Voice-native RAG · MSMARCO-XI

Ask anything.
Grounded in truth.

A voice pipeline that transcribes, retrieves, and answers only from what it can prove — hybrid lexical retrieval over IndicRAGSuite's MSMARCO-XI corpus, six chunking strategies, and hard guardrails that know when to stay silent.

STT → ElevenLabs
Hybrid BM25 + TF-IDF
6 chunking strategies
3 independent guardrails
Sub-200ms target
ENTER THE PIPELINE
Live console

Speak, or type, your question

Runs entirely against the bundled MSMARCO-XI-style sample corpus (29 query/passage pairs, EN + HI) using the same chunking + hybrid retrieval + guardrail logic as the production Python pipeline.

Mic uses your browser's on-device speech recognition for this live demo (fallback when no ElevenLabs key is configured). Production transcription runs on ElevenLabs Scribe — see stt.py.
Pipeline trace

Answer

—

Retrieved context

Latency — this query

Pipeline shape

Voice → text → retrieve → guardrail → answer

01

Speech-to-text

ElevenLabs Scribe with structured retry/backoff on 429/5xx. Browser Web Speech API stands in for this live demo.

02

Multi-strategy chunking

Fixed-size, sentence-window, semantic (topic-shift), recursive, and query-passage-aware — routed per document.

03

Hybrid retrieval

TF-IDF cosine + BM25, fused with Reciprocal Rank Fusion, over a numpy vector store built for embedding drop-in.

04

Guardrails

Unsafe-input filter, off-topic confidence floor, and a grounding check that verifies the answer against retrieved text.

Corpus: ai4bharat/MSMARCO-XI (huggingface.co/datasets/ai4bharat/MSMARCO-XI) · sample corpus bundled for offline demo · full pipeline in rag_pipeline/