Upload main.py

main.py

@@ -1,6 +1,7 @@
 from fastapi import FastAPI, Request
 import json
 import os
+import re
 import firebase_admin
 from firebase_admin import credentials, firestore
 from datetime import datetime
@@ -10,7 +11,7 @@ app = FastAPI()
 # --- SETUP ---
 COLLECTION_KNOWLEDGE = "knowledge_base"
 COLLECTION_RULES = "availability_rules"
-COLLECTION_INBOX = "inbox"  # NEU: Hier landen ungelöste Fragen
+COLLECTION_INBOX = "inbox"
 KNOWLEDGE_CACHE = []
 
 # --- FIREBASE VERBINDUNG ---
@@ -43,12 +44,21 @@ def reload_knowledge():
 async def startup():
     reload_knowledge()
 
+# --- HELPER: STEMMING & TOKENIZING ---
 def get_stem(word):
+    # Einfaches Stemming (Wortstamm finden)
     w = word.lower().strip()
-    for end in ["ern", "en", "er", "es", "st", "te", "e", "s", "t"]:
-        if w.endswith(end) and len(w) > len(end)+2: return w[:-len(end)]
+    suffixes = ["ungen", "innen", "keit", "sch", "ern", "en", "er", "es", "st", "te", "e", "s", "t"]
+    for end in suffixes:
+        if w.endswith(end) and len(w) > len(end) + 2: 
+            return w[:-len(end)]
     return w
 
+def tokenize(text):
+    # Zerlegt Text in saubere Wort-Liste (ohne Punkt/Komma)
+    clean_text = re.sub(r'[^\w\s]', '', text.lower())
+    return [get_stem(w) for w in clean_text.split() if w]
+
 # --- HELPER: VAPI REQUEST PARSER ---
 def parse_vapi_request(data):
     tool_call_id = "unknown"
@@ -90,7 +100,6 @@ async def check_availability(request: Request):
                 rd = r.to_dict()
                 if rd.get('start_date') <= today <= rd.get('end_date'):
                     print(f"🛑 REGEL AKTIV: {rd.get('name')}")
-                    # Einfache Logik: Wenn "Ferien" im Namen -> Limited, sonst Unavailable
                     if "ferien" in rd.get('name', '').lower():
                         status = "limited"
                     else:
@@ -105,7 +114,7 @@ async def check_availability(request: Request):
     }
 
 # ==========================================
-# TOOL 2: SUCHE (Mit Inbox-Speicherung!)
+# TOOL 2: SUCHE (KEYWORD ROUTING)
 # ==========================================
 @app.post("/search")
 async def search(request: Request):
@@ -113,47 +122,75 @@ async def search(request: Request):
     tool_call_id, args = parse_vapi_request(data)
     query = args.get("search_query") or args.get("query") or data.get("search_query")
     
-    print(f"🔎 FRAGE (ID: {tool_call_id}): '{query}'")
+    print(f"🔎 QUERY: '{query}'")
     
     answer_text = "Dazu habe ich leider keine Informationen in meiner Datenbank."
     
     if query:
-        STOP_WORDS = ["hallo", "guten", "tag", "moin", "bitte", "danke", "frage"]
-        q_words = [get_stem(w) for w in query.lower().split() if len(w)>2]
-        relevant_words = [w for w in q_words if w not in STOP_WORDS]
+        # Stop-Wörter: Diese ignorieren wir komplett, damit sie keine falschen Hits erzeugen
+        STOP_WORDS = [
+            "hallo", "guten", "tag", "moin", "bitte", "danke", "frage", "sagen", 
+            "kannst", "du", "mir", "was", "ist", "wer", "wie", "wo", "wann",
+            "anbieten", "machen", "tun", "geben", "helfen", "möchte", "will", "haben"
+        ]
+        
+        # 1. Query zerlegen
+        query_stems = [w for w in tokenize(query) if w not in STOP_WORDS and len(w) > 2]
+        print(f"🧐 Tokens: {query_stems}")
         
         found = False
         
-        if relevant_words:
+        if query_stems:
             best_doc = None
             best_score = 0
+            
             for doc in KNOWLEDGE_CACHE:
                 score = 0
-                title = doc.get("question", "").lower()
-                content = doc.get("answer", "").lower()
-                keywords = [k.lower() for k in doc.get("keywords", [])]
+                hits = []
+                
+                # Dokument zerlegen
+                title_stems = tokenize(doc.get("question", ""))
+                keyword_stems = tokenize(", ".join(doc.get("keywords", [])))
+                content_stems = tokenize(doc.get("answer", "")) # Inhalt ist jetzt unwichtig
                 
-                for word in relevant_words:
-                    if word in title: score += 50 
-                    for k in keywords:
-                        if get_stem(k) == get_stem(word): score += 30
-                    if word in content: score += 5
+                for q_stem in query_stems:
+                    
+                    # 1. KEYWORD MATCH (Der "Router") -> 100 Punkte!
+                    if q_stem in keyword_stems:
+                        score += 100
+                        hits.append(f"KEYWORD '{q_stem}'")
+                        
+                    # 2. TITEL MATCH -> 50 Punkte
+                    elif q_stem in title_stems:
+                        score += 50
+                        hits.append(f"TITLE '{q_stem}'")
+                        
+                    # 3. CONTENT MATCH -> Nur 5 Punkte (reicht alleine nicht)
+                    elif q_stem in content_stems:
+                        score += 5
+                        # hits.append("content") # Loggen wir nicht, spammt sonst
 
                 if score > best_score:
                     best_score = score
                     best_doc = doc
+                    if score > 10:
+                        print(f"   Kandidat: {score} Pkt ({hits}) -> {doc.get('question')[:40]}...")
 
-            # SCHWELLE: 20 PUNKTE
-            if best_doc and best_score >= 20:
-                print(f"🏆 TREFFER ({best_score}): {best_doc.get('question')}")
+            # SCHWELLE: 40 PUNKTE
+            # Ein Keyword-Treffer (100) gewinnt immer.
+            # Ein Titel-Treffer (50) gewinnt immer.
+            # Nur Fließtext (5 Pkt pro Wort) braucht mind. 8 Treffer -> Unwahrscheinlich bei falschen Dokus.
+            
+            if best_doc and best_score >= 40:
+                print(f"🏆 GEWINNER ({best_score} Pkt): {best_doc.get('question')}")
                 answer_text = best_doc.get("answer")
                 found = True
             else:
-                 print(f"⚠️ Zu wenig Relevanz (Max: {best_score})")
+                 print(f"⚠️ Kein eindeutiger Treffer (Max Score: {best_score})")
         
-        # --- NEU: SPEICHERN WENN NICHT GEFUNDEN ---
+        # --- INBOX LOGIK ---
         if not found and db:
-            print("📥 Speichere in Inbox...")
+            print("📥 Ab in die Inbox.")
             db.collection(COLLECTION_INBOX).add({
                 "query": query,
                 "timestamp": datetime.now(),