Update rag_hf.py
Browse files
rag_hf.py
CHANGED
@@ -8,7 +8,6 @@ import os
|
|
8 |
import requests
|
9 |
from rdflib import Graph as RDFGraph, Namespace
|
10 |
from sentence_transformers import SentenceTransformer
|
11 |
-
#from dotenv import load_dotenv
|
12 |
|
13 |
# === STREAMLIT UI CONFIG ===
|
14 |
st.set_page_config(
|
@@ -23,9 +22,7 @@ st.set_page_config(
|
|
23 |
)
|
24 |
|
25 |
# === CONFIGURATION ===
|
26 |
-
|
27 |
-
ENDPOINT_URL = "https://api-inference.huggingface.co/models/NousResearch/Nous-Hermes-2-Mistral-7B-DPO"
|
28 |
-
#ENDPOINT_URL = "https://api-inference.huggingface.co/models/mistralai/Mistral-7B-Instruct-v0.3"
|
29 |
HF_API_TOKEN = os.getenv("HF_API_TOKEN")
|
30 |
if not HF_API_TOKEN:
|
31 |
st.error("⚠️ No se cargó el token HF_API_TOKEN desde los Secrets.")
|
@@ -38,61 +35,6 @@ EX = Namespace("http://example.org/lang/")
|
|
38 |
# === CUSTOM CSS ===
|
39 |
st.markdown("""
|
40 |
<style>
|
41 |
-
.header {
|
42 |
-
color: #2c3e50;
|
43 |
-
border-bottom: 2px solid #4f46e5;
|
44 |
-
padding-bottom: 0.5rem;
|
45 |
-
margin-bottom: 1.5rem;
|
46 |
-
}
|
47 |
-
.feature-card {
|
48 |
-
background-color: #f8fafc;
|
49 |
-
border-radius: 8px;
|
50 |
-
padding: 1rem;
|
51 |
-
margin: 0.5rem 0;
|
52 |
-
border-left: 3px solid #4f46e5;
|
53 |
-
}
|
54 |
-
.response-card {
|
55 |
-
background-color: #fdfdfd;
|
56 |
-
color: #1f2937;
|
57 |
-
border-radius: 8px;
|
58 |
-
padding: 1.5rem;
|
59 |
-
box-shadow: 0 2px 6px rgba(0,0,0,0.08);
|
60 |
-
margin: 1rem 0;
|
61 |
-
font-size: 1rem;
|
62 |
-
line-height: 1.5;
|
63 |
-
}
|
64 |
-
.language-card {
|
65 |
-
background-color: #f9fafb;
|
66 |
-
border-radius: 8px;
|
67 |
-
padding: 1rem;
|
68 |
-
margin: 0.5rem 0;
|
69 |
-
border: 1px solid #e5e7eb;
|
70 |
-
}
|
71 |
-
.sidebar-section {
|
72 |
-
margin-bottom: 1.5rem;
|
73 |
-
}
|
74 |
-
.sidebar-title {
|
75 |
-
font-weight: 600;
|
76 |
-
color: #4f46e5;
|
77 |
-
}
|
78 |
-
.suggested-question {
|
79 |
-
padding: 0.5rem;
|
80 |
-
margin: 0.25rem 0;
|
81 |
-
border-radius: 4px;
|
82 |
-
cursor: pointer;
|
83 |
-
transition: all 0.2s;
|
84 |
-
}
|
85 |
-
.suggested-question:hover {
|
86 |
-
background-color: #f1f5f9;
|
87 |
-
}
|
88 |
-
.metric-badge {
|
89 |
-
display: inline-block;
|
90 |
-
background-color: #e8f4fc;
|
91 |
-
padding: 0.25rem 0.5rem;
|
92 |
-
border-radius: 4px;
|
93 |
-
font-size: 0.85rem;
|
94 |
-
margin-right: 0.5rem;
|
95 |
-
}
|
96 |
.tech-badge {
|
97 |
background-color: #ecfdf5;
|
98 |
color: #065f46;
|
@@ -109,7 +51,6 @@ st.markdown("""
|
|
109 |
def load_all_components():
|
110 |
embedder = SentenceTransformer(EMBEDDING_MODEL, device=DEVICE)
|
111 |
methods = {}
|
112 |
-
# Solo carga el método LinkGraph
|
113 |
label, suffix, ttl, matrix_path = ("LinkGraph", "_hybrid_graphsage", "grafo_ttl_hibrido_graphsage.ttl", "embed_matrix_hybrid_graphsage.npy")
|
114 |
with open(f"id_map{suffix}.pkl", "rb") as f:
|
115 |
id_map = pickle.load(f)
|
@@ -149,6 +90,24 @@ def query_rdf(rdf, lang_id):
|
|
149 |
except Exception as e:
|
150 |
return [("error", str(e))]
|
151 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
152 |
def generate_response(matrix, id_map, G, rdf, user_question, k, embedder):
|
153 |
ids = get_top_k(matrix, id_map, user_question, k, embedder)
|
154 |
context = [get_context(G, i) for i in ids]
|
@@ -156,186 +115,49 @@ def generate_response(matrix, id_map, G, rdf, user_question, k, embedder):
