Kumbh Mela 2025 · kumbh.gov.in
Kumbh Sahayak
Multilingual assistant for pilgrims, built on hybrid retrieval, tuned prompts and guardrails, running on an 8B open model.
10M+ users · 90%+ accuracy
Overview
Kumbh Sahayak answered pilgrims’ questions about routes, services and rituals across languages during Kumbh Mela 2025. It handled 10k+ concurrent requests at peak. The main constraint was quality on a small open model, so most of the work went into retrieval, prompt optimization and evaluation rather than a bigger LLM.
Architecture
Separate services for the RAG pipeline, guardrails and media generation. Retrieval is hybrid: BM25 for lexical matches and fine-tuned BGE embeddings for semantic ones, followed by a BGE reranker.
- Hybrid retrieval: BM25 + vector search + BGE reranker
- Input and output guardrails modelled on AWS Guardrails, plus custom rules
- Prompt optimization with TextGrad and DSPy against a labelled eval set
- Fuzzy-match cache (Levenshtein) for repeated questions
- Similar-chunk merging using Jensen–Shannon divergence
- Voice cloning (VITS, Coqui-TTS, ElevenLabs) and Wav2Lip lip-sync for official video messages
Query → input guardrails → hybrid retrieval → rerank → LLM → output guardrails → user
What I did
- Built the question-answering pipeline combining BM25, fuzzy and vector search
- Fine-tuned BGE embeddings and applied the BGE reranker for ranking quality
- Optimized prompts with TextGrad and DSPy, measured on a curated evaluation set
- Designed the safety layer for a public, government-facing deployment
- Built the speech synthesis and lip-sync pipeline with the Samagra team, including audio clean-up with NVIDIA NeMo
References
- FlagEmbedding (BGE)
Embedding and reranker models used for retrieval.
- TextGrad
Automatic “differentiation” via text, used for prompt optimization.
- Wav2Lip
Lip-sync model from IIIT Hyderabad.
- VITS
End-to-end text-to-speech used for voice cloning.