Update app.py
Browse files
app.py
CHANGED
@@ -1,63 +1,83 @@
|
|
1 |
-
|
2 |
-
|
3 |
-
|
4 |
-
from
|
|
|
|
|
5 |
from langchain_huggingface import HuggingFaceEmbeddings
|
|
|
|
|
|
|
|
|
6 |
from langchain import hub
|
7 |
-
|
8 |
|
9 |
-
#
|
10 |
-
|
|
|
|
|
11 |
|
12 |
-
|
13 |
-
|
|
|
|
|
|
|
14 |
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
max_new_tokens=512,
|
20 |
-
temperature=0.7
|
21 |
-
)
|
22 |
|
23 |
-
|
24 |
-
|
|
|
|
|
|
|
25 |
|
26 |
-
#
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
)
|
31 |
|
32 |
-
|
33 |
-
|
34 |
-
embedding_function=embedding_function
|
35 |
-
)
|
36 |
|
37 |
-
# RAG
|
38 |
-
def
|
39 |
-
|
40 |
-
prompt = hub.pull("rlm/rag-prompt")
|
41 |
-
rag_chain = prompt | llm | parser
|
42 |
|
|
|
|
|
|
|
|
|
43 |
context = []
|
44 |
for doc, score in docs:
|
45 |
if score < 7:
|
46 |
-
|
47 |
-
|
48 |
-
if context:
|
49 |
-
|
50 |
-
|
51 |
-
|
|
|
|
|
|
|
|
|
52 |
else:
|
53 |
-
return "No tengo informaci贸n
|
54 |
-
|
55 |
-
#
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
63 |
|
|
|
1 |
+
import os
|
2 |
+
import subprocess
|
3 |
+
import gradio as gr
|
4 |
+
from langchain.text_splitter import RecursiveCharacterTextSplitter
|
5 |
+
from langchain_community.document_loaders import WebBaseLoader
|
6 |
+
from langchain_community.vectorstores import Chroma
|
7 |
from langchain_huggingface import HuggingFaceEmbeddings
|
8 |
+
from langchain.document_loaders import PyPDFLoader
|
9 |
+
import requests
|
10 |
+
from rerankers import Reranker
|
11 |
+
from langchain_community.chat_models import ChatOllama
|
12 |
from langchain import hub
|
13 |
+
from langchain_core.output_parsers import StrOutputParser
|
14 |
|
15 |
+
# Paso 1: Instalar ollama
|
16 |
+
subprocess.run("curl -fsSL https://ollama.com/install.sh | sh", shell=True, check=True)
|
17 |
+
subprocess.run("ollama serve &", shell=True, check=True)
|
18 |
+
subprocess.run("ollama pull llama3.2:1b", shell=True, check=True)
|
19 |
|
20 |
+
# Paso 2: Descargar el documento PDF
|
21 |
+
URL = "https://gruposdetrabajo.sefh.es/gefp/images/stories/documentos/4-ATENCION-FARMACEUTICA/Nutricion/Manual_basico_N_clinica_y_Dietetica_Valencia_2012.pdf"
|
22 |
+
response = requests.get(URL)
|
23 |
+
with open("Manual_de_nutrici贸n_clinica.pdf", "wb") as f:
|
24 |
+
f.write(response.content)
|
25 |
|
26 |
+
# Paso 3: Inicializar el modelo y los embeddings
|
27 |
+
local_llm = "llama3.2:1b"
|
28 |
+
llm = ChatOllama(model=local_llm, temperature=0, top_k=50, top_p=0.95)
|
29 |
+
chain = llm | StrOutputParser()
|
|
|
|
|
|
|
30 |
|
31 |
+
# Cargar y procesar el PDF
|
32 |
+
loader = PyPDFLoader("Manual_de_nutrici贸n_clinica.pdf")
|
33 |
+
documents = loader.load()
|
34 |
+
text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=20)
|
35 |
+
all_splits = text_splitter.split_documents(documents)
|
36 |
|
37 |
+
# Crear embeddings y la base de datos vectorial
|
38 |
+
model_name = "sentence-transformers/paraphrase-multilingual-mpnet-base-v2"
|
39 |
+
embeddings = HuggingFaceEmbeddings(model_name=model_name)
|
40 |
+
vectordb = Chroma.from_documents(documents=all_splits, embedding=embeddings, persist_directory="chroma_db")
|
|
|
41 |
|
42 |
+
# Inicializar el reranker
|
43 |
+
ranker = Reranker("answerdotai/answerai-colbert-small-v1", model_type='colbert')
|
|
|
|
|
44 |
|
45 |
+
# Paso 4: Definir la funci贸n RAG con reranking
|
46 |
+
def format_docs(docs):
|
47 |
+
return "\n\n".join(doc[0].page_content for doc in docs)
|
|
|
|
|
48 |
|
49 |
+
def test_rag_reranking(query, ranker):
|
50 |
+
docs = vectordb.similarity_search_with_score(query)
|
51 |
+
prompt = hub.pull("rlm/rag-prompt")
|
52 |
+
rag_chain = prompt | llm | StrOutputParser()
|
53 |
context = []
|
54 |
for doc, score in docs:
|
55 |
if score < 7:
|
56 |
+
doc_details = doc.to_json()['kwargs']
|
57 |
+
context.append(doc_details['page_content'])
|
58 |
+
if len(context) > 0:
|
59 |
+
# Aplicar reranking
|
60 |
+
ranking = ranker.rank(query=query, docs=context)
|
61 |
+
# Tomar el contexto m谩s relevante
|
62 |
+
useful_context = ranking[0].text
|
63 |
+
# Generar la respuesta
|
64 |
+
generation = rag_chain.invoke({"context": useful_context, "question": query})
|
65 |
+
return generation
|
66 |
else:
|
67 |
+
return "No tengo informaci贸n para responder a esta pregunta"
|
68 |
+
|
69 |
+
# Paso 5: Crear una interfaz con Gradio
|
70 |
+
def answer_query(query):
|
71 |
+
return test_rag_reranking(query, ranker)
|
72 |
+
|
73 |
+
interface = gr.Interface(
|
74 |
+
fn=answer_query,
|
75 |
+
inputs=gr.Textbox(label="Ingresa tu pregunta sobre nutrici贸n:"),
|
76 |
+
outputs=gr.Textbox(label="Respuesta:"),
|
77 |
+
title="Respuesta a Preguntas sobre Nutrici贸n",
|
78 |
+
description="Haz preguntas sobre nutrici贸n basadas en el Manual B谩sico de Nutrici贸n Cl铆nica y Diet茅tica (Valencia, 2012)."
|
79 |
+
)
|
80 |
+
|
81 |
+
# Lanzar la interfaz
|
82 |
+
interface.launch()
|
83 |
|