Update app.py
Browse files
app.py
CHANGED
@@ -10,6 +10,10 @@ from sentence_transformers import SentenceTransformer
|
|
10 |
import faiss
|
11 |
import json
|
12 |
import numpy as np
|
|
|
|
|
|
|
|
|
13 |
|
14 |
logging.basicConfig(
|
15 |
level=logging.DEBUG,
|
@@ -23,7 +27,7 @@ if not API_KEY:
|
|
23 |
logger.error("HF_API_KEY non impostata.")
|
24 |
raise EnvironmentError("HF_API_KEY non impostata.")
|
25 |
|
26 |
-
client = InferenceClient(
|
27 |
|
28 |
RDF_FILE = "Ontologia.rdf"
|
29 |
HF_MODEL = "Qwen/Qwen2.5-72B-Instruct"
|
@@ -137,7 +141,7 @@ async def call_hf_model(messages, temperature=0.5, max_tokens=1024)->str:
|
|
137 |
max_tokens=max_tokens,
|
138 |
top_p=0.9
|
139 |
)
|
140 |
-
raw=resp["choices"][0]["message"]["content"]
|
141 |
# Forziamo la query su linea singola se multiline
|
142 |
single_line = " ".join(raw.splitlines())
|
143 |
logger.debug(f"Risposta HF single-line: {single_line}")
|
@@ -146,100 +150,101 @@ async def call_hf_model(messages, temperature=0.5, max_tokens=1024)->str:
|
|
146 |
logger.error(f"HuggingFace error: {e}")
|
147 |
raise HTTPException(status_code=500, detail=str(e))
|
148 |
|
149 |
-
app=FastAPI()
|
150 |
|
151 |
class QueryRequest(BaseModel):
|
152 |
-
message:str
|
153 |
-
max_tokens:int=1024
|
154 |
-
temperature:float=0.5
|
155 |
|
156 |
@app.post("/generate-response/")
|
157 |
-
async def generate_response(req:QueryRequest):
|
158 |
-
user_input=req.message
|
159 |
logger.info(f"Utente dice: {user_input}")
|
160 |
|
161 |
# Recupera documenti rilevanti usando RAG
|
162 |
relevant_docs = retrieve_relevant_documents(user_input, top_k=3)
|
163 |
retrieved_text = "\n".join([doc['text'] for doc in relevant_docs])
|
164 |
|
165 |
-
sys_msg=create_system_message(knowledge_text, retrieved_text)
|
166 |
-
msgs=[
|
167 |
-
{"role":"system","content":sys_msg},
|
168 |
-
{"role":"user","content":user_input}
|
169 |
]
|
|
|
170 |
# Primo tentativo
|
171 |
-
r1=await call_hf_model(msgs, req.temperature, req.max_tokens)
|
172 |
logger.info(f"PRIMA RISPOSTA:\n{r1}")
|
173 |
|
174 |
# Se non parte con "PREFIX base:"
|
175 |
if not r1.startswith("PREFIX base:"):
|
176 |
-
sc=f"Non hai risposto con query SPARQL su una sola riga. Riprova. Domanda: {user_input}"
|
177 |
-
msgs2=[
|
178 |
-
{"role":"system","content":sys_msg},
|
179 |
-
{"role":"assistant","content":r1},
|
180 |
-
{"role":"user","content":sc}
|
181 |
]
|
182 |
-
r2=await call_hf_model(msgs2,req.temperature,req.max_tokens)
|
183 |
logger.info(f"SECONDA RISPOSTA:\n{r2}")
|
184 |
if r2.startswith("PREFIX base:"):
|
185 |
-
sparql_query=r2
|
186 |
else:
|
187 |
-
return {"type":"NATURAL","response": r2}
|
188 |
else:
|
189 |
-
sparql_query=r1
|
190 |
|
191 |
# Esegui la query con rdflib
|
192 |
-
g=rdflib.Graph()
|
193 |
try:
|
194 |
-
g.parse(RDF_FILE,format="xml")
|
195 |
except Exception as e:
|
196 |
logger.error(f"Parsing RDF error: {e}")
|
197 |
-
return {"type":"ERROR","response":f"Parsing RDF error: {e}"}
|
198 |
|
199 |
try:
|
200 |
-
results=g.query(sparql_query)
|
201 |
except Exception as e:
|
202 |
-
fallback=f"La query SPARQL ha fallito. Riprova. Domanda: {user_input}"