|
|
156 |
for i in ids:
|
157 |
rdf_facts.extend([f"{p}: {v}" for p, v in query_rdf(rdf, i)])
|
158 |
|
159 |
-
|
160 |
-
|
161 |
-
|
162 |
-
|
163 |
-
|
164 |
-
###
|
165 |
-
|
166 |
-
### PREGUNTA: {user_question}
|
167 |
-
Respuesta: [/INST]"""
|
168 |
-
|
169 |
-
# Prompt para generar respuesta en inglés
|
170 |
-
prompt_en = f"""<s>[INST] You are an expert in South American indigenous languages. Use strictly and only the information below to answer the user question in **English**.
|
171 |
-
- Do not infer or assume facts that are not explicitly stated.
|
172 |
-
- If the answer is unknown or insufficient, say \"I cannot answer with the available data.\"
|
173 |
-
- Limit your answer to 100 words.
|
174 |
-
### CONTEXT: {chr(10).join(context)}
|
175 |
-
### RDF RELATIONS: {chr(10).join(rdf_facts)}
|
176 |
-
### QUESTION: {user_question}
|
177 |
-
Answer: [/INST]"""
|
178 |
-
|
179 |
-
response_es = "Error al generar respuesta en español."
|
180 |
-
response_en = "Error generating response in English."
|
181 |
-
|
182 |
-
try:
|
183 |
-
# Generar respuesta en español
|
184 |
-
res_es = requests.post(
|
185 |
-
ENDPOINT_URL,
|
186 |
-
headers={"Authorization": f"Bearer {HF_API_TOKEN}", "Content-Type": "application/json"},
|
187 |
-
json={"inputs": prompt_es}, timeout=60
|
188 |
-
)
|
189 |
-
out_es = res_es.json()
|
190 |
-
if isinstance(out_es, list) and "generated_text" in out_es[0]:
|
191 |
-
# Limpiar la respuesta para asegurar buen formato de markdown
|
192 |
-
response_es = out_es[0]["generated_text"].replace(prompt_es.strip(), "").strip()
|
193 |
-
response_es = response_es.replace('\n', ' ').replace(' ', ' ').strip() # Reemplazar saltos de línea con espacios, limpiar espacios dobles
|
194 |
|
195 |
-
|
196 |
-
|
197 |
-
|
198 |
-
|
199 |
-
|
200 |
-
|
201 |
-
|
202 |
-
if isinstance(out_en, list) and "generated_text" in out_en[0]:
|
203 |
-
# Limpiar la respuesta para asegurar buen formato de markdown
|
204 |
-
response_en = out_en[0]["generated_text"].replace(prompt_en.strip(), "").strip()
|
205 |
-
response_en = response_en.replace('\n', ' ').replace(' ', ' ').strip() # Reemplazar saltos de línea con espacios, limpiar espacios dobles
|
206 |
|
207 |
-
|
208 |
-
|
209 |
-
f"<b>Respuesta en español:</b><br>" # Usamos <br> para el salto de línea HTML
|
210 |
-
f"{response_es}<br><br>" # Dos <br> para un doble salto de línea
|
211 |
-
f"<b>Answer in English:</b><br>"
|
212 |
-
f"{response_en}"
|
213 |
-
)
|
214 |
-
return full_response, ids, context, rdf_facts
|
215 |
-
#except Exception as e:
|
216 |
-
# return f"Ocurrió un error al generar la respuesta: {str(e)}", ids, context, rdf_facts
|
217 |
-
except Exception as e:
|
218 |
-
st.error(f"❌ Error técnico: {str(e)}")
|
219 |
-
st.stop() # Detiene la ejecución para que veas el error
|
220 |
|
|
|
|
|
|
|
|
|
|
|
221 |
|
222 |
-
# === MAIN APP ===
|
223 |
def main():
|
224 |
methods, embedder = load_all_components()
|
225 |
-
|
226 |
-
st.markdown(""
|
227 |
-
|
228 |
-
|
229 |
-
|
230 |
-
|
231 |
-
|
232 |
-
with st.expander("📌 **Resumen General**", expanded=True):
|
233 |
-
st.markdown("""
|
234 |
-
Esta aplicación ofrece **análisis impulsado por IA, Grafos y RAGs (GraphRAGs)** de lenguas indígenas de América del Sur,
|
235 |
-
integrando información de **Glottolog, Wikipedia y Wikidata**.
