ResearchMind

01

The problem

One-shot LLM answers to research questions are unsourced and hard to trust. Splitting the work into search, reading, writing and review steps makes every stage inspectable and gives the final report an explicit quality score.

02

Key features

• 01Search agent breaks the topic into sub-queries and gathers unique sources through Tavily
• 02Reader agent scrapes each source with BeautifulSoup and extracts relevant insights
• 03Writer agent turns the source notes into a structured Markdown report
• 04Critic agent scores the report from 0 to 10 with written feedback
• 05Live dashboard streams each agent's progress over Server-Sent Events

03

Architecture highlights

A LangGraph StateGraph runs START → search → read → write → critique → END over a typed shared state. Each node emits progress events that FastAPI streams to the browser.

• Shared ResearchState (topic, sources, summaries, report, critique, events) is the only contract between agents.
• The pipeline runs off the event loop and its events are bridged into an SSE generator, so the UI updates as each agent finishes.

04

Tech stack