please

README.md

@@ -4,8 +4,8 @@ emoji: 🏦
 colorFrom: yellow
 colorTo: green
 sdk: gradio
-sdk_version: 5.16.1
-python_version: 3.11
+sdk_version: 5.9.1
+python: 13.1
 app_file: app.py
 pinned: true
 license: mit

app.py

@@ -1,88 +1,101 @@
-# STARTUP:
-# 1. pip install -r requirements.txt
-# 2. python app.py      ← start the app (auto-builds index)
+"""
+Hindi Banking Voice Assistant
+CPU Basic safe — uses HF InferenceClient with providers that actually work.
+
+STT  : openai/whisper-large-v3        (hf-inference, ASR)
+LLM  : mistralai/Mistral-7B-Instruct-v0.3  (featherless provider, free tier)
+TTS  : facebook/mms-tts-hin           (hf-inference, Hindi TTS, small model)
+"""
 
 import os
 import io
 import numpy as np
 import soundfile as sf
 import gradio as gr
-
-import stt
-import llm
-import tts
-import rag_pipeline
+from huggingface_hub import InferenceClient
+
+from rag_pipeline import (
+    normalize_jargon,
+    translate_to_retrieval_query,
+    build_rag_prompt,
+    format_response_for_tts,
+    get_retriever,
+)
 
 # ─────────────────────────────────────────────
-# AUTO INGEST — builds ChromaDB on first startup
-# Runs automatically if chroma_db folder not found
+# CONFIG
 # ─────────────────────────────────────────────
-import os
-if not os.path.exists("./chroma_db"):
-    print("ChromaDB not found — building knowledge base index...")
-    try:
-        from knowledge_base import KNOWLEDGE_BASE
-        from sentence_transformers import SentenceTransformer
-        import chromadb
+HF_TOKEN = os.environ.get("HF_TOKEN", "")
 
-        print("Loading embedding model for ingest...")
-        _embedder = SentenceTransformer(
-            'sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2'
-        )
-        _chroma = chromadb.PersistentClient(path="./chroma_db")
-        _collection = _chroma.get_or_create_collection("banking_hindi")
-
-        _documents = []
-        _metadatas = []
-        _ids = []
-
-        for doc in KNOWLEDGE_BASE:
-            full_text = f"{doc['title']}\n{doc['content']}"
-            _documents.append(full_text)
-            _metadatas.append({
-                "id": doc["id"],
-                "title": doc["title"],
-                "category": doc["category"]
-            })
-            _ids.append(doc["id"])
-
-        print(f"Embedding {len(_documents)} documents... (takes 2-3 min on first run)")
-        _embeddings = _embedder.encode(
-            _documents,
-            show_progress_bar=True,
-            batch_size=8
-        ).tolist()
-
-        _collection.add(
-            documents=_documents,
-            embeddings=_embeddings,
-            metadatas=_metadatas,
-            ids=_ids
-        )
-        print(f"✅ ChromaDB ready — {len(_documents)} documents indexed")
-        
-        # Cleanup temp variables
-        del _embedder, _chroma, _collection
-        del _documents, _metadatas, _ids, _embeddings
+STT_MODEL = "openai/whisper-large-v3"
+LLM_MODEL = "mistralai/Mistral-7B-Instruct-v0.3"
+TTS_MODEL = "facebook/mms-tts-hin"
+
+retriever = get_retriever()
+print(f"Knowledge base ready: {len(retriever.doc_ids)} documents")
 
-    except Exception as e:
-        print(f"⚠️ Auto-ingest failed: {e}. Will use keyword retrieval fallback.")
-else:
-    print("✅ ChromaDB found — skipping ingest")
 
 # ─────────────────────────────────────────────
-# CONFIG
+# INFERENCE FUNCTIONS
 # ─────────────────────────────────────────────
-HF_TOKEN = (
-    os.environ.get("HF_TOKEN") or
-    os.environ.get("HF_API_TOKEN") or
-    os.environ.get("HUGGINGFACE_TOKEN") or
-    ""
-)
-print(f"DEBUG app.py: HF_TOKEN loaded = {bool(HF_TOKEN)}, length = {len(HF_TOKEN)}")
-
 
