Spaces:
Running
Running
# modules/semantic/semantic_live_interface.py | |
import streamlit as st | |
from streamlit_float import * | |
from streamlit_antd_components import * | |
import pandas as pd | |
import logging | |
# Configuración del logger | |
logger = logging.getLogger(__name__) | |
# Importaciones locales | |
from .semantic_process import ( | |
process_semantic_input, | |
format_semantic_results | |
) | |
from ..utils.widget_utils import generate_unique_key | |
from ..database.semantic_mongo_db import store_student_semantic_result | |
from ..database.chat_mongo_db import store_chat_history, get_chat_history | |
def display_semantic_live_interface(lang_code, nlp_models, semantic_t): | |
""" | |
Interfaz para el análisis semántico en vivo con proporciones de columna ajustadas | |
""" | |
try: | |
# 1. Inicializar el estado de la sesión de manera más robusta | |
if 'semantic_live_state' not in st.session_state: | |
st.session_state.semantic_live_state = { | |
'analysis_count': 0, | |
'current_text': '', | |
'last_result': None, | |
'text_changed': False | |
} | |
# 2. Función para manejar cambios en el texto | |
def on_text_change(): | |
current_text = st.session_state.semantic_live_text | |
st.session_state.semantic_live_state['current_text'] = current_text | |
st.session_state.semantic_live_state['text_changed'] = True | |
# 3. Crear columnas con nueva proporción (1:3) | |
input_col, result_col = st.columns([1, 3]) | |
# Columna izquierda: Entrada de texto | |
with input_col: | |
st.subheader(semantic_t.get('enter_text', 'Ingrese su texto')) | |
# Área de texto con manejo de eventos | |
text_input = st.text_area( | |
semantic_t.get('text_input_label', 'Escriba o pegue su texto aquí'), | |
height=500, | |
key="semantic_live_text", | |
value=st.session_state.semantic_live_state.get('current_text', ''), | |
on_change=on_text_change, | |
label_visibility="collapsed" # Oculta el label para mayor estabilidad | |
) | |
# Botón de análisis y procesamiento | |
analyze_button = st.button( | |
semantic_t.get('analyze_button', 'Analizar'), | |
key="semantic_live_analyze", | |
type="primary", | |
icon="🔍", | |
disabled=not text_input, | |
use_container_width=True | |
) | |
if analyze_button and text_input: | |
try: | |
with st.spinner(semantic_t.get('processing', 'Procesando...')): | |
analysis_result = process_semantic_input( | |
text_input, | |
lang_code, | |
nlp_models, | |
semantic_t | |
) | |
if analysis_result['success']: | |
st.session_state.semantic_live_state['last_result'] = analysis_result | |
st.session_state.semantic_live_state['analysis_count'] += 1 | |
st.session_state.semantic_live_state['text_changed'] = False | |
store_student_semantic_result( | |
st.session_state.username, | |
text_input, | |
analysis_result['analysis'] | |
) | |
else: | |
st.error(analysis_result.get('message', 'Error en el análisis')) | |
except Exception as e: | |
logger.error(f"Error en análisis: {str(e)}") | |
st.error(semantic_t.get('error_processing', 'Error al procesar el texto')) | |
# Columna derecha: Visualización de resultados | |
with result_col: | |
st.subheader(semantic_t.get('live_results', 'Resultados en vivo')) | |
if 'last_result' in st.session_state.semantic_live_state and \ | |
st.session_state.semantic_live_state['last_result'] is not None: | |
analysis = st.session_state.semantic_live_state['last_result']['analysis'] | |
if 'key_concepts' in analysis and analysis['key_concepts'] and \ | |
'concept_graph' in analysis and analysis['concept_graph'] is not None: | |
st.markdown(""" | |
<style> | |
.unified-container { | |
background-color: white; | |
border-radius: 10px; | |
overflow: hidden; | |
box-shadow: 0 2px 4px rgba(0,0,0,0.1); | |
width: 100%; | |
margin-bottom: 1rem; | |
} | |
.concept-table { | |
display: flex; | |
flex-wrap: nowrap; /* Evita el wrap */ | |
gap: 6px; /* Reducido el gap */ | |
padding: 10px; | |
background-color: #f8f9fa; | |
overflow-x: auto; /* Permite scroll horizontal si es necesario */ | |
white-space: nowrap; /* Mantiene todo en una línea */ | |
} | |
.concept-item { | |
background-color: white; | |
border-radius: 4px; | |
padding: 4px 8px; /* Padding reducido */ | |
display: inline-flex; /* Cambiado a inline-flex */ | |
align-items: center; | |
gap: 4px; /* Gap reducido */ | |
box-shadow: 0 1px 2px rgba(0,0,0,0.1); | |
flex-shrink: 0; /* Evita que los items se encojan */ | |
} | |
.concept-name { | |
font-weight: 500; | |
color: #1f2937; | |
font-size: 0.8em; /* Tamaño de fuente reducido */ | |
} | |
.concept-freq { | |
color: #6b7280; | |
font-size: 0.75em; /* Tamaño de fuente reducido */ | |
} | |
.graph-section { | |
padding: 20px; | |
background-color: white; | |
} | |
</style> | |
""", unsafe_allow_html=True) | |
with st.container(): | |
# Conceptos en una sola línea | |
concepts_html = """ | |
<div class="unified-container"> | |
<div class="concept-table"> | |
""" | |
concepts_html += ''.join( | |
f'<div class="concept-item"><span class="concept-name">{concept}</span>' | |
f'<span class="concept-freq">({freq:.2f})</span></div>' | |
for concept, freq in analysis['key_concepts'] | |
) | |
concepts_html += "</div></div>" | |
st.markdown(concepts_html, unsafe_allow_html=True) | |
# Grafo | |
if 'concept_graph' in analysis and analysis['concept_graph'] is not None: | |
st.image( | |
analysis['concept_graph'], | |
use_container_width=True | |
) | |
# Botones y controles | |
button_col, spacer_col = st.columns([1,5]) | |
with button_col: | |
st.download_button( | |
label="📥 " + semantic_t.get('download_graph', "Download"), | |
data=analysis['concept_graph'], | |
file_name="semantic_live_graph.png", | |
mime="image/png", | |
use_container_width=True | |
) | |
with st.expander("📊 " + semantic_t.get('graph_help', "Graph Interpretation")): | |
st.markdown(""" | |
- 🔀 Las flechas indican la dirección de la relación entre conceptos | |
- 🎨 Los colores más intensos indican conceptos más centrales en el texto | |
- ⭕ El tamaño de los nodos representa la frecuencia del concepto | |
- ↔️ El grosor de las líneas indica la fuerza de la conexión | |
""") | |
else: | |
st.info(semantic_t.get('no_graph', 'No hay datos para mostrar')) | |
except Exception as e: | |
logger.error(f"Error general en interfaz semántica en vivo: {str(e)}") | |
st.error(semantic_t.get('general_error', "Se produjo un error. Por favor, intente de nuevo.")) | |