fix: Tamil to English RAG query translation for accurate retrieval

app.py

@@ -57,8 +57,42 @@ def get_lang(lang):
     if "English Only" in lang: return "Answer ONLY in English"
     return "Answer in Tamil + English mix (Tanglish)"
 
+# Tamil → English query mapping for better RAG retrieval
+TAMIL_TO_ENGLISH = {
+    "தேவை விதி": "law of demand",
+    "தேவை": "demand",
+    "விநியோகம்": "supply",
+    "விநியோக விதி": "law of supply",
+    "இரட்டை பதிவு": "double entry",
+    "இரட்டை பதிவு முறை": "double entry bookkeeping",
+    "மொத்த உள்நாட்டு உற்பத்தி": "GDP gross domestic product",
+    "தொழில் முனைவோர்": "entrepreneur entrepreneurship",
+    "நுகர்வோர் உரிமைகள்": "consumer rights",
+    "பணவீக்கம்": "inflation",
+    "வங்கி": "bank banking",
+    "வரி": "tax taxation",
+    "இலாபம்": "profit",
+    "நஷ்டம்": "loss",
+    "சொத்து": "assets",
+    "பொறுப்பு": "liability",
+    "மூலதனம்": "capital",
+}
+
+def get_rag_query(message):
+    """Convert Tamil query to English for better vector search"""
+    msg_lower = message.lower()
+    for tamil, english in TAMIL_TO_ENGLISH.items():
+        if tamil in message:
+            return english
+    # If mostly Tamil, use subject+grade as context
+    tamil_chars = sum(1 for c in message if '஀' <= c <= '௿')
+    if tamil_chars > len(message) * 0.3:
+        return message  # Let Gemini embed handle it
+    return message
+
 def generate(message, mode, grade, subject, language, story_style):
-    ctx = rag_search(message, grade=grade, subject=subject)
+    rag_query = get_rag_query(message)
+    ctx = rag_search(rag_query, grade=grade, subject=subject)
     ctx_text = f"Textbook Context:\n{ctx}" if ctx else "(Use general TN Board knowledge)"
     system = SYSTEM_PROMPT.format(language=get_lang(language))
     if mode == "📚 Chat Tutor":