-# (STT, LLM, and TTS functions moved to separate modules)
+def get_client(provider: str) -> InferenceClient:
+    return InferenceClient(provider=provider, api_key=HF_TOKEN)
+
+
+def stt_whisper(audio_array: np.ndarray, sample_rate: int) -> str:
+    """Convert Hindi audio to text using Whisper via hf-inference."""
+    buf = io.BytesIO()
+    sf.write(buf, audio_array, sample_rate, format="WAV", subtype="PCM_16")
+    buf.seek(0)
+
+    client = get_client("hf-inference")
+    result = client.automatic_speech_recognition(
+        audio=buf.read(),
+        model=STT_MODEL,
+    )
+    # result is an ASROutput object with .text
+    return result.text.strip() if hasattr(result, "text") else str(result).strip()
+
+
+def llm_generate(prompt: str) -> str:
+    """Generate Hindi answer using Mistral via featherless (free tier)."""
+    client = get_client("featherless")
+
+    # Build messages in chat format
+    messages = [
+        {
+            "role": "system",
+            "content": (
+                "आप एक सहायक बैंकिंग सहायक हैं। केवल दी गई जानकारी के आधार पर "
+                "सरल हिंदी में 3-4 वाक्यों मे��� उत्तर दें। "
+                "यदि जानकारी नहीं है तो कहें: 'यह जानकारी मेरे पास नहीं है। "
+                "कृपया अपने बैंक से संपर्क करें।' हमेशा हिंदी में उत्तर दें।"
+            ),
+        },
+        {"role": "user", "content": prompt},
+    ]
+
+    response = client.chat.completions.create(
+        model=LLM_MODEL,
+        messages=messages,
+        max_tokens=350,
+        temperature=0.7,
+    )
+    return response.choices[0].message.content.strip()
+
+
+def tts_hindi(text: str):
+    """Convert Hindi text to speech using MMS-TTS-HIN via hf-inference."""
+    client = get_client("hf-inference")
+    try:
+        audio_bytes = client.text_to_speech(text=text, model=TTS_MODEL)
+        buf = io.BytesIO(audio_bytes)
+        audio_array, sample_rate = sf.read(buf)
+        return int(sample_rate), audio_array.astype(np.float32)
+    except Exception as e:
+        print(f"TTS error: {e}")
+        return None
 
 
 # ─────────────────────────────────────────────
@@ -90,98 +103,108 @@ print(f"DEBUG app.py: HF_TOKEN loaded = {bool(HF_TOKEN)}, length = {len(HF_TOKEN
 # ─────────────────────────────────────────────
 
 def run_voice_pipeline(audio_input):
-    """Voice mode: mic audio → transcript, answer, citations, audio, status."""
+    """Voice mode: mic audio → transcript, answer, citations, audio."""
     if audio_input is None:
-        yield "", "कृपया माइक्रोफोन बटन दबाकर अपना प्रश्न पूछें।", "", None, "🔴 *तैयार*"
+        yield "", "कृपया माइक्रोफोन बटन दबाकर अपना प्रश्न पूछें।", "", None
         return
     if not HF_TOKEN:
-        yield "", "⚠️ HF_TOKEN Secret नहीं मिला। Space Settings → Secrets में जोड़ें।", "", None, "❌ *त्रुटि*"
+        yield "", "⚠️ HF_TOKEN Secret नहीं मिला। Space Settings → Secrets में जोड़ें।", "", None
         return
 