|
236 |
-
""")
|
237 |
-
#st.markdown("*Puedes preguntar en **español o inglés**, y el modelo responderá en **ambos idiomas**.*")
|
238 |
-
|
239 |
-
with st.sidebar:
|
240 |
-
st.markdown("### 📚 Información de Contacto")
|
241 |
-
st.markdown("""
|
242 |
-
- <span class="tech-badge">Correo: jxvera@gmail.com</span>
|
243 |
-
""", unsafe_allow_html=True)
|
244 |
-
st.markdown("---")
|
245 |
-
st.markdown("### 🚀 Inicio Rápido")
|
246 |
-
st.markdown("""
|
247 |
-
1. **Escribe una pregunta** en el cuadro de entrada
|
248 |
-
2. **Haz clic en 'Analizar'** para obtener la respuesta
|
249 |
-
3. **Explora los resultados** con los detalles expandibles
|
250 |
-
""")
|
251 |
-
|
252 |
-
st.markdown("---")
|
253 |
-
st.markdown("### 🔍 Preguntas de Ejemplo")
|
254 |
-
questions = [
|
255 |
-
"¿Qué idiomas están en peligro en Brasil? (What languages are endangered in Brazil?)",
|
256 |
-
"¿Qué idiomas se hablan en Perú? (What languages are spoken in Perú?)",
|
257 |
-
"¿Cuáles idiomas están relacionados con el Quechua? (Which languages are related to Quechua?)",
|
258 |
-
"¿Dónde se habla el Mapudungun? (Where is Mapudungun spoken?)"
|
259 |
-
]
|
260 |
-
|
261 |
-
for q in questions:
|
262 |
-
if st.button(q, key=f"suggested_{q}", use_container_width=True):
|
263 |
-
st.session_state.query = q.split(" (")[0]
|
264 |
-
|
265 |
-
st.markdown("---")
|
266 |
-
st.markdown("### ⚙️ Detalles Técnicos")
|
267 |
-
st.markdown("""
|
268 |
-
- <span class="tech-badge">Embeddings</span> GraphSAGE
|
269 |
-
- <span class="tech-badge">Modelo de Lenguaje</span> Mistral-7B-Instruct
|
270 |
-
- <span class="tech-badge">Grafo de Conocimiento</span> Integración basada en RDF
|
271 |
-
""", unsafe_allow_html=True)
|
272 |
-
|
273 |
-
st.markdown("---")
|
274 |
-
st.markdown("### 📂 Fuentes de Datos")
|
275 |
-
st.markdown("""
|
276 |
-
- **Glottolog** (Clasificación de idiomas)
|
277 |
-
- **Wikipedia** (Resúmenes textuales)
|
278 |
-
- **Wikidata** (Hechos estructurados)
|
279 |
-
""")
|
280 |
-
|
281 |
-
st.markdown("---")
|
282 |
-
st.markdown("### 📊 Parámetros de Análisis")
|
283 |
-
k = st.slider("Número de idiomas a analizar", 1, 10, 3)
|
284 |
-
st.markdown("---")
|
285 |
-
st.markdown("### 🔧 Opciones Avanzadas")
|
286 |
-
show_ctx = st.checkbox("Mostrar información de contexto", False)
|
287 |
-
show_rdf = st.checkbox("Mostrar hechos estructurados", False)
|
288 |
-
|
289 |
-
st.markdown("### 📝 Haz una pregunta sobre lenguas indígenas")
|
290 |
-
st.markdown("*(Puedes preguntar en español o inglés, y el modelo responderá en **ambos idiomas**.)*")
|
291 |
-
query = st.text_input(
|
292 |
-
"Ingresa tu pregunta:",
|
293 |
-
value=st.session_state.get("query", ""),
|
294 |
-
label_visibility="collapsed",
|
295 |
-
placeholder="Ej. ¿Qué lenguas se hablan en Perú?"
