Deep Research Station — 深度研究台

Deep Research Station is a fully local, privacy-first deep research platform that runs multi-agent AI workflows entirely on consumer GPU hardware. It replaces cloud-dependent research tools like Perplexity or ChatGPT Deep Research with a self-hosted system where all LLM inference happens on a local RTX 5090, eliminating API costs and data privacy concerns. The custom-built Next.js frontend connects to a LangGraph-powered multi-agent backend, providing real-time visualization of the research process as a Supervisor agent coordinates multiple parallel Researcher agents to investigate complex topics and produce comprehensive reports.

• Zero-cost local inference

  All four LLM roles (research, summarization, compression, report generation) run on a single RTX 5090 via Ollama. No API keys, no per-token billing, and no data ever leaves the machine.

• Multi-agent architecture

  A Supervisor agent decomposes a research question into sub-tasks, delegates them to multiple Researcher agents running in parallel, collects and compresses their findings, then generates a final structured report — all orchestrated as a LangGraph state machine.

• Real-time execution visualization

  The frontend shows a live graph-execution panel with the currently active node (Clarify → Brief → Research → Report), streaming token output, and per-node status indicators over SSE.

• Custom frontend, no LangSmith login

  A purpose-built Next.js UI with a three-panel layout — session history, chat interface, and graph state + runtime configuration — that removes the LangSmith account dependency of the default LangGraph Studio.

• Dynamic runtime configuration

  Users can switch Ollama models, search engines, research depth (iteration count), parallelism (concurrent research units), and clarification behavior straight from the UI, without restarting any services.

• Session persistence

  Thread-based conversation management with full state checkpointing — research can be paused, resumed, or branched from any point.

A research question flows through a four-stage LangGraph state machine — clarify, brief, research, report — where a Supervisor agent plans strategy and dispatches parallel Researcher agents, each running its own ReAct search loop, then compresses findings and loops until the brief is satisfied.

1. 1

   Clarify

   clarify_with_user analyzes whether the question is specific enough; if it is ambiguous the graph asks a clarifying question and waits, otherwise it proceeds.

2. 2

   Brief

   write_research_brief transforms the user's messages into a structured research brief with scope, sub-questions, and success criteria.

3. 3

   Research — Supervisor + parallel Researchers

   The Supervisor reads the brief, reflects with a think_tool, and delegates via the ConductResearch tool to multiple Researcher agents dispatched in parallel. Each Researcher runs a ReAct loop — search, analyze, reflect on gaps, search again — then compresses its findings. The Supervisor evaluates completeness and either loops back for more research or signals ResearchComplete.

4. 4

   Report

   final_report_generation synthesizes all compressed research notes into a structured markdown report with citations and findings.