     try:
-        yield "", "", "", None, "🎙️ **सुन रहा हूँ...** (STT)"
         sample_rate, audio_array = audio_input
-        
-        # Ensure STT gets proper data
-        transcript = stt.stt_whisper(audio_array, sample_rate)
-        
-        if not transcript.strip():
-            yield "", "आवाज़ समझ नहीं आई। कृपया फिर से बोलें।", "", None, "⚠️ *फिर से प्रयास करें*"
+        audio_float = audio_array.astype(np.float32)
+        if np.abs(audio_float).max() > 1.0:
+            audio_float /= 32768.0
+        if audio_float.ndim > 1:
+            audio_float = audio_float.mean(axis=1)
+
+        yield "⏳ आवाज़ पहचाना जा रहा है...", "", "", None
+
+        hindi_text = stt_whisper(audio_float, sample_rate)
+        if not hindi_text:
+            yield "", "आवाज़ स्पष्ट नहीं सुनाई दी। कृपया दोबारा कोशिश करें।", "", None
             return
-            
-        yield transcript, "", "", None, "🔎 **जानकारी खोज रहा हूँ...** (RAG)"
-        normalized = rag_pipeline.normalize_jargon(transcript)
-        eng_query  = rag_pipeline.translate_to_retrieval_query(normalized)
-        context_docs = rag_pipeline.retrieve_with_embeddings(eng_query, top_k=3)
-        
-        citations = "\n".join([f"• {doc['title']}" for doc in context_docs]) if context_docs else ""
-        
-        yield transcript, "", "", None, "🤖 **उत्तर तैयार कर रहा हूँ...** (LLM)"
-        
+
+        yield hindi_text, "⏳ जानकारी खोजी जा रही है...", "", None
+
+        normalized = normalize_jargon(hindi_text)
+        eng_query  = translate_to_retrieval_query(normalized)
+        docs       = retriever.retrieve(eng_query, top_k=3) or retriever.retrieve(hindi_text, top_k=2)
+        citations  = "\n".join(f"• {d['title']}" for d in docs) if docs else "कोई स्रोत नहीं मिला।"
+
+        yield hindi_text, "⏳ उत्तर तैयार किया जा रहा है...", citations, None
+
+        # Build a concise context-augmented prompt for the chat model
         context = "\n\n".join(
             f"[{i+1}] {d['title']}\n{d['content'][:500]}"
-            for i, d in enumerate(context_docs)
-        ) if context_docs else "कोई प्रासंगिक जानकारी नहीं मिली।"
+            for i, d in enumerate(docs)
+        ) if docs else "कोई प्रासंगिक जानकारी नहीं मिली।"
 
         full_prompt = (
             f"नीचे दी गई जानकारी के आधार पर प्रश्न का उत्तर दें:\n\n"
             f"{context}\n\n"
-            f"प्रश्न: {transcript}"
+            f"प्रश्न: {hindi_text}"
         )
-        answer = llm.llm_generate(full_prompt)
-        
-        yield transcript, answer, citations, None, "🔊 **आवाज़ बना रहा हूँ...** (TTS)"
-        clean_text = rag_pipeline.format_response_for_tts(answer)
-        audio_out = tts.tts_hindi(clean_text)
-        
-        yield transcript, answer, citations, audio_out, "✅ **पूरा हुआ**"
-        
+
+        answer = llm_generate(full_prompt)
+        clean_answer = format_response_for_tts(answer) or \
+            "यह जानकारी मेरे पास नहीं है। कृपया अपने बैंक से संपर्क करें।"
+
+        yield hindi_text, clean_answer, citations, None
+
+        yield hindi_text, clean_answer, "⏳ ऑडियो बनाया जा रहा है...", None
+        audio_out = tts_hindi(clean_answer)
+        yield hindi_text, clean_answer, citations, audio_out
+
     except Exception as e:
         import traceback; traceback.print_exc()
-        print(f"Pipeline error: {e}")
-        yield "", f"⚠️ त्रुटि: {e}", "", None, "❌ *विफल*"
+        yield "", f"⚠️ त्रुटि: {e}", "", None
 
 
-def run_text_pipeline(text_input):
-    """Text mode: typed question → answer, citations, audio, status."""
+def run_text_pipeline(text_input: str):
+    """Text mode: typed question → answer, citations, audio."""
     if not text_input or not text_input.strip():
-        yield text_input, "कृपया एक प्रश्न लिखें।", "", None, "🔴 *तैयार*"
+        yield text_input, "कृपया एक प्रश्न लिखें।", "", None
         return
     if not HF_TOKEN:
-        yield text_input, "⚠️ HF_TOKEN Secret नहीं मिला।", "", None, "❌ *त्रुटि*"
+        yield text_input, "⚠️ HF_TOKEN Secret नहीं मिला।", "", None
         return
 
