verbessertes Prompt, und chat-Router optimiert

This commit is contained in:
Lars committed 2025-12-09 13:50:18 +01:00
1 parent bd44af2b68
commit 6c2074166c
2 files changed
+50 -27

No files matched your search

+35 -18
View File
@@ -1,6 +1,6 @@
"""
app/routers/chat.py — RAG Endpunkt (WP-06 Hybrid Router)
Version: 0.2.1 (Fix: System Prompt Separation)
app/routers/chat.py — RAG Endpunkt (WP-06 Hybrid Router v2)
Update: Robusteres LLM-Parsing für Small Language Models (SLMs).
"""
from fastapi import APIRouter, HTTPException, Depends
@@ -88,6 +88,11 @@ def _build_enriched_context(hits: List[QueryHit]) -> str:
return "\n\n".join(context_parts)
async def _classify_intent(query: str, llm: LLMService) -> str:
"""
Hybrid Router v2:
1. Keyword Check (Best/Longest Match) -> FAST
2. LLM Fallback (Robust Parsing) -> SMART
"""
config = get_full_config()
strategies = config.get("strategies", {})
settings = config.get("settings", {})
@@ -96,7 +101,7 @@ async def _classify_intent(query: str, llm: LLMService) -> str:
best_intent = None
max_match_length = 0
# 1. FAST PATH
# 1. FAST PATH: Keywords
for intent_name, strategy in strategies.items():
if intent_name == "FACT": continue
keywords = strategy.get("trigger_keywords", [])
@@ -110,29 +115,41 @@ async def _classify_intent(query: str, llm: LLMService) -> str:
logger.info(f"Intent detected via KEYWORD: {best_intent}")
return best_intent
# 2. SLOW PATH
# 2. SLOW PATH: LLM Router
if settings.get("llm_fallback_enabled", False):
router_prompt_template = settings.get("llm_router_prompt", "")
if router_prompt_template:
prompt = router_prompt_template.replace("{query}", query)
logger.info("Keywords failed. Asking LLM for Intent...")
# Router braucht keinen System-Prompt, nur den Classifier-Prompt
llm_decision = await llm.generate_raw_response(prompt)
# Kurzer Raw Call
raw_response = await llm.generate_raw_response(prompt)
llm_decision = llm_decision.strip().upper()
if ":" in llm_decision:
llm_decision = llm_decision.split(":")[-1].strip()
# --- Robust Parsing für SLMs ---
# Wir suchen nach den bekannten Strategie-Namen im Output
llm_output_upper = raw_response.upper()
logger.info(f"LLM Router Raw Output: '{raw_response}'") # Debugging
# Satzzeichen entfernen für sauberen Match
llm_decision = ''.join(filter(str.isalnum, llm_decision))
if llm_decision in strategies:
logger.info(f"Intent detected via LLM: {llm_decision}")
return llm_decision
found_intents = []
for strat_key in strategies.keys():
# Wir prüfen, ob der Strategie-Name (z.B. "EMPATHY") im Text vorkommt
if strat_key in llm_output_upper:
found_intents.append(strat_key)
# Entscheidung
final_intent = "FACT"
if len(found_intents) == 1:
# Eindeutiger Treffer
final_intent = found_intents[0]
logger.info(f"Intent detected via LLM (Parsed): {final_intent}")
return final_intent
elif len(found_intents) > 1:
# Mehrere Treffer (z.B. "Es ist FACT oder DECISION") -> Nimm den ersten oder Fallback
logger.warning(f"LLM returned multiple intents {found_intents}. Using first match: {found_intents[0]}")
return found_intents[0]
else:
logger.warning(f"LLM predicted unknown intent '{llm_decision}', falling back to FACT.")
logger.warning(f"LLM did not return a valid strategy name. Falling back to FACT.")
return "FACT"
@router.post("/", response_model=ChatResponse)
@@ -200,7 +217,7 @@ async def chat_endpoint(
logger.info(f"[{query_id}] Sending to LLM (Intent: {intent}, Template: {prompt_key})...")
# FIX: System-Prompt separat übergeben!
# System-Prompt separat übergeben
answer_text = await llm.generate_raw_response(prompt=final_prompt, system=system_prompt)
duration_ms = int((time.time() - start_time) * 1000)