Upload 2 files

rag_pipeline.py

@@ -0,0 +1,281 @@
+"""
+RAG Pipeline for Hindi Banking Voice Assistant
+Uses: Whisper (STT) + Indic-Gemma (LLM) + Indic-Parler-TTS (TTS)
+All retrieval in English, all responses in Hindi.
+"""
+
+import re
+import numpy as np
+from typing import Optional
+import torch
+from knowledge_base import get_document_texts, KNOWLEDGE_BASE
+
+
+# ─────────────────────────────────────────────
+# FINANCE JARGON NORMALIZATION MAP
+# Hindi/Hinglish → English canonical terms
+# ─────────────────────────────────────────────
+JARGON_MAP = {
+    # EMI variants
+    "kist": "EMI equated monthly installment",
+    "maahik kist": "monthly installment EMI",
+    "maasik bhugtan": "monthly payment EMI",
+    "kisten": "installments EMI",
+    "maahik bhugtan": "monthly installment EMI",
+
+    # Loan
+    "karz": "loan",
+    "udhaar": "loan credit",
+    "rin": "loan",
+    "loan lena": "apply for loan",
+    "loan milega": "loan eligibility",
+    "paise chahiye": "need money loan",
+    "paisa": "money funds",
+    "raqam": "amount loan",
+
+    # Interest
+    "byaj": "interest rate",
+    "sudh": "interest",
+    "byaj dar": "interest rate",
+    "faixed byaj": "fixed interest rate",
+    "badlav wala byaj": "floating interest rate",
+
+    # Bank account
+    "khata": "bank account",
+    "bachat khata": "savings account",
+    "khata kholna": "open bank account",
+    "bank mein khata": "bank account",
+
+    # Collateral/Guarantee
+    "zamanat": "collateral guarantee security",
+    "zamanatdar": "guarantor",
+    "girwi": "mortgage pledge",
+    "girvi rakhna": "pledge collateral",
+
+    # Documents
+    "kaagaz": "documents",
+    "dastavej": "documents",
+    "pehchaan patra": "identity proof",
+    "niwas praman": "address proof",
+    "aay praman": "income proof",
+
+    # Credit/CIBIL
+    "saakh": "credit score CIBIL",
+    "credit score kya hai": "what is credit score CIBIL",
+    "score": "CIBIL credit score",
+
+    # Principal/Tenure
+    "mool rashi": "principal amount",
+    "avadhi": "loan tenure duration",
+    "muddat": "loan tenure period",
+    "kitne saal": "how many years tenure",
+    "kitne mahine": "how many months tenure",
+
+    # Repayment
+    "wapasi": "repayment",
+    "bhugtan": "payment repayment",
+    "chukana": "repay loan",
+    "ada karna": "pay repay",
+
+    # Government schemes
+    "sarkar ki yojana": "government scheme",
+    "yojana": "scheme",
+    "sarkari loan": "government loan scheme",
+    "subsidy": "subsidy government benefit",
+    "anudan": "grant subsidy",
+
+    # Specific schemes
+    "mudra": "mudra loan PMMY",
+    "kisaan": "farmer kisan",
+    "kisan": "farmer kisan credit card",
+    "jan dhan": "PMJDY jan dhan account",
+    "bima": "insurance",
+    "jeevan bima": "life insurance PMJJBY",
+    "suraksha bima": "accident insurance PMSBY",
+    "pension": "pension APY Atal Pension Yojana",
+    "gramin bank": "rural bank RRB",
+    "shg": "self help group SHG women loan",
+    "samuh": "self help group SHG",
+    "mahila samuh": "women self help group SHG microfinance",
+
+    # Defaults/issues
+    "default": "loan default NPA",
+    "band ho gaya": "account closed loan default",
+    "paise nahin de paya": "unable to repay loan default",
+    "chhoot": "waiver loan waiver",
+
+    # Property
+    "ghar lena": "home purchase home loan",
+    "makan": "house home property",
+    "zameen": "land property",
+    "ghar banana": "home construction loan",
+    "flat": "apartment home loan",
+
+    # Grievance
+    "shikayat": "complaint grievance",
+    "problem": "complaint issue grievance",
+    "dhoka": "fraud complaint",
+    "pareshan": "problem issue complaint",
+}
+
+
+def normalize_jargon(text: str) -> str:
+    """
+    Replace Hindi/Hinglish finance jargon with standardized English terms
+    to improve retrieval quality.
+    """
+    text_lower = text.lower()
+    for hindi_term, english_term in JARGON_MAP.items():
+        if hindi_term in text_lower:
+            text_lower = text_lower.replace(hindi_term, english_term)
+    return text_lower
+
+
+def translate_to_retrieval_query(hindi_text: str) -> str:
+    """
+    Convert normalized Hindi/Hinglish text to an English retrieval query.
+    This uses rule-based translation for common patterns.
+    For production: could use a translation model.
+    """
+    normalized = normalize_jargon(hindi_text)
+
+    # Extract key English terms that emerged from normalization
+    # plus keep any already-English words in the original
+    english_words = []
+
+    # Words already in English or normalized
+    for word in normalized.split():
+        if any(c.isalpha() for c in word):