     try:
-        yield text_input, "", "", None, "🔎 **जानकारी खोज रहा हूँ...** (RAG)"
-        normalized = rag_pipeline.normalize_jargon(text_input)
-        eng_query  = rag_pipeline.translate_to_retrieval_query(normalized)
-        context_docs = rag_pipeline.retrieve_with_embeddings(eng_query, top_k=3)
-        
-        citations = "\n".join([f"• {doc['title']}" for doc in context_docs]) if context_docs else ""
-        
-        yield text_input, "", "", None, "🤖 **उत्तर तैयार कर रहा हूँ...** (LLM)"
-        
+        yield text_input, "⏳ जानकारी खोजी जा रही है...", "", None
+
+        normalized = normalize_jargon(text_input)
+        eng_query  = translate_to_retrieval_query(normalized)
+        docs       = retriever.retrieve(eng_query, top_k=3) or retriever.retrieve(text_input, top_k=2)
+        citations  = "\n".join(f"• {d['title']}" for d in docs) if docs else "कोई स्रोत नहीं मिला।"
+
+        yield text_input, "⏳ उत्तर तैयार किया जा रहा है...", citations, None
+
         context = "\n\n".join(
             f"[{i+1}] {d['title']}\n{d['content'][:500]}"
-            for i, d in enumerate(context_docs)
-        ) if context_docs else "कोई प्रासंगिक जानकारी नहीं मिली।"
+            for i, d in enumerate(docs)
+        ) if docs else "कोई प्रासंगिक जानकारी नहीं मिली।"
 
         full_prompt = (
             f"नीचे दी गई जानकारी के आधार पर प्रश्न का उत्तर दें:\n\n"
             f"{context}\n\n"
             f"प्रश्न: {text_input}"
         )
-        answer = llm.llm_generate(full_prompt)
-        
-        yield text_input, answer, citations, None, "🔊 **आवाज़ बना रहा हूँ...** (TTS)"
-        clean_text = rag_pipeline.format_response_for_tts(answer)
-        audio_out = tts.tts_hindi(clean_text)
-        
-        yield text_input, answer, citations, audio_out, "✅ **पूरा हुआ**"
-        
+
+        answer = llm_generate(full_prompt)
+        clean_answer = format_response_for_tts(answer) or \
+            "यह जानकारी मेरे पास नहीं है। कृपया अपने बैंक से संपर्क करें।"
+
+        yield text_input, clean_answer, citations, None
+
+        yield text_input, clean_answer, "⏳ ऑडियो बन��या जा रहा है...", None
+        audio_out = tts_hindi(clean_answer)
+        yield text_input, clean_answer, citations, audio_out
+
     except Exception as e:
         import traceback; traceback.print_exc()
-        yield text_input, f"⚠️ त्रुटि: {e}", "", None, "❌ *विफल*"
+        yield text_input, f"⚠️ त्रुटि: {e}", "", None
 
 
 # ─────────────────────────────────────────────
@@ -316,9 +339,8 @@ def build_ui():
                         v_audio_out = gr.Audio(label="🔊 उत्तर सुनें", type="numpy", autoplay=True)
                         v_citations = gr.Textbox(label="📚 स्रोत", interactive=False, lines=3, elem_classes="sources-area")
 
-                v_status = gr.Markdown("🔴 *तैयार*", elem_classes="status-bar")
                 v_btn.click(fn=run_voice_pipeline, inputs=[v_audio_in],
-                            outputs=[v_transcript, v_answer, v_citations, v_audio_out, v_status])
+                            outputs=[v_transcript, v_answer, v_citations, v_audio_out])
 
             with gr.Tab("⌨️ Text Mode"):
                 with gr.Row():
@@ -332,11 +354,10 @@ def build_ui():
                         t_audio_out = gr.Audio(label="🔊 उत्तर सुनें", type="numpy", autoplay=True)
                         t_citations = gr.Textbox(label="📚 स्रोत", interactive=False, lines=3, elem_classes="sources-area")
 
