Spaces:
Sleeping
Sleeping
fix con file ingest
Browse files- app.py +8 -16
- rag_ingest.py +254 -0
app.py
CHANGED
|
@@ -11,25 +11,17 @@ st.title("π€ RAG Chatbot β INSIEL")
|
|
| 11 |
|
| 12 |
import os
|
| 13 |
import streamlit as st
|
|
|
|
| 14 |
|
| 15 |
-
|
| 16 |
-
def run_notebook_once():
|
| 17 |
if not os.path.exists("vectorstore"):
|
| 18 |
-
st.info("Inizializzazione: generazione vectorstore in
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 19 |
|
| 20 |
-
import nbformat
|
| 21 |
-
from nbconvert.preprocessors import ExecutePreprocessor
|
| 22 |
-
|
| 23 |
-
with open("rag.ipynb") as f:
|
| 24 |
-
nb = nbformat.read(f, as_version=4)
|
| 25 |
-
|
| 26 |
-
ep = ExecutePreprocessor(timeout=600)
|
| 27 |
-
ep.preprocess(nb, {"metadata": {"path": "./"}})
|
| 28 |
-
|
| 29 |
-
st.success("Vectorstore generata correttamente β
")
|
| 30 |
-
|
| 31 |
-
# Chiama la funzione prima di tutto
|
| 32 |
-
run_notebook_once()
|
| 33 |
|
| 34 |
|
| 35 |
|
|
|
|
| 11 |
|
| 12 |
import os
|
| 13 |
import streamlit as st
|
| 14 |
+
import subprocess
|
| 15 |
|
| 16 |
+
def run_ingest_if_needed():
|
|
|
|
| 17 |
if not os.path.exists("vectorstore"):
|
| 18 |
+
st.info("Inizializzazione: generazione vectorstore in corso...")
|
| 19 |
+
try:
|
| 20 |
+
subprocess.run(["python", "rag_ingest.py"], check=True)
|
| 21 |
+
st.success("Vectorstore generata correttamente β
")
|
| 22 |
+
except subprocess.CalledProcessError:
|
| 23 |
+
st.error("Errore durante la generazione della vectorstore.")
|
| 24 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 25 |
|
| 26 |
|
| 27 |
|
rag_ingest.py
ADDED
|
@@ -0,0 +1,254 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python
|
| 2 |
+
# coding: utf-8
|
| 3 |
+
|
| 4 |
+
# In[1]:
|
| 5 |
+
|
| 6 |
+
|
| 7 |
+
get_ipython().system('pip install docling chromadb sentence-transformers')
|
| 8 |
+
|
| 9 |
+
|
| 10 |
+
# In[2]:
|
| 11 |
+
|
| 12 |
+
|
| 13 |
+
get_ipython().system('pip install pymupdf tqdm spacy')
|
| 14 |
+
get_ipython().system('python -m spacy download it_core_news_sm')
|
| 15 |
+
|
| 16 |
+
|
| 17 |
+
# In[3]:
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
get_ipython().system('pip install transformers')
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
# In[4]:
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
import fitz # PyMuPDF
|
| 27 |
+
from tqdm.auto import tqdm
|
| 28 |
+
import pandas as pd
|
| 29 |
+
|
| 30 |
+
def text_formatter(text: str) -> str:
|
| 31 |
+
# Pulizia semplice
|
| 32 |
+
import re
|
| 33 |
+
text = text.replace("\n", " ").strip()
|
| 34 |
+
text = re.sub(r"[ \t]{2,}", " ", text)
|
| 35 |
+
text = re.sub(r"\.{2,}", " ", text) # sostituisce ... con spazio
|
| 36 |
+
text = re.sub(r"Pagina\s+\d+\s+di\s+\d+", "", text, flags=re.IGNORECASE)
|
| 37 |
+
text = re.sub(r"Creazione VM su Cloud INSIEL","", text)
|
| 38 |
+
text = re.sub(r"IO_XX_00_XX ISTRUZIONE OPERATIVA 22/10/2024", "",text)
|
| 39 |
+
text = re.sub(r"IO_XX_00_XX ISTRUZIONE OPERATIVA 22/10/2024 ","",text)
|
| 40 |
+
return text.strip()
|
| 41 |
+
|
| 42 |
+
def open_and_read_pdf(pdf_path: str):
|
| 43 |
+
doc = fitz.open(pdf_path)
|
| 44 |
+
pages = []
|
| 45 |
+
for page_number, page in tqdm(enumerate(doc), total=len(doc), desc="π Lettura pagine PDF"):
|
| 46 |
+
text = text_formatter(page.get_text())
|
| 47 |
+
pages.append({
|
| 48 |
+
