README.md

metadata

title: Hindi Banking Voice Assistant
emoji: 🏦
colorFrom: yellow
colorTo: green
sdk: gradio
sdk_version: 5.16.1
python_version: 3.11
app_file: app.py
pinned: true
license: mit
suggested_hardware: cpu-basic
short_description: Hindi voice assistant for banking & govt schemes

🏦 बैंकिंग सहायक — Hindi Banking Voice Assistant

A voice-powered AI assistant for Indian users to understand banking, loans, and government financial schemes — entirely in Hindi. Runs on CPU Basic (free tier) via HF Inference API.

🔑 Setup (Required — One Time)

1. Go to your Space → Settings → Secrets
2. Add secret → Name: HF_TOKEN | Value: your free HF read token from hf.co/settings/tokens

That's it. All 3 models are called via API — nothing loads into the Space's RAM.

---

🤖 Models Used

Role	Model
Speech-to-Text	openai/whisper-large-v3-turbo
LLM	Telugu-LLM-Labs/Indic-gemma-7b-finetuned-sft-Navarasa-2.0
Text-to-Speech	ai4bharat/indic-parler-tts

---

🔄 Pipeline

🎤 Hindi Audio
    ↓
[HF API] openai/whisper-large-v3-turbo      →  Hindi transcript
    ↓
Jargon Normalization (local, pure Python)   →  kist→EMI, zamanat→collateral ...
    ↓
Keyword RAG Retrieval (local, pure Python)  →  top 3 relevant KB passages
    ↓
[HF API] Indic-gemma-7b (Navarasa 2.0)     →  Hindi answer grounded in KB
    ↓
[HF API] ai4bharat/indic-parler-tts         →  Hindi speech audio
    ↓
🔊 Answer (audio + text + citations)

Local compute: ~50 MB RAM for knowledge base + retriever. Zero GPU required.

---

⏱️ Expected Response Times (free tier)

Step	Time
STT (Whisper API)	5–10s
RAG Retrieval	< 1s
LLM (Indic-Gemma API)	15–40s (cold start up to 60s)
TTS (Parler API)	10–20s
Total	~30–70s per query

First call may be slower due to HF cold start. The app retries automatically.

---

📖 Knowledge Base Coverage

Banks (12)

SBI, HDFC Bank, ICICI Bank, Axis Bank, Kotak Mahindra Bank, PNB, Bank of Baroda, Canara Bank, Union Bank of India, Indian Bank, Small Finance Banks (AU / Ujjivan / Equitas / Jana), Regional Rural Banks & Co-operative Banks, NBFCs (Bajaj Finance / Muthoot / Manappuram)

Government Credit Schemes (11)

Scheme	Coverage
PM Mudra Yojana	Shishu / Kishore / Tarun loans up to ₹10L (Tarun Plus ₹20L)
Stand-Up India	SC/ST & women entrepreneurs ₹10L–₹1Cr
PM SVANidhi	Street vendors ₹10K → ₹20K → ₹50K
PMEGP + KVIC	Employment generation with 15–35% govt subsidy
CGTMSE + ECLGS	Collateral-free MSME loans
PM Awas Yojana CLSS	Home loan interest subsidy for EWS / LIG / MIG
Kisan Credit Card	Agricultural credit at effective 4% p.a.
PM Vishwakarma	Artisan loans at 5% concessional rate
PMJDY + Overdraft	Zero-balance accounts + ₹10,000 overdraft
Vidya Lakshmi + CSIS	Education loans + interest subsidy
NRLM-SHG	Women Self Help Group bank linkage program

Social Security Schemes (6)

Scheme	Benefit
PMJJBY	₹2 lakh life insurance at ₹436/year
PMSBY	₹2 lakh accident insurance at ₹20/year
Atal Pension Yojana	Guaranteed monthly pension for unorganized sector
PM-KISAN	₹6,000/year direct income support for farmers
EPFO / EPF / EPS	Provident fund & pension for formal employees
PMFBY	Crop insurance at 1.5–2% farmer premium
MGNREGS	100 days guaranteed employment per year

Financial Terms & Concepts

EMI, CIBIL Score, KYC documents, MCLR & Repo Rate, Collateral & Guarantor, Moratorium & Loan Restructuring, Processing Fees & Prepayment Charges, NPA & OTS, Loan types (home / personal / vehicle / gold / education / LAP), UPI / NEFT / RTGS, FD / RD, Microfinance (JLG / SHG), Consumer Rights & RBI Ombudsman

---

📁 Repository Structure

app.py              Gradio app + HF Inference API pipeline (STT → RAG → LLM → TTS)
rag_pipeline.py     Jargon normalization map, keyword retriever, prompt builder
knowledge_base.py   40+ documents in English (retrieved, answers given in Hindi)
requirements.txt    Minimal: gradio, requests, soundfile, numpy
README.md           This file (HF Space config + documentation)

---

💡 Design Principles

• Grounded answers only — LLM responds only from retrieved KB content, not hallucinations
• Always in Hindi — STT, LLM prompt, and TTS all enforce Hindi throughout
• One follow-up at a time — conversational, not overwhelming
• Spoken-friendly output — short sentences, no technical jargon in answers
• Graceful fallback — says "बैंक से संपर्क करें" when information is unavailable
• Extensible — adding a new language requires only changing STT language code + TTS voice description

---

📝 Example Questions

• मुझे होम लोन के लिए कौन से दस्तावेज़ चाहिए?
• EMI क्या होती है और कैसे calculate होती है?
• मुद्रा लोन कैसे मिलेगा?
• CIBIL score क्या होता है?
• PM Awas Yojana में subsidy कैसे मिलती है?
• जन धन खाता कैसे खोलें?
• किसान क्रेडिट कार्ड क्या है?
• महिला स्वयं सहायता समूह को loan कैसे मिलेगा?

---

⚠️ Disclaimer: This assistant provides general information only. For actual financial decisions, please consult your bank or a certified financial advisor. यह assistant केवल सामान्य जानकारी देता है। वित्तीय निर्णयों के लिए अपने बैंक या certified advisor से सलाह लें।