-                t_status = gr.Markdown("🔴 *तैयार*", elem_classes="status-bar")
                 t_btn.click(fn=run_text_pipeline, inputs=[t_input],
-                            outputs=[t_transcript, t_answer, t_citations, t_audio_out, t_status])
+                            outputs=[t_transcript, t_answer, t_citations, t_audio_out])
                 t_input.submit(fn=run_text_pipeline, inputs=[t_input],
-                               outputs=[t_transcript, t_answer, t_citations, t_audio_out, t_status])
+                               outputs=[t_transcript, t_answer, t_citations, t_audio_out])
 
         gr.HTML('<div style="padding:6px 10px 0;"><p style="color:var(--gold);font-size:11px;font-weight:600;letter-spacing:1px;text-transform:uppercase;margin:8px 0 4px 32px;">📌 उदाहरण प्रश्न</p></div>')
         gr.Examples(examples=[[q] for q in EXAMPLES], inputs=[t_input], label="", cache_examples=False)
@@ -378,4 +399,4 @@ def build_ui():
 
 if __name__ == "__main__":
     demo = build_ui()
-    demo.queue().launch(show_api=False)
+    demo.launch(server_name="0.0.0.0", server_port=7860, show_error=True)

ingest.py

@@ -1,55 +0,0 @@
-"""
-One-time script to build ChromaDB vector store from knowledge base.
-Run once before starting the app: python ingest.py
-Embeds all 37 knowledge base documents using multilingual MiniLM.
-"""
-from knowledge_base import KNOWLEDGE_BASE
-from sentence_transformers import SentenceTransformer
-import chromadb
-
-print("Loading embedding model...")
-embedder = SentenceTransformer(
-    'sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2'
-)
-
-print("Setting up ChromaDB...")
-chroma_client = chromadb.PersistentClient(path="./chroma_db")
-
-# Delete existing collection to rebuild fresh
-try:
-    chroma_client.delete_collection("banking_hindi")
-    print("Deleted existing collection.")
-except:
-    pass
-
-collection = chroma_client.get_or_create_collection("banking_hindi")
-
-documents = []
-metadatas = []
-ids = []
-
-for doc in KNOWLEDGE_BASE:
-    full_text = f"{doc['title']}\n{doc['content']}"
-    documents.append(full_text)
-    metadatas.append({
-        "id": doc["id"],
-        "title": doc["title"],
-        "category": doc["category"]
-    })
-    ids.append(doc["id"])
-
-print(f"Embedding {len(documents)} documents...")
-embeddings = embedder.encode(
-    documents, 
-    show_progress_bar=True,
-    batch_size=8
-).tolist()
-
-collection.add(
-    documents=documents,
-    embeddings=embeddings,
-    metadatas=metadatas,
-    ids=ids
-)
-print(f"✅ Successfully ingested {len(documents)} documents into ChromaDB")
-print(f"Collection count: {collection.count()}")

