Documentation Index
Fetch the complete documentation index at: https://mintlify.com/diazdavilajesus16-stack/IA-LUMINA/llms.txt
Use this file to discover all available pages before exploring further.
When a student sends a message, chatbot.py’s predict_class() function runs a cascade of matching strategies to ensure there is always a relevant response, even when the neural network is not confident.
Three-stage cascade
def predict_class(sentence: str) -> list[dict]:
if not sentence.strip():
return []
bow = bag_of_words(sentence)
raw = model.predict(np.array([bow]), verbose=0)[0]
# Stage 1: high-confidence neural network prediction
high_results = [
{'intent': classes[i], 'probability': str(r)}
for i, r in enumerate(raw)
if r > _THRESHOLD_HIGH
]
high_results.sort(key=lambda x: float(x['probability']), reverse=True)
if high_results:
_context.append(high_results[0]['intent'])
return high_results
# Stage 2: keyword fallback
kw_tag = _keyword_fallback(sentence)
if kw_tag:
_context.append(kw_tag)
return [{'intent': kw_tag, 'probability': '0.50'}]
# Stage 3: low-confidence neural network
low_results = [
{'intent': classes[i], 'probability': str(r)}
for i, r in enumerate(raw)
if r > _THRESHOLD_LOW
]
low_results.sort(key=lambda x: float(x['probability']), reverse=True)
if low_results:
_context.append(low_results[0]['intent'])
return low_results
return []
| Stage | Method | Threshold | Fallback trigger |
|---|
| 1 | Neural network | > 0.60 | Score too low |
| 2 | Keyword map | Fixed 0.50 | No keyword match |
| 3 | Neural network | > 0.35 | Score too low |
| 4 | Contextual fallback | — | Always |
Keyword map
_keyword_fallback() normalizes the input (removing accents for Spanish text) then scans for keywords:
_KEYWORD_MAP: dict[str, list[str]] = {
"saludo": ["hola", "buenas", "hey", "saludos", ...],
"matematicas_general":["matematica", "matematicas", "numeros", ...],
"fisica": ["fisica", "newton", "fuerza", "velocidad", ...],
# ... 26 total categories
}
Context tracking
The last 3 matched intents are stored in a deque:
_context: deque[str] = deque(maxlen=3)
_build_fallback() uses the most recent context to offer a relevant follow-up:
_CONTEXT_FOLLOWUP = {
"matematicas_general": "Si quieres, podemos continuar con matemáticas...",
"biologia": "Si tienes más preguntas sobre biología...",
"fisica": "¿Quieres que sigamos con física?...",
"historia": "¿Seguimos explorando la historia?...",
"tecnicas_estudio": "¿Quieres que diseñemos juntos un plan de estudio?...",
}
Main.py also contains a simpler match_intent() function that uses word-overlap scoring directly against respuestas.json patterns — this is the UI’s lightweight path. chatbot.py’s predict_class() is the full ML engine, designed for deeper integration.