"page_number": page_number + 1,
|
| 49 |
+
"page_char_count": len(text),
|
| 50 |
+
"page_word_count": len(text.split()),
|
| 51 |
+
"page_token_estimate": len(text) // 4,
|
| 52 |
+
"text": text
|
| 53 |
+
})
|
| 54 |
+
return pages
|
| 55 |
+
|
| 56 |
+
pdf_path = "data/insiel.pdf" # Cambia se il tuo file Γ¨ altrove
|
| 57 |
+
pages_and_texts = open_and_read_pdf(pdf_path)
|
| 58 |
+
|
| 59 |
+
|
| 60 |
+
# In[5]:
|
| 61 |
+
|
| 62 |
+
|
| 63 |
+
import spacy
|
| 64 |
+
nlp = spacy.load("it_core_news_sm")
|
| 65 |
+
|
| 66 |
+
# Spezza il testo di ogni pagina in frasi
|
| 67 |
+
for page in tqdm(pages_and_texts, desc="βοΈ Split in frasi"):
|
| 68 |
+
doc = nlp(page["text"])
|
| 69 |
+
sentences = [sent.text.strip() for sent in doc.sents if sent.text.strip()]
|
| 70 |
+
page["sentence_chunks"] = []
|
| 71 |
+
|
| 72 |
+
CHUNK_SIZE = 10 # Gruppi da 5 frasi
|
| 73 |
+
for i in range(0, len(sentences), CHUNK_SIZE):
|
| 74 |
+
chunk = sentences[i:i + CHUNK_SIZE]
|
| 75 |
+
page["sentence_chunks"].append(chunk)
|
| 76 |
+
|
| 77 |
+
|
| 78 |
+
# In[6]:
|
| 79 |
+
|
| 80 |
+
|
| 81 |
+
pages_and_texts[65]
|
| 82 |
+
|
| 83 |
+
|
| 84 |
+
# In[7]:
|
| 85 |
+
|
| 86 |
+
|
| 87 |
+
df = pd.DataFrame(pages_and_texts)
|
| 88 |
+
df["chunk_id"] = df.index.astype(str)
|
| 89 |
+
|
| 90 |
+
# Mostra i primi
|
| 91 |
+
df.tail()
|
| 92 |
+
|
| 93 |
+
|
| 94 |
+
# In[8]:
|
| 95 |
+
|
| 96 |
+
|
| 97 |
+
df.shape
|
| 98 |
+
|
| 99 |
+
|
| 100 |
+
# In[9]:
|
| 101 |
+
|
| 102 |
+
|
| 103 |
+
df[df['page_token_estimate'] < 60].count()
|
| 104 |
+
|
| 105 |
+
|
| 106 |
+
# In[10]:
|
| 107 |
+
|
| 108 |
+
|
| 109 |
+
final = df[df['page_token_estimate'] > 60]
|
| 110 |
+
|
| 111 |
+
|
| 112 |
+
# In[11]:
|
| 113 |
+
|
| 114 |
+
|
| 115 |
+
final.describe().round(2)
|
| 116 |
+
|
| 117 |
+
|
| 118 |
+
# In[12]:
|
| 119 |
+
|
| 120 |
+
|
| 121 |
+
get_ipython().system('pip install sentence-transformers chromadb')
|
| 122 |
+
|
| 123 |
+
|
| 124 |
+
# In[13]:
|
| 125 |
+
|
| 126 |
+
|
| 127 |
+
from sentence_transformers import SentenceTransformer
|
| 128 |
+
from tqdm.notebook import tqdm
|
| 129 |
+
|
| 130 |
+
embedding_model = SentenceTransformer("sentence-transformers/distiluse-base-multilingual-cased-v1")
|
| 131 |
+
|
| 132 |
+
texts = final["text"].tolist()
|
| 133 |
+
chunk_ids = final["chunk_id"].tolist()
|
| 134 |
+
metadatas = [{"page": int(p)} for p in final["page_number"]]
|
| 135 |
+
|
| 136 |
+
embeddings = embedding_model.encode(texts, show_progress_bar=True)
|
| 137 |
+
|
| 138 |
+
|
| 139 |
+
# In[14]:
|
| 140 |
+
|
| 141 |
+
|
| 142 |
+
import chromadb
|
| 143 |
+
|
| 144 |
+
# nuovo client
|
| 145 |
+
client = chromadb.PersistentClient(path="./vectorstore")
|
| 146 |
+
|
| 147 |
+
# collection
|
| 148 |
+
collection = client.get_or_create_collection("insiel_chunks")
|
| 149 |
+
|
| 150 |
+
# aggiunta
|
| 151 |
+
collection.add(
|
| 152 |
+
documents=texts,
|
| 153 |
+
embeddings=embeddings.tolist(),
|
| 154 |
+
metadatas=metadatas,
|
| 155 |
+
ids=chunk_ids
|
| 156 |
+
)
|
| 157 |
+
|
| 158 |
+
|
| 159 |
+
# In[16]:
|
| 160 |
+
|
| 161 |
+
|
| 162 |
+
"""query = input("Domanda: ")
|
| 163 |
+
query_embedding = embedding_model.encode([query])
|
| 164 |
+
"""
|
| 165 |
+
results = collection.query(
|
| 166 |
+
query_embeddings=query_embedding,
|
| 167 |
+
n_results=3 # puoi aumentare a 5, 10, ecc.