|
296 |
-
)
|
297 |
-
|
298 |
-
if st.button("Analizar", type="primary", use_container_width=True):
|
299 |
-
if not query:
|
300 |
-
st.warning("Por favor, ingresa una pregunta")
|
301 |
-
return
|
302 |
-
|
303 |
-
label = "LinkGraph"
|
304 |
-
method = methods[label]
|
305 |
-
|
306 |
-
#st.markdown(f"#### Método {label}")
|
307 |
-
#st.caption("Embeddings de GraphSAGE que capturan patrones en el grafo de conocimiento")
|
308 |
-
|
309 |
start = datetime.datetime.now()
|
310 |
response, lang_ids, context, rdf_data = generate_response(*method, query, k, embedder)
|
311 |
duration = (datetime.datetime.now() - start).total_seconds()
|
312 |
-
|
313 |
-
st.
|
314 |
-
|
315 |
-
|
316 |
-
|
317 |
-
|
318 |
-
|
319 |
-
</div>
|
320 |
-
</div>
|
321 |
-
""", unsafe_allow_html=True)
|
322 |
-
|
323 |
-
if show_ctx:
|
324 |
-
with st.expander(f"📖 Contexto de {len(lang_ids)} idiomas"):
|
325 |
-
for lang_id, ctx in zip(lang_ids, context):
|
326 |
-
st.markdown(f"<div class='language-card'>{ctx}</div>", unsafe_allow_html=True)
|
327 |
-
|
328 |
-
if show_rdf:
|
329 |
-
with st.expander("🔗 Hechos estructurados (RDF)"):
|
330 |
-
st.code("\n".join(rdf_data))
|
331 |
-
|
332 |
-
st.markdown("---")
|
333 |
-
st.markdown("""
|
334 |
-
<div style="font-size: 0.8rem; color: #64748b; text-align: center;">
|
335 |
-
<b>📌 Nota:</b> Esta herramienta está diseñada para investigadores, lingüistas y preservacionistas culturales.
|
336 |
-
Para mejores resultados, usa preguntas específicas sobre idiomas, familias o regiones.
|
337 |
-
</div>
|
338 |
-
""", unsafe_allow_html=True)
|
339 |
|
340 |
if __name__ == "__main__":
|
341 |
-
main()
|
|
|
8 |
import requests
|
9 |
from rdflib import Graph as RDFGraph, Namespace
|
10 |
from sentence_transformers import SentenceTransformer
|
|
|
11 |
|
12 |
# === STREAMLIT UI CONFIG ===
|
13 |
st.set_page_config(
|
|
|
22 |
)
|
23 |
|
24 |
# === CONFIGURATION ===
|
25 |
+
ENDPOINT_URL = "https://api-inference.huggingface.co/models/HuggingFaceH4/zephyr-7b-beta"
|
|
|
|
|
26 |
HF_API_TOKEN = os.getenv("HF_API_TOKEN")
|
27 |
if not HF_API_TOKEN:
|
28 |
st.error("⚠️ No se cargó el token HF_API_TOKEN desde los Secrets.")
|
|
|
35 |
# === CUSTOM CSS ===
|
36 |
st.markdown("""
|
37 |
<style>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
38 |
.tech-badge {
|
39 |
background-color: #ecfdf5;
|
40 |
color: #065f46;
|
|
|
51 |
def load_all_components():
|
52 |
embedder = SentenceTransformer(EMBEDDING_MODEL, device=DEVICE)
|
53 |
methods = {}
|
|
|
54 |
label, suffix, ttl, matrix_path = ("LinkGraph", "_hybrid_graphsage", "grafo_ttl_hibrido_graphsage.ttl", "embed_matrix_hybrid_graphsage.npy")
|
55 |
with open(f"id_map{suffix}.pkl", "rb") as f:
|
56 |
id_map = pickle.load(f)
|
|
|
90 |
except Exception as e:
|
91 |
return [("error", str(e))]
|
92 |
|
93 |
+
def query_llm(prompt):
|
94 |
+
try:
|
95 |
+
res = requests.post(
|
96 |
+
ENDPOINT_URL,
|
97 |
+
headers={"Authorization": f"Bearer {HF_API_TOKEN}", "Content-Type": "application/json"},
|
98 |
+
json={"inputs": prompt}, timeout=60
|
99 |
+
)
|
100 |
+
res.raise_for_status()
|
101 |
+
out = res.json()
|
102 |
+
if isinstance(out, list):
|
103 |
+
if len(out) > 0 and isinstance(out[0], dict) and "generated_text" in out[0]:
|
104 |
+
return out[0]["generated_text"].strip()
|
105 |
+
elif isinstance(out, dict) and "generated_text" in out:
|
106 |
+
return out["generated_text"].strip()
|
107 |
+
return "Sin respuesta del modelo."