|
203 |
-
msgs3=[
|
204 |
-
{"role":"system","content":sys_msg},
|
205 |
-
{"role":"assistant","content":sparql_query},
|
206 |
-
{"role":"user","content":fallback}
|
207 |
]
|
208 |
-
r3=await call_hf_model(msgs3,req.temperature,req.max_tokens)
|
209 |
if r3.startswith("PREFIX base:"):
|
210 |
-
sparql_query=r3
|
211 |
try:
|
212 |
-
results=g.query(sparql_query)
|
213 |
except Exception as e2:
|
214 |
-
return {"type":"ERROR","response":f"Query fallita di nuovo: {e2}"}
|
215 |
else:
|
216 |
-
return {"type":"NATURAL","response":r3}
|
217 |
|
218 |
-
if len(results)==0:
|
219 |
-
return {"type":"NATURAL","sparql_query":sparql_query,"response":"Nessun risultato."}
|
220 |
|
221 |
# Confeziona risultati
|
222 |
-
row_list=[]
|
223 |
for row in results:
|
224 |
-
row_str=", ".join([f"{k}:{v}" for k,v in row.asdict().items()])
|
225 |
row_list.append(row_str)
|
226 |
-
results_str="\n".join(row_list)
|
227 |
|
228 |
# Spiegazione
|
229 |
-
exp_prompt=create_explanation_prompt(results_str)
|
230 |
-
msgs4=[
|
231 |
-
{"role":"system","content":exp_prompt},
|
232 |
-
{"role":"user","content":""}
|
233 |
]
|
234 |
-
explanation=await call_hf_model(msgs4,req.temperature,req.max_tokens)
|
235 |
|
236 |
return {
|
237 |
-
"type":"NATURAL",
|
238 |
-
"sparql_query":sparql_query,
|
239 |
-
"sparql_results":row_list,
|
240 |
-
"explanation":explanation
|
241 |
}
|
242 |
|
243 |
@app.get("/")
|
244 |
def home():
|
245 |
-
return {"message":"Prompt lascia libertà su come chiamare la proprietà del materiale, ma suggerisce un possibile 'materialeOpera'."}
|
|
|
10 |
import faiss
|
11 |
import json
|
12 |
import numpy as np
|
13 |
+
from dotenv import load_dotenv
|
14 |
+
|
15 |
+
# Carica le variabili d'ambiente
|
16 |
+
load_dotenv()
|
17 |
|
18 |
logging.basicConfig(
|
19 |
level=logging.DEBUG,
|
|
|
27 |
logger.error("HF_API_KEY non impostata.")
|
28 |
raise EnvironmentError("HF_API_KEY non impostata.")
|
29 |
|
30 |
+
client = InferenceClient(token=API_KEY)
|
31 |
|
32 |
RDF_FILE = "Ontologia.rdf"
|
33 |
HF_MODEL = "Qwen/Qwen2.5-72B-Instruct"
|
|
|
141 |
max_tokens=max_tokens,
|
142 |
top_p=0.9
|
143 |
)
|
144 |
+
raw = resp["choices"][0]["message"]["content"]
|
145 |
# Forziamo la query su linea singola se multiline
|
146 |
single_line = " ".join(raw.splitlines())
|
147 |
logger.debug(f"Risposta HF single-line: {single_line}")
|
|
|
150 |
logger.error(f"HuggingFace error: {e}")
|
151 |
raise HTTPException(status_code=500, detail=str(e))
|
152 |
|
153 |
+
app = FastAPI()
|
154 |
|
155 |
class QueryRequest(BaseModel):
|
156 |
+
message: str
|
157 |
+
max_tokens: int = 1024
|
158 |
+
temperature: float = 0.5
|
159 |
|
160 |
@app.post("/generate-response/")
|
161 |
+
async def generate_response(req: QueryRequest):
|
162 |
+
user_input = req.message
|
163 |
logger.info(f"Utente dice: {user_input}")
|
164 |
|
165 |
# Recupera documenti rilevanti usando RAG
|
166 |
relevant_docs = retrieve_relevant_documents(user_input, top_k=3)
|
167 |
retrieved_text = "\n".join([doc['text'] for doc in relevant_docs])
|
168 |
|
169 |
+
