Samuele E. Locatelli
54 lines
1.9 KiB
Python
54 lines
1.9 KiB
Python
# services/history_manager.py
|
|
import httpx
|
|
from config import settings
|
|
from services import redis_service
|
|
|
|
MAX_HISTORY_TURNS = 10 # ultimi turni da mantenere
|
|
SUMMARY_TRIGGER_TURNS = 20 # soglia per fare summarization
|
|
|
|
|
|
async def summarize_messages(messages):
|
|
"""
|
|
Usa LM Studio per riassumere i messaggi in forma compatta (condensed history).
|
|
"""
|
|
prompt = [
|
|
{"role": "system", "content": "Riassumi la seguente conversazione in forma di elenco puntato, mantenendo solo i fatti chiave e le informazioni rilevanti per continuare il dialogo."},
|
|
*messages
|
|
]
|
|
async with httpx.AsyncClient(timeout=settings.REQUEST_TIMEOUT) as client:
|
|
resp = await client.post(
|
|
settings.LM_STUDIO_URL,
|
|
json={"model": settings.MODEL_NAME, "messages": prompt}
|
|
)
|
|
resp.raise_for_status()
|
|
data = resp.json()
|
|
return data["choices"][0]["message"]["content"]
|
|
|
|
|
|
async def prepare_history(user_id: str, session_id: str):
|
|
"""
|
|
Recupera la history da Redis, applica windowing e summarization se necessario,
|
|
e restituisce la lista di messaggi da passare al modello.
|
|
"""
|
|
full_history = redis_service.get_chat(user_id, session_id, limit=1000)
|
|
|
|
if len(full_history) > SUMMARY_TRIGGER_TURNS:
|
|
old_messages = full_history[:-MAX_HISTORY_TURNS]
|
|
last_messages = full_history[-MAX_HISTORY_TURNS:]
|
|
|
|
summary_text = await summarize_messages(old_messages)
|
|
|
|
condensed_history = [{"role": "system", "content": f"Riassunto conversazione: {summary_text}"}]
|
|
condensed_history.extend(last_messages)
|
|
|
|
# Salva la nuova history condensata
|
|
redis_service.clear_chat(user_id, session_id)
|
|
for msg in condensed_history:
|
|
redis_service.save_chat(user_id, session_id, msg)
|
|
|
|
return condensed_history
|
|
else:
|
|
# Solo windowing
|
|
return full_history[-MAX_HISTORY_TURNS:]
|
|
|