← All workNLP / Retrieval
IndicRAG-QA — Cross-lingual RAG
A RAG pipeline for English, Hindi and code-mixed queries that cites a passage or abstains, with script-aware RRF fusion.
- Stack
- Python · BM25 · Dense Retrieval · LLMs
- Key figure
- 0.153 — Recall@5, from 0.022
DG Labs
IndicRAG-QA
0.153Recall@5, from 0.022
- English, Hindi and code-mixed queries
- Cites a passage or abstains
- Script-aware RRF fusion
Python · BM25 · Dense Retrieval · LLMs
01
The problem
Retrieval breaks down when the question and the evidence are in different scripts — English, Hindi and code-mixed queries rarely find the right passage, and a model that answers anyway will make things up.
02
The approach
Built a RAG pipeline combining BM25 and dense retrieval, with a script-aware reciprocal-rank-fusion step. The generator must cite a passage — or abstain.
03
The outcome
Cross-lingual Recall@5 rose from 0.022 to 0.153 (p < 0.001).
Of note
- ✦ English, Hindi and code-mixed queries
- ✦ Cites a passage or abstains
- ✦ Script-aware RRF fusion
- ✦ Recall@5: 0.022 → 0.153 (p < 0.001)