+            english_words.append(word)
+
+    return " ".join(english_words[:20])  # cap at 20 words for retrieval
+
+
+# ─────────────────────────────────────────────
+# SIMPLE BM25-STYLE KEYWORD RETRIEVAL
+# (No heavy dependencies — pure Python TF-IDF-like scoring)
+# ─────────────────────────────────────────────
+class SimpleRetriever:
+    def __init__(self):
+        self.documents = []
+        self.doc_ids = []
+        self.doc_words = []
+        self._build_index()
+
+    def _build_index(self):
+        """Build keyword index over knowledge base."""
+        for doc_id, text in get_document_texts():
+            self.doc_ids.append(doc_id)
+            self.documents.append(text.lower())
+            words = set(re.findall(r'\b\w+\b', text.lower()))
+            self.doc_words.append(words)
+
+        # Also index tags for better matching
+        for i, doc in enumerate(KNOWLEDGE_BASE):
+            tags_text = " ".join(doc.get("tags", []))
+            self.documents[i] += " " + tags_text.lower()
+            self.doc_words[i].update(re.findall(r'\b\w+\b', tags_text.lower()))
+
+    def retrieve(self, query: str, top_k: int = 3) -> list[dict]:
+        """Retrieve top_k most relevant documents for query."""
+        query_words = set(re.findall(r'\b\w+\b', query.lower()))
+
+        if not query_words:
+            return []
+
+        scores = []
+        for i, doc_words in enumerate(self.doc_words):
+            # Jaccard-like overlap + length normalization
+            overlap = len(query_words & doc_words)
+            score = overlap / (len(query_words) + 0.5)
+
+            # Bonus for exact phrase matches
+            for qw in query_words:
+                if len(qw) > 4 and qw in self.documents[i]:
+                    score += 0.3
+
+            scores.append((score, i))
+
+        scores.sort(reverse=True)
+        top_indices = [idx for score, idx in scores[:top_k] if score > 0]
+
+        results = []
+        for idx in top_indices:
+            doc = KNOWLEDGE_BASE[idx]
+            results.append({
+                "id": doc["id"],
+                "title": doc["title"],
+                "content": doc["content"],
+                "category": doc["category"],
+                "score": scores[idx][0] if idx < len(scores) else 0,
+            })
+
+        return results
+
+
+# Singleton retriever
+_retriever = None
+
+def get_retriever() -> SimpleRetriever:
+    global _retriever
+    if _retriever is None:
+        _retriever = SimpleRetriever()
+    return _retriever
+
+
+def build_rag_prompt(user_question: str, retrieved_docs: list[dict]) -> str:
+    """
+    Build the prompt for the Indic-Gemma LLM.
+    System prompt enforces Hindi, grounded answers, brevity.
+    """
+    context_parts = []
+    for i, doc in enumerate(retrieved_docs, 1):
+        context_parts.append(f"[{i}] {doc['title']}\n{doc['content'][:600]}")
+
+    context = "\n\n".join(context_parts) if context_parts else "कोई प्रासंगिक जानकारी नहीं मिली।"
+
+    prompt = f"""<|system|>
+आप एक सहायक बैंकिंग सहायक हैं जो भारतीय बैंकिंग, लोन, और सरकारी योजनाओं के बारे में सरल हिंदी में जानकारी देते हैं।
+
+नियम:
+1. केवल नीचे दी गई जानकारी के आधार पर उत्तर दें। अनुमान न लगाएं।
+2. उत्तर छोटा, सरल और बोलने योग्य हो — 3-4 वाक्यों में।
+3. यदि जानकारी उपलब्ध नहीं है, तो कहें: "यह जानकारी मेरे पास नहीं है। कृपया अपने बैंक से संपर्क करें।"
+4. अंत में केवल एक जरूरी follow-up प्रश्न पूछें (यदि आवश्यक हो)।
+5. हमेशा हिंदी में उत्तर दें।
+
+संदर्भ जानकारी:
+{context}
+<|end|>
+<|user|>
+{user_question}
+<|end|>
+<|assistant|>"""
+
+    return prompt
+
+
+def format_response_for_tts(text: str) -> str:
+    """
+    Clean LLM output for TTS: remove special chars, ensure clean Hindi.
+    """
+    # Remove markdown
+    text = re.sub(r'\*+', '', text)
+    text = re.sub(r'#+\s*', '', text)
+    text = re.sub(r'\[[\d]+\]', '', text)  # remove citation markers
+    text = re.sub(r'\n+', ' ', text)
+    text = re.sub(r'\s+', ' ', text)
+    text = text.strip()
+    return text
+
+
+def get_tts_description(text: str) -> str:
+    """
+    Parler-TTS requires a description of the speaker/style.
+    Returns appropriate description for Hindi banking assistant.
+    """
+    return (
+        "A calm, clear female voice speaking in Hindi. "
+        "The speech is measured and helpful, like a bank customer service representative. "
+        "Very clear pronunciation, moderate pace, friendly tone."
+    )

requirements.txt

@@ -0,0 +1,6 @@
+# CPU Basic safe — all models via HF Inference API, nothing loads locally
+gradio>=4.44.0
+requests>=2.31.0
+soundfile>=0.12.1
+numpy>=1.24.0
+# NO torch, NO transformers, NO parler-tts — all inference is API-based