llm.py

@@ -1,95 +0,0 @@
-"""
-LLM Module - Llama 3.1 70B via HuggingFace Router
-Uses router.huggingface.co OpenAI-compatible endpoint.
-Same pattern as verified working production code.
-Includes retry logic for model cold starts.
-"""
-import requests
-import time
-import os
-
-HF_TOKEN = (
-    os.environ.get("HF_TOKEN") or
-    os.environ.get("HF_API_TOKEN") or
-    os.environ.get("HUGGINGFACE_TOKEN") or
-    ""
-)
-print(f"DEBUG LLM: HF_TOKEN loaded = {bool(HF_TOKEN)}")
-
-LLM_ROUTER_URL = "https://router.huggingface.co/v1/chat/completions"
-LLM_MODEL = "meta-llama/Llama-3.1-70B-Instruct"
-
-SYSTEM_PROMPT = """आप एक भारतीय बैंकिंग सहायक हैं।
-आपका काम है ग्रामीण और शहरी भारतीय उपयोगकर्ताओं को 
-बैंकिंग और वित्तीय जानकारी सरल हिंदी में देना।
-
-सख्त नियम:
-1. केवल हिंदी में उत्तर दें — अंग्रेजी बिल्कुल नहीं
-2. केवल दिए गए संदर्भ से उत्तर दें
-3. यदि संदर्भ में जानकारी नहीं है तो कहें: 
-   यह जानकारी मेरे पास नहीं है। कृपया अपने बैंक से संपर्क करें।
-4. ब्याज दर, EMI, या कोई भी राशि केवल संदर्भ से बताएं
-5. उत्तर 3-4 वाक्यों में दें — TTS के लिए छोटा रखें
-6. सरल भाषा — जैसे किसी गांव के व्यक्ति को समझाना हो"""
-
-def llm_generate(prompt: str) -> str:
-    """
-    Generate Hindi answer using Llama 3.1 70B.
-    Keeps same function signature as existing llm_generate() in app.py.
-    """
-    if not HF_TOKEN:
-        return ("HF_TOKEN नहीं मिला। "
-                "Space Settings → Secrets में HF_TOKEN जोड़ें।")
-    
-    headers = {"Authorization": f"Bearer {HF_TOKEN}"}
-    payload = {
-        "model": LLM_MODEL,
-        "messages": [
-            {"role": "system", "content": SYSTEM_PROMPT},
-            {"role": "user", "content": prompt}
-        ],
-        "max_tokens": 350,
-        "temperature": 0.3
-    }
-    
-    for i in range(3):
-        try:
-            res = requests.post(
-                LLM_ROUTER_URL,
-                headers=headers,
-                json=payload,
-                timeout=45
-            )
-            print(f"DEBUG LLM status: {res.status_code}")
-            
-            if res.status_code != 200:
-                print(f"DEBUG LLM error body: {res.text[:300]}")
-            
-            if not res.text.strip():
-                print(f"Empty response, retry {i+1}...")
-                time.sleep(5)
-                continue
-            
-            result = res.json()
-            
-            if isinstance(result, dict) and "choices" in result:
-                answer = result["choices"][0]["message"]["content"].strip()
-                print(f"DEBUG LLM answer preview: {answer[:100]}")
-                return answer
-            
-            if isinstance(result, dict) and "error" in result:
-                err = result.get("error", "")
-                if isinstance(err, dict):
-                    err = err.get("message", str(err))
-                if "loading" in str(err).lower():
-                    print(f"Model loading, retry {i+1} in 10s...")
-                    time.sleep(10)
-                    continue
-                print(f"LLM API error: {err}")
-                break
-                
-        except Exception as e:
-            print(f"LLM exception retry {i+1}: {e}")
-            time.sleep(2)
-    
-    return "माफ करें, अभी उत्तर देने में समस्या हो रही है। कृपया दोबारा प्रयास करें।"

rag_pipeline.py

@@ -106,8 +106,6 @@ JARGON_MAP = {
     "zameen": "land property",
     "ghar banana": "home construction loan",
     "flat": "apartment home loan",
-    "awas yojana": "awas yojana housing scheme pmay subsidy",
-    "awas": "housing pmay",
 
     # Grievance
     "shikayat": "complaint grievance",
@@ -127,15 +125,9 @@ def normalize_jargon(text: str) -> str:
 
 
 def translate_to_retrieval_query(normalized_text: str) -> str:
-    """Extract English words and key Hindi terms from normalized text for retrieval."""
-    # Keep English words (alpha) and common Hindi keywords that are in KB tags
-    words = [str(w) for w in normalized_text.split() if any(c.isalpha() for c in w)]
-    if not words:
-        # Fallback to the original text if no normalization happened
-        return str(normalized_text)
-    # Standard slicing for list of strings
-    result_words = words[0:20]
-    return " ".join(result_words)
+    """Extract English words from normalized text for retrieval."""
+    words = [w for w in normalized_text.split() if any(c.isalpha() for c in w)]
+    return " ".join(words[:20])
 
