Socrates_docker / classify_chat.py
AlessandroAmodioNGI's picture
rename sories with tales and make the call more neutral
411c910
Raw
History Blame Contribute Delete
3.4 kB
from classify_chat_helper import call_classification_llm
from classify_parameters import TRACKED_FIELDS, ALLOWED_STORY_TOPICS, TOPIC_DESCRIPTIONS
from db_user import load_user_info, save_user_info, update_country
from db_5_process_session import _load_history
_VALID_TOPICS = set(TOPIC_DESCRIPTIONS.keys())
RECENT_HISTORY_LIMIT = 5
def analyze_message(user_id: str, query: str, character_id: str = "socrates"):
# --- Load current user_info for this user ---
user_info = load_user_info(user_id)
profile = user_info.get("user_profile", {})
missing_fields = [f for f in ["name", "living_country", "origin_country"] if not profile.get(f)]
# --- Trim history to the most recent turns ---
history = _load_history("chat_history_short", user_id)
all_messages = []
for session in history.get("sessions", []):
all_messages.extend(session.get("messages", []))
recent_history = all_messages[-RECENT_HISTORY_LIMIT:] if all_messages else []
# --- Call LLM for classification + extraction ---
analysis = call_classification_llm(query, recent_history, user_info, missing_fields, character_id=character_id)
# Strip empty/blank values — classifier returns all fields, most empty
raw_extracted = analysis.get("extracted_user_info", {})
extracted_info = {
k: v for k, v in raw_extracted.items()
if v not in (None, "", [], {})
}
relevant_missing = analysis.get("relevant_missing_fields", [])
topic = analysis.get("topic", "personal")
# Guard: LLM sometimes bleeds response_mode values (e.g. "dialogic") into topic.
if topic not in _VALID_TOPICS:
topic = "personal"
response_mode = analysis.get("response_mode", "dialogic")
topic_for_tale = analysis.get("topic_for_tale", "none")
if topic_for_tale not in ALLOWED_STORY_TOPICS:
topic_for_tale = "none"
# --- Update user_info in Supabase only when genuinely new info was extracted ---
if extracted_info:
_prev_living = profile.get("living_country", "")
_prev_origin = profile.get("origin_country", "")
profile.update(extracted_info)
user_info["user_profile"] = profile
save_user_info(user_info, user_id)
# Only update countries when the LLM extracted a value that differs from what was stored
if "living_country" in extracted_info and extracted_info["living_country"] != _prev_living:
update_country("living", extracted_info["living_country"], user_id)
if "origin_country" in extracted_info and extracted_info["origin_country"] != _prev_origin:
update_country("origin", extracted_info["origin_country"], user_id)
return {
"topic": topic,
"response_mode": response_mode,
"user_info": user_info,
"relevant_missing": relevant_missing,
"topic_for_tale": topic_for_tale,
"chunks": [],
"needs_news_fetch": analysis.get("needs_news_fetch", False),
"news_topic": analysis.get("news_topic", ""),
"news_question": analysis.get("news_question", ""),
"news_temporal_context": analysis.get("news_temporal_context", ""),
"socratic_trigger": analysis.get("socratic_trigger", "none") or "none",
"trigger_subtype": analysis.get("trigger_subtype", "") or "",
"socratic_alignment": analysis.get("socratic_alignment", "none") or "none",
}