|
108 |
+
except Exception as e:
|
109 |
+
return f"Error al consultar el modelo: {str(e)}"
|
110 |
+
|
111 |
def generate_response(matrix, id_map, G, rdf, user_question, k, embedder):
|
112 |
ids = get_top_k(matrix, id_map, user_question, k, embedder)
|
113 |
context = [get_context(G, i) for i in ids]
|
|
|
115 |
for i in ids:
|
116 |
rdf_facts.extend([f"{p}: {v}" for p, v in query_rdf(rdf, i)])
|
117 |
|
118 |
+
prompt_es = (
|
119 |
+
"Eres un experto en lenguas indígenas sudamericanas.\n"
|
120 |
+
"Usa solo la información del contexto y hechos RDF siguientes.\n\n"
|
121 |
+
+ "### CONTEXTO:\n" + "\n".join(context) + "\n\n"
|
122 |
+
+ "### RELACIONES RDF:\n" + "\n".join(rdf_facts) + "\n\n"
|
123 |
+
+ f"### PREGUNTA:\n{user_question}\n\nRespuesta breve en español:"
|
124 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
125 |
|
126 |
+
prompt_en = (
|
127 |
+
"You are an expert in South American indigenous languages.\n"
|
128 |
+
"Use only the following context and RDF facts to answer.\n\n"
|
129 |
+
+ "### CONTEXT:\n" + "\n".join(context) + "\n\n"
|
130 |
+
+ "### RDF RELATIONS:\n" + "\n".join(rdf_facts) + "\n\n"
|
131 |
+
+ f"### QUESTION:\n{user_question}\n\nShort answer in English:"
|
132 |
+
)
|
|
|
|
|
|
|
|
|
133 |
|
134 |
+
response_es = query_llm(prompt_es)
|
135 |
+
response_en = query_llm(prompt_en)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
136 |
|
137 |
+
full_response = (
|
138 |
+
f"<b>Respuesta en español:</b><br>{response_es}<br><br>"
|
139 |
+
f"<b>Answer in English:</b><br>{response_en}"
|
140 |
+
)
|
141 |
+
return full_response, ids, context, rdf_facts
|
142 |
|
|
|
143 |
def main():
|
144 |
methods, embedder = load_all_components()
|
145 |
+
st.title("Atlas de Lenguas: Lenguas Indígenas Sudamericanas")
|
146 |
+
st.markdown("<span class='tech-badge'>Correo: jxvera@gmail.com</span>", unsafe_allow_html=True)
|
147 |
+
query = st.text_input("Escribe tu pregunta sobre lenguas indígenas:")
|
148 |
+
k = st.slider("Número de lenguas similares a recuperar", min_value=1, max_value=10, value=3)
|
149 |
+
if st.button("Analizar"):
|
150 |
+
method = methods["LinkGraph"]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
151 |
start = datetime.datetime.now()
|
152 |
response, lang_ids, context, rdf_data = generate_response(*method, query, k, embedder)
|
153 |
duration = (datetime.datetime.now() - start).total_seconds()
|
154 |
+
st.markdown(response, unsafe_allow_html=True)
|
155 |
+
st.caption(f"⏱️ {duration:.2f} segundos | 🌐 {len(lang_ids)} idiomas analizados")
|
156 |
+
with st.expander("📖 Contexto"):
|
157 |
+
for ctx in context:
|
158 |
+
st.markdown(ctx)
|
159 |
+
with st.expander("🔗 Hechos RDF"):
|
160 |
+
st.code("\n".join(rdf_data))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
161 |
|
162 |
if __name__ == "__main__":
|
163 |
+
main()
|