sys_msg = create_system_message(knowledge_text, retrieved_text)
|
170 |
+
msgs = [
|
171 |
+
{"role": "system", "content": sys_msg},
|
172 |
+
{"role": "user", "content": user_input}
|
173 |
]
|
174 |
+
|
175 |
# Primo tentativo
|
176 |
+
r1 = await call_hf_model(msgs, req.temperature, req.max_tokens)
|
177 |
logger.info(f"PRIMA RISPOSTA:\n{r1}")
|
178 |
|
179 |
# Se non parte con "PREFIX base:"
|
180 |
if not r1.startswith("PREFIX base:"):
|
181 |
+
sc = f"Non hai risposto con query SPARQL su una sola riga. Riprova. Domanda: {user_input}"
|
182 |
+
msgs2 = [
|
183 |
+
{"role": "system", "content": sys_msg},
|
184 |
+
{"role": "assistant", "content": r1},
|
185 |
+
{"role": "user", "content": sc}
|
186 |
]
|
187 |
+
r2 = await call_hf_model(msgs2, req.temperature, req.max_tokens)
|
188 |
logger.info(f"SECONDA RISPOSTA:\n{r2}")
|
189 |
if r2.startswith("PREFIX base:"):
|
190 |
+
sparql_query = r2
|
191 |
else:
|
192 |
+
return {"type": "NATURAL", "response": r2}
|
193 |
else:
|
194 |
+
sparql_query = r1
|
195 |
|
196 |
# Esegui la query con rdflib
|
197 |
+
g = rdflib.Graph()
|
198 |
try:
|
199 |
+
g.parse(RDF_FILE, format="xml")
|
200 |
except Exception as e:
|
201 |
logger.error(f"Parsing RDF error: {e}")
|
202 |
+
return {"type": "ERROR", "response": f"Parsing RDF error: {e}"}
|
203 |
|
204 |
try:
|
205 |
+
results = g.query(sparql_query)
|
206 |
except Exception as e:
|
207 |
+
fallback = f"La query SPARQL ha fallito. Riprova. Domanda: {user_input}"
|
208 |
+
msgs3 = [
|
209 |
+
{"role": "system", "content": sys_msg},
|
210 |
+
{"role": "assistant", "content": sparql_query},
|
211 |
+
{"role": "user", "content": fallback}
|
212 |
]
|
213 |
+
r3 = await call_hf_model(msgs3, req.temperature, req.max_tokens)
|
214 |
if r3.startswith("PREFIX base:"):
|
215 |
+
sparql_query = r3
|
216 |
try:
|
217 |
+
results = g.query(sparql_query)
|
218 |
except Exception as e2:
|
219 |
+
return {"type": "ERROR", "response": f"Query fallita di nuovo: {e2}"}
|
220 |
else:
|
221 |
+
return {"type": "NATURAL", "response": r3}
|
222 |
|
223 |
+
if len(results) == 0:
|
224 |
+
return {"type": "NATURAL", "sparql_query": sparql_query, "response": "Nessun risultato."}
|
225 |
|
226 |
# Confeziona risultati
|
227 |
+
row_list = []
|
228 |
for row in results:
|
229 |
+
row_str = ", ".join([f"{k}:{v}" for k, v in row.asdict().items()])
|
230 |
row_list.append(row_str)
|
231 |
+
results_str = "\n".join(row_list)
|
232 |
|
233 |
# Spiegazione
|
234 |
+
exp_prompt = create_explanation_prompt(results_str)
|
235 |
+
msgs4 = [
|
236 |
+
{"role": "system", "content": exp_prompt},
|
237 |
+
{"role": "user", "content": ""}
|
238 |
]
|
239 |
+
explanation = await call_hf_model(msgs4, req.temperature, req.max_tokens)
|
240 |
|
241 |
return {
|
242 |
+
"type": "NATURAL",
|
243 |
+
"sparql_query": sparql_query,
|
244 |
+
"sparql_results": row_list,
|
245 |
+
"explanation": explanation
|
246 |
}
|
247 |
|
248 |
@app.get("/")
|
249 |
def home():
|
250 |
+
return {"message": "Prompt lascia libertà su come chiamare la proprietà del materiale, ma suggerisce un possibile 'materialeOpera'."}
|