 
 # ─────────────────────────────────────────────
@@ -171,15 +163,10 @@ class SimpleRetriever:
         for i, doc_words in enumerate(self.doc_words):
             overlap = len(query_words & doc_words)
             score = overlap / (len(query_words) + 0.5)
-            
-            # Substantial bonus for exact phrase matching in document
-            if query.lower() in self.documents[i]:
-                score += 1.0
-            
             # Bonus for longer exact word matches
             for qw in query_words:
-                if len(qw) > 3 and qw in self.documents[i]:
-                    score += 0.2
+                if len(qw) > 4 and qw in self.documents[i]:
+                    score += 0.3
             scores.append((score, i))
 
         scores.sort(reverse=True)
@@ -225,10 +212,10 @@ def build_rag_prompt(user_question: str, retrieved_docs: list) -> str:
     return 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. नीचे दी गई जानकारी का उपयोग करें। यदि जानकारी बिल्कुल स्पष्ट नहीं है, तो सामान्य बैंकिंग ज्ञान का उपयोग कर सहायता करें लेकिन "यह जानकारी मेरे पास नहीं है" कहने से बचें यदि आप बेसिक सलाह दे सकते 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 बिल्कुल भी उपलब्ध नहीं है, तभी कहें: "यह जानकारी मेरे पास नहीं है। कृपया अपने बैंक से संपर्क करें।"
+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।
 
@@ -251,89 +238,10 @@ def format_response_for_tts(text: str) -> str:
     return text.strip()
 
 
-
-# ─────────────────────────────────────────────
-# EMBEDDING RETRIEVER — multilingual, Hindi-aware
-# Uses paraphrase-multilingual-MiniLM-L12-v2
-# Falls back to SimpleRetriever if ChromaDB not ready
-# ─────────────────────────────────────────────
-
-try:
-    from sentence_transformers import SentenceTransformer
-    from FlagEmbedding import FlagReranker
-    import chromadb
-
-    print("Loading embedding model (multilingual MiniLM)...")
-    embedder = SentenceTransformer(
-        'sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2'
-    )
-    print("Loading reranker (bge-reranker-v2-m3)...")
-    reranker = FlagReranker('BAAI/bge-reranker-v2-m3', use_fp16=False)
-    
-    chroma_client = chromadb.PersistentClient(path="./chroma_db")
-    EMBEDDING_READY = True
-    print("Embedding retriever ready.")
-
-except Exception as e:
-    print(f"Embedding retriever not available: {e}. Using keyword retriever.")
-    EMBEDDING_READY = False
-
-
-def retrieve_with_embeddings(query: str, top_k: int = 3) -> list:
-    """
-    Two-stage retrieval:
-    Stage 1 - ChromaDB embedding search (top 10)
-    Stage 2 - bge-reranker picks best 3
-    Falls back to SimpleRetriever if embeddings not ready.
-    Critical for Hinglish queries like 
-    'home loan ka interest kitna hai for salaried?'
-    """
-    if not EMBEDDING_READY:
-        print("DEBUG RAG: falling back to keyword retriever")
-        retriever = get_retriever()
-        return retriever.retrieve(query, top_k=top_k)
-    
-    try:
-        collection = chroma_client.get_collection("banking_hindi")
-    except Exception:
-        print("DEBUG RAG: ChromaDB collection not found, run ingest.py first")
-        print("DEBUG RAG: falling back to keyword retriever")
-        retriever = get_retriever()
-        return retriever.retrieve(query, top_k=top_k)
-    
-    try:
-        # Stage 1: embedding similarity search
-        query_embedding = embedder.encode([query]).tolist()
-        results = collection.query(
-            query_embeddings=query_embedding,
-            n_results=min(10, collection.count())
-        )
-        candidates = results['documents'][0]
-        metadatas = results['metadatas'][0]
-        
-        print(f"DEBUG RAG: {len(candidates)} candidates from ChromaDB")
-        
-        # Stage 2: rerank
-        pairs = [[query, doc] for doc in candidates]
-        scores = reranker.compute_score(pairs)
-        ranked = sorted(
-            zip(scores, candidates, metadatas), 
-            reverse=True
-        )
-        
-        top_results = [
-            {
-                "id": meta.get("id", ""),
-                "title": meta.get("title", ""),
-                "content": doc,
-                "category": meta.get("category", "")
-            }
-            for _, doc, meta in ranked[:top_k]
-        ]
-        print(f"DEBUG RAG: top result = {top_results[0]['title'] if top_results else 'none'}")
-        return top_results
-        
-    except Exception as e:
-        print(f"DEBUG RAG: embedding retrieval error: {e}")
-        retriever = get_retriever()
-        return retriever.retrieve(query, top_k=top_k)
+def get_tts_description(text: str) -> str:
+    """Speaker description for Indic-Parler-TTS."""
+    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

