Upload 5 files

README.md

@@ -1,12 +1,27 @@
----
-title: Chatbot
-emoji: 👁
-colorFrom: indigo
-colorTo: pink
-sdk: gradio
-sdk_version: 6.10.0
-app_file: app.py
-pinned: false
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# RAG Question Answering System
+
+##  Descripción del proyecto
+
+Este proyecto implementa un sistema de **Retrieval-Augmented Generation (RAG)** que permite responder preguntas en inglés basándose exclusivamente en una base de conocimiento local.
+
+El sistema combina:
+- Un modelo de **embeddings** para recuperar documentos relevantes.
+- Un modelo de **lenguaje (LLM)** para generar respuestas basadas en esos documentos.
+
+Flujo del sistema:
+1. El usuario introduce una pregunta.
+2. Se recuperan los documentos más relevantes mediante similitud coseno.
+3. Se construye un prompt con ese contexto.
+4. El modelo genera una respuesta basada únicamente en los documentos.
+
+La interfaz se implementa con **Gradio**.
+
+---
+
+## ⚙️ Instalación
+
+### 1. Clonar el repositorio
+
+```bash
+git clone <repo-url>
+cd <repo-name>

app.py

@@ -0,0 +1,43 @@
+import gradio as gr
+from rag_engine import recuperar_documentos, generar_respuesta
+
+def ask(query, top_k, umbral):
+    docs = recuperar_documentos(query, int(top_k), float(umbral))
+    respuesta = generar_respuesta(query, docs)
+
+    if docs:
+        docs_formateados = "\n\n---\n\n".join(docs)
+    else:
+        docs_formateados = "No relevant documents found."
+
+    return respuesta, docs_formateados
+
+
+with gr.Blocks() as demo:
+    gr.Markdown("# RAG Question Answering System")
+    gr.Markdown("Ask a question based on the provided documents.")
+
+    query = gr.Textbox(
+        label="Your Question",
+        placeholder="Type your question..."
+    )
+
+    top_k = gr.Slider(1, 5, value=5, step=1, label="Top K Documents")
+
+    umbral = gr.Slider(0.0, 1.0, value=0.55, step=0.05, label="Similarity Threshold")
+
+    respuesta = gr.Textbox(label="Answer", lines=3)
+
+    docs = gr.Textbox(label="Retrieved Documents", lines=6, max_lines=15)
+
+    boton = gr.Button("Enviar")
+
+    boton.click(
+        fn=ask,
+        inputs=[query, top_k, umbral],
+        outputs=[respuesta, docs]
+    )
+
+
+if __name__ == "__main__":
+    demo.launch()

documents.json

@@ -0,0 +1,7 @@
+{
+    "doc1": "Hospital contact details: You can contact the hospital at email testing@gmail.com, phone +911234567890, or visit us at xyz, abc, 1234, Nepal.",
+    "doc2": "Hospital's working hours: The hospital's working hours are 7:00 AM - 8:00 PM daily.",
+    "doc3": "Official email address: The official email address to contact the hospital is testing@gmail.com.",
+    "doc4": "Main services: We provide comprehensive healthcare services including emergency care, diagnostic testing, surgical procedures, maternity services, and specialized treatments.",
+    "doc5": "Hospital location: The hospital is located at xyz, abc, 1234, Nepal."
+}

rag_engine.py

@@ -0,0 +1,88 @@
+import json
+import torch
+import numpy as np
+from sentence_transformers import SentenceTransformer
+from transformers import AutoTokenizer, AutoModelForCausalLM
+
+# ================================
+# CARGA DE MODELOS
+# ================================
+
+embedding_model = SentenceTransformer("MongoDB/mdbr-leaf-ir")
+
+tokenizer = AutoTokenizer.from_pretrained("PleIAs/Pleias-RAG-350M")
+llm_model = AutoModelForCausalLM.from_pretrained("PleIAs/Pleias-RAG-350M")
+
+# ================================
+# CARGA DE DOCUMENTOS (FIX AQUÍ)
+# ================================
+
+with open("documents.json", "r", encoding="utf-8") as f:
+    documents_dict = json.load(f)
+
+# 🔥 IMPORTANTE: convertir dict → lista de textos
+documents = list(documents_dict.values())
+
+# ================================
+# PRECOMPUTO DE EMBEDDINGS
+# ================================
+
+doc_embeddings = embedding_model.encode(documents, convert_to_tensor=True)
+
+# ================================
+# RECUPERAR DOCUMENTOS
+# ================================
+
+def recuperar_documentos(consulta, top_k=2, umbral=0.4):
+    query_embedding = embedding_model.encode(consulta, convert_to_tensor=True)
+
+    similitudes = torch.nn.functional.cosine_similarity(
+        query_embedding.unsqueeze(0),
+        doc_embeddings,
+        dim=1
+    )
+
+    indices_ordenados = torch.argsort(similitudes, descending=True)
+
+    resultados = []
+    for idx in indices_ordenados:
+        if similitudes[idx].item() >= umbral:
+            resultados.append(documents[idx])
+        if len(resultados) >= top_k:
+            break
+
+    return resultados
+
+# ================================
+# GENERAR RESPUESTA
+# ================================
+
+def generar_respuesta(consulta, documentos_recuperados):
+    contexto = " ".join(documentos_recuperados)
+
+    prompt = f"""Answer the question based only on the context provided
+Context: {contexto}
+Question: {consulta}
+Answer:"""
+
+    inputs = tokenizer(prompt, return_tensors="pt")
+
+    outputs = llm_model.generate(
+        **inputs,
+        max_new_tokens=150
+    )
+
+    respuesta = tokenizer.decode(outputs[0], skip_special_tokens=True)
+
+    # eliminar el prompt del output
+    respuesta = respuesta.replace(prompt, "").strip()
+
+    return respuesta
+
+# ================================
+# FUNCIÓN FINAL
+# ================================
+
+def preguntar(consulta, top_k=2, umbral=0.4):
+    docs = recuperar_documentos(consulta, top_k, umbral)
+    return generar_respuesta(consulta, docs)

requirements.txt

@@ -0,0 +1,8 @@
+torch
+transformers
+sentence-transformers
+scikit-learn
+fastapi
+uvicorn
+gradio
+pydantic