Minimum Viable Product (MVP)
The MVP is a conversational AI chat interface where users ask questions about the Vedas and receive contextual, referenced answers powered by the Vedic knowledge base, web search, and Sarvam AI.
Explore: veda-guru.com — landing page with all entry points
Scope
| Area | Included in MVP | Future |
|---|---|---|
| Chat | Text-based Q&A with AI | Voice, guided tutorials |
| Vedic KB | DharmicData (Rig/Yajur/Atharva) + HF IndianVedasOriginal (Sama, Krishna Yajur) — BM25 search | Upanishads, Brahmanas |
| Search | Vedic corpus (tried first) → Tavily → Wikipedia → DuckDuckGo → Google | Fine-tuned Vedic model |
| LLM | Sarvam AI (sarvam-105b, 128K context) |
Fine-tuned Vedic model |
| TTS | Sarvam Bulbul v3 with browser fallback | Multiple voices, languages |
| Languages | English | 10+ Indian languages |
| References | Source labels from Vedic corpus (text citations) + web URLs | Clickable verse references |
| Auth | Email/password register + login | SSO, OAuth providers |
| History | Auto-saved conversations per user | Search, export, share |
| Voice | Maharshi page with mic + TTS (Beta) | Full voice conversation |
| Security | Rate limiting, CORS lock, JWT auth | Per-session tokens, 2FA |
| UI/UX | Landing page, auto-resize textarea, smart scroll, stop button, follow-up suggestions | Full design system |
| Learning | Gurukula learning page (/learn) with 3 tiers, 36 adhyayas, daily reading, pronunciation guides, auto-generated quizzes, progress tracking (localStorage + server sync) |
Advanced analytics, peer discussions |
How it works
- User lands on the dharmic-themed landing page (
/) and clicks Enter Chat - Chat page (
/chat) loads, connects to the backend - User asks a question (e.g. “What does the Rig Veda say about creation?”)
- Backend searches the Vedic corpus first (BM25 over DharmicData + HuggingFace Vedas)
- If corpus results found → used as context; if not → falls back to web search
- Context + question is fed to Sarvam AI (
sarvam-105b) - Sarvam AI synthesizes an answer with source citations
- Response is rendered with markdown formatting in the chat UI
- If logged in, the exchange is auto-saved to conversation history
- Follow-up suggestions are fetched in the background and shown as clickable chips
User flow
Landing page (/)
│
▼ Click "Enter Chat"
Chat interface (/chat)
│
▼ Ask a question
Vedic corpus search (BM25 — tried first)
│
├── Found? → Use as context
│
└── Not found? → Web search (Tavily → Wikipedia → DuckDuckGo)
│
▼
Context + question sent to Sarvam AI
│
▼
AI synthesizes answer with source citations
│
▼
Response rendered with markdown in chat UI
│
├── Auto-saved to history (if logged in)
│
└── Follow-up suggestions fetched and displayed
Tech stack (MVP)
| Component | Choice |
|---|---|
| LLM | Sarvam AI (sarvam-105b, 128K context) |
| TTS | Sarvam AI bulbul:v3, shubh voice |
| Vedic KB | BM25 in-memory (DharmicData JSON + HuggingFace IndianVedasOriginal) |
| Backend | Python FastAPI (GCP Cloud Run) |
| Frontend | Vanilla HTML/CSS/JS (Cloudflare Pages) — 4 pages: landing, chat, learning, maharshi |
| Database | Neon Postgres (persistent) or SQLite fallback |
| Auth | HMAC-signed JWT tokens, SHA-256 password hashing |
| Web search | Tavily (primary) + Wikipedia API + DuckDuckGo + Google |
| Rate limiting | slowapi (30 req/min per IP) |
| Markdown render | marked.js with Vedic-themed code block styles |
| Speech-to-text | Browser Web Speech API (free, no API calls) |
| Corpus download | GitHub API (version-tracked) + HuggingFace direct download |
Success metrics
- Accuracy: Responses are factually grounded in search results (>90% precision)
- Relevance: User finds the answer helpful (>80% satisfaction)
- Latency: First response within 5s (or up to 60s on cold start)