Multilingual RAG with Bhashini

Cross-lingual retrieval across Hindi, Marathi, Tamil, Bengali, and English

Multilingual RAG with Bhashini Cross-lingual retrieval across Hindi, Marathi, Tamil, Bengali, and English 01 / Query 02 / Translate 03 / Retrieve 04 / Generate 05 / Respond User Query · any Indian language · 01 / Query · 5 languages User Query any Indian language 5 languages Bhashini · query translation · 02 / Translate · govt API Bhashini query translation govt API Vector Index · language-agnostic · 03 / Retrieve · shared Vector Index language-agnostic shared Cross-lingual · multilingual retrieval · 03 / Retrieve Cross-lingual multilingual retrieval Answer Generation · grounded on passages · 04 / Generate Answer Generation grounded on passages Bhashini · response translation · 05 / Respond Bhashini response translation Response · user's own language · 05 / Respond · localized Response user's own language localized native query pivot query doc vectors ranked passages answer (pivot) translated answer Legend primary data policy / PII async batch data store

One Corpus, Many Languages

  • • Bhashini normalizes any Indian-language query to a pivot
  • • Multilingual embeddings retrieve across a single shared index
  • • No per-language corpus duplication is required

Answer in the User's Language

  • • Generation grounds on cross-lingual passages
  • • Bhashini translates the answer back to the query language
  • • The round trip is invisible to the user