RAG over mixed Hindi-English documents: what breaks and what to check

Indian document sets often mix Hindi, English and transliterated text on the same page. Standard embedding models handle this unevenly — what to test before trusting them.

Written by

Pranjul Rathour

GenAI Engineer at SCULT INDIA, based in Kanpur, India. Builds production RAG systems, fine-tunes open LLMs and ships agentic AI products; has mentored 200+ college developers. Open to GenAI roles, hackathon judging, mentorship sessions and guest talks.

pranjulrathour41@gmail.com LinkedIn GitHub X Instagram Bluesky Dev.to

At an Integral Startup Foundation hackathon

Indian document sets — government forms, college notices, WhatsApp exports — often mix Hindi, English and transliterated Hindi-in-Roman-script on the same page. A standard embedding model, benchmarked mostly on clean monolingual English text, handles this unevenly.

What to actually check

Does the embedding model have documented multilingual support, and specifically Hindi — not just a claim of '100+ languages' with no per-language evaluation.
Test retrieval with a query in one language against documents in the other; cross-lingual retrieval quality varies a lot between models.
Transliterated text (Hindi words spelled in Roman script) is the hardest case — test it explicitly rather than assuming a multilingual model handles it like proper Devanagari.

A practical fallback

If cross-lingual retrieval quality is poor, translating queries and documents into one common language before embedding, using a small dedicated translation step, is a reliable workaround.

See embeddings models for RAG: how to choose.

— Pranjul Rathour, GenAI Engineer from Kanpur, India. Open to GenAI roles, hackathon judging, mentorship sessions and guest talks at any campus: pranjulrathour41@gmail.com.