@@ -1,12 +1,4 @@
-# Stability Fixes (Deep Lock)
-pydantic==2.9.2
-fastapi==0.115.5
-
-# Core Logic Dependencies
+gradio>=5.9.1
+huggingface_hub>=0.34.0
 soundfile>=0.12.1
-faster-whisper
-chromadb
-sentence-transformers
-FlagEmbedding
-requests
-edge-tts
+numpy>=1.24.0

stt.py

@@ -1,36 +0,0 @@
-"""
-STT Module - faster-whisper medium int8 on CPU
-Transcribes Hindi audio to text.
-Faster than openai/whisper via API, runs locally, no API cost.
-"""
-from faster_whisper import WhisperModel
-import numpy as np
-import soundfile as sf
-import io
-
-print("Loading Whisper model (medium int8)... first run downloads ~500MB")
-whisper_model = WhisperModel("medium", device="cpu", compute_type="int8")
-print("Whisper model loaded.")
-
-def stt_whisper(audio_array: np.ndarray, sample_rate: int) -> str:
-    """
-    Convert Hindi audio array to text.
-    Keeps same function signature as existing stt_whisper() in app.py.
-    """
-    try:
-        buf = io.BytesIO()
-        sf.write(buf, audio_array, sample_rate, format="WAV", subtype="PCM_16")
-        buf.seek(0)
-        
-        segments, info = whisper_model.transcribe(
-            buf,
-            language="hi",
-            beam_size=5
-        )
-        transcript = " ".join([s.text for s in segments]).strip()
-        print(f"DEBUG STT transcript: {transcript}")
-        print(f"DEBUG STT detected language: {info.language} confidence: {info.language_probability:.2f}")
-        return transcript
-    except Exception as e:
-        print(f"STT error: {e}")
-        return ""

tts.py

@@ -1,63 +0,0 @@
-"""
-TTS Module - Microsoft Edge TTS
-Uses hi-IN-SwaraNeural — natural Hindi female voice.
-No compilation needed, works on Python 3.14 + Windows.
-No API key required.
-Returns (sample_rate, audio_array) tuple — same as existing tts_hindi().
-"""
-import edge_tts
-import asyncio
-import numpy as np
-import soundfile as sf
-import io
-import tempfile
-import os
-
-HINDI_VOICE = "hi-IN-SwaraNeural"
-
-def tts_hindi(text: str):
-    """
-    Convert Hindi text to audio using Edge TTS.
-    Keeps same function signature as existing tts_hindi() in app.py.
-    Returns (sample_rate, audio_array) tuple or None on failure.
-    """
-    try:
-        if not text or not text.strip():
-            return None
-        
-        # Limit text length
-        if len(text) > 500:
-            text = text[:500]
-            print("DEBUG TTS: text truncated to 500 chars")
-        
-        # Edge TTS is async — run it synchronously
-        async def _generate():
-            communicate = edge_tts.Communicate(text, HINDI_VOICE)
-            # Use tempfile to get a proper temp path
-            fd, tmp_path = tempfile.mkstemp(suffix=".mp3")
-            os.close(fd) # Close file descriptor immediately
-            try:
-                await communicate.save(tmp_path)
-                return tmp_path
-            except Exception as e:
-                if os.path.exists(tmp_path):
-                    os.remove(tmp_path)
-                raise e
-        
-        # Run async function
-        tmp_path = asyncio.run(_generate())
-        
-        # Read audio file
-        audio_array, sample_rate = sf.read(tmp_path)
-        audio_array = audio_array.astype(np.float32)
-        
-        # Cleanup temp file
-        if os.path.exists(tmp_path):
-            os.remove(tmp_path)
-        
-        print(f"DEBUG TTS: generated audio at {sample_rate}Hz")
-        return (int(sample_rate), audio_array)
-        
-    except Exception as e:
-        print(f"TTS error: {e}")
-        return None