context)
|
|
|
|
| 81 |
if len(answer.strip().split()) < 8 or len(answer.split('.')) < 2:
|
| 82 |
answer += "\n\n" + context
|
| 83 |
return answer
|
|
|
|
| 96 |
"""
|
| 97 |
# Русскоязычный FAQ-чат-бот на базе вашей методички и нейросетевой модели
|
| 98 |
|
| 99 |
+
Задайте вопрос — получите развернутый AI-ответ (бот анализирует ваш документ, подборка абзацев и генерация выполняются автоматически).
|
| 100 |
"""
|
| 101 |
)
|
| 102 |
question = gr.Textbox(label="Ваш вопрос", lines=2)
|
| 103 |
ask_btn = gr.Button("Получить ответ")
|
| 104 |
answer = gr.Markdown(label="Ответ", visible=True)
|
|
|
|
|
|
|
|
|
|
| 105 |
|
| 106 |
+
# Активация спиннера "Чат-бот думает..."
|
| 107 |
+
def with_spinner(q):
|
| 108 |
+
yield "Чат-бот думает..."
|
| 109 |
+
yield ask_chatbot(q)
|
| 110 |
+
|
| 111 |
+
ask_btn.click(with_spinner, question, answer)
|
| 112 |
+
question.submit(with_spinner, question, answer)
|
| 113 |
+
|
| 114 |
+
gr.Markdown("#### Примеры вопросов:")
|
| 115 |
gr.Examples(EXAMPLES, inputs=question)
|
| 116 |
|
| 117 |
gr.Markdown("""
|