Build a Deep Researcher That Beats OpenAI, Gemini, and Perplexity

A walkthrough of building a fully open-source, self-hostable deep research stack using Onyx (retrieval), CrewAI (orchestration), and Voxtral (voice). The architecture separates research into distinct stages — clarification, planning, and iterative execution — to avoid context bleed and hallucinations. Onyx's 6-stage retrieval pipeline uses hybrid search, LLM filtering, and adaptive reflection loops, and is benchmarked #1 on DeepResearch Bench ahead of OpenAI, Gemini, and Perplexity. CrewAI Flows enforce clean stage separation between Researcher, Analyst, and Report Writer agents. Voxtral adds voice input and report narration. The key motivation is data sovereignty: all queries and internal documents stay on your own infrastructure.