Sub-1.5s first audible word
Audio starts streaming before the LLM has finished writing — perceived latency drops to near-zero.
Voice + RAG·Production showcase
Speak the question.
Hear the answer.
Upload PDFs, ask out loud, hear cited answers streamed back as natural speech. Full voice loop in production — mic capture, Whisper STT, hybrid retrieval, low-latency TTS.
Showcase mode · 5 queries / 3 docs / 5-min session TTL
Pipeline
Five stages, end to end
Every voice query travels this exact path. No black box.
01
Upload PDF
Document is split into semantic chunks via LangChain text splitters.
02
Embed Locally
FastEmbed (BAAI/bge-small-en-v1.5) runs on the server — no external embedding API.
03
Hybrid Retrieval
Question is embedded and fused with full-text search via pgvector + tsvector.
04
AI Synthesis
Processor Agent (GPT-4.1-mini) writes a grounded answer with citations.
05
Stream Speech
GPT-4o-mini-TTS streams PCM audio over SSE into the Web Audio API.
FastEmbed local · pgvector + tsvector · OpenAI Agents SDK · SSE audio streaming
What makes it different
Engineered for production, not demos
Tenant-isolated retrieval
Every chunk row carries a session_id. Vector and keyword queries WHERE-filter on it before fusion. No cross-tenant leak.
Hybrid search, not just vectors
pgvector HNSW + tsvector GIN, merged via RRF. Catches both semantic similarity and exact-term matches.
Local embeddings
FastEmbed runs on the server in ONNX. No OpenAI cost for ingestion, full data control.
9 distinct voices
Coral, alloy, echo, fable, onyx, nova, sage, shimmer, verse — pick a voice that fits the use case.
Stack
Built with
Next.js 16React 19FastAPIPostgreSQL 17pgvectorFastEmbed (ONNX)OpenAI WhisperGPT-4.1-miniGPT-4o-mini-TTSOpenAI Agents SDKServer-Sent EventsWeb Audio API
Ready to talk to your documents?
Upload a PDF, pick a voice, ask anything. The demo runs on a 5-minute session.
Open the app