|
| 168 |
+
)
|
| 169 |
+
|
| 170 |
+
|
| 171 |
+
# In[17]:
|
| 172 |
+
|
| 173 |
+
|
| 174 |
+
"""for i, (doc, meta) in enumerate(zip(results["documents"][0], results["metadatas"][0])):
|
| 175 |
+
print(f"\nπΉ RISULTATO {i+1} (pagina {meta['page']}):")
|
| 176 |
+
print(doc[:500] + "...\n---")
|
| 177 |
+
"""
|
| 178 |
+
|
| 179 |
+
|
| 180 |
+
# In[18]:
|
| 181 |
+
|
| 182 |
+
|
| 183 |
+
import torch
|
| 184 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
|
| 185 |
+
|
| 186 |
+
model_id = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
|
| 187 |
+
|
| 188 |
+
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
| 189 |
+
model = AutoModelForCausalLM.from_pretrained(model_id).to(torch.device("cpu"))
|
| 190 |
+
|
| 191 |
+
rag_chat = pipeline("text-generation", model=model, tokenizer=tokenizer, max_new_tokens=200, device=-1)
|
| 192 |
+
|
| 193 |
+
|
| 194 |
+
# In[ ]:
|
| 195 |
+
|
| 196 |
+
|
| 197 |
+
def generate_rag_response_local(query, retrieved_chunks):
|
| 198 |
+
context = "\n\n".join(retrieved_chunks)
|
| 199 |
+
|
| 200 |
+
prompt = f"""[INST] Usa solo le informazioni fornite nel contesto qui sotto per rispondere alla domanda, la risposta deve finire sempre con un punto.
|
| 201 |
+
Se la risposta non Γ¨ presente, di' chiaramente che non Γ¨ specificato nel documento.
|
| 202 |
+
|
| 203 |
+
Contesto:
|
| 204 |
+
{context}
|
| 205 |
+
|
| 206 |
+
Domanda: {query}
|
| 207 |
+
Risposta: [/INST]
|
| 208 |
+
"""
|
| 209 |
+
result = rag_chat(prompt)[0]["generated_text"]
|
| 210 |
+
return result.split("Risposta:")[-1].strip()
|
| 211 |
+
|
| 212 |
+
|
| 213 |
+
# In[ ]:
|
| 214 |
+
|
| 215 |
+
|
| 216 |
+
# π§ Inserisci la domanda
|
| 217 |
+
query = input("Domanda: ")
|
| 218 |
+
|
| 219 |
+
# π Ottieni l'embedding della query (usa sentence-transformers, NON il modello generativo!)
|
| 220 |
+
query_embedding = embedding_model.encode([query])
|
| 221 |
+
|
| 222 |
+
# π Retrieval dei chunk piΓΉ simili da Chroma
|
| 223 |
+
results = collection.query(
|
| 224 |
+
query_embeddings=query_embedding,
|
| 225 |
+
n_results=3
|
| 226 |
+
)
|
| 227 |
+
|
| 228 |
+
# π§± Estrai i chunk di contesto
|
| 229 |
+
retrieved_chunks = results["documents"][0]
|
| 230 |
+
|
| 231 |
+
# π€ Genera la risposta usando il modello open-source locale
|
| 232 |
+
response = generate_rag_response_local(query, retrieved_chunks)
|
| 233 |
+
|
| 234 |
+
# π¨οΈ Mostra la risposta
|
| 235 |
+
print("π€ Risposta:\n", response)
|
| 236 |
+
|
| 237 |
+
|
| 238 |
+
# In[ ]:
|
| 239 |
+
|
| 240 |
+
|
| 241 |
+
retrieved_chunks
|
| 242 |
+
|
| 243 |
+
|
| 244 |
+
# In[ ]:
|
| 245 |
+
|
| 246 |
+
|
| 247 |
+
results
|
| 248 |
+
|
| 249 |
+
|
| 250 |
+
# In[ ]:
|
| 251 |
+
|
| 252 |
+
|
| 253 |
+
|
| 254 |
+
|