V2 vision: evolve WhisperDocs from a working voice + PDF prototype into a production-oriented document intelligence platform that can support 1000+ users, paid subscriptions, reliable retrieval, measurable latency, and a deployable real-world operating model.

1. V2 goals

Product goals

• Support multi-user document libraries with isolated tenant/document access.
• Add subscriptions with plan-based quotas and usage tracking.
• Support reliable uploads and long-running document ingestion.
• Provide grounded answers with source/page references.
• Preserve the core experience: upload a PDF → talk to it naturally by voice.

Engineering goals

• Improve RAG quality instead of relying only on basic text matching.
• Reduce retrieval and end-to-end voice latency.
• Make ingestion asynchronous and retryable.
• Scale interactive voice traffic separately from heavy document processing.
• Add observability, evaluation, rate limiting, and failure recovery.
• Establish measurable baselines before large-scale optimization.

2. RAG V2 - from basic search to a production retrieval funnel

The current system stores page-indexed text chunks and searches them through the application server. V2 introduces a multi-stage retrieval funnel: filtering, hybrid retrieval, reranking, evaluation, and traceability. Production RAG architecture commonly separates ingestion, retrieval, augmentation, and generation, with retrieval quality being a major source of overall answer quality.

Target retrieval flow

User question → query understanding/rewrite → tenant + document authorization filter → lexical + semantic candidate retrieval → merge/deduplicate → reranker → context compression/top-k → grounded prompt → LLM → citations/source pages.

Retrieval funnel