Multi-Agent AI Research

Four AI agents turn a topic into a sourced, peer-reviewed research report:

#	Agent	Does	Tool
1	Scout	Searches the live web for reliable, recent sources	web_search (Tavily)
2	Reader	Opens the best links and extracts the readable text	scrape_url
3	Writer	Synthesises everything into a structured report	llama-3.3-70b on Groq
4	Critic	Scores the report out of ten and flags what to improve	llama-3.3-70b on Groq

Setup

Put your keys in .env:

TAVILY_API_KEY=...
GROQ_API_KEY=...

Install dependencies:

uv pip install -r backend/requirements.txt     # or: pip install -r backend/requirements.txt

Run the website

Build the React interface once, then start the server:

cd frontend && npm install && npm run build && cd ..
python backend/server.py

Then open http://127.0.0.1:8000. The landing page is at /; the research workspace is at /app. Enter a topic and each agent reports its status live as it works — the page follows the run over server-sent events, with a pipeline graph, per-agent timings and a live log feed alongside.

backend/server.py serves frontend/dist. If you start it before building, the API at /api still runs and the root path returns the build instructions rather than crashing. See frontend/README.md for the dev-server workflow and for which features are backed by the API versus stored in the browser.

Before publishing the landing page: the metrics, testimonials and paid pricing tiers in frontend/src/content/placeholders.ts are invented placeholders. Replace them or delete those sections — publishing fabricated statistics or reviews as real is deceptive. The sections currently render a visible note saying so.

When the run finishes you get:

• Report — the rendered Markdown report
• Critique — the critic's score, strengths and gaps
• Sources — every URL the scout found, as clickable cards
• Raw output — the unedited search and reader agent output

Reports can be copied or downloaded as Markdown, and completed runs are listed under Recent research (stored in your browser only).

Other entry points

python backend/pipeline.py           # CLI: prompts for a topic, prints each stage
streamlit run backend/Streamlit.py

Layout

backend/
  server.py       FastAPI app — JSON + SSE API, serves the frontend
  pipeline.py     The four-step run, emits progress events
  agents.py       Agent + chain definitions
  tools.py        web_search and scrape_url
  Streamlit.py    Alternative Streamlit UI
  requirements.txt
frontend/         React + TypeScript interface (see frontend/README.md)
.env              API keys, shared by everything above

Troubleshooting

"Could not reach the research server" — the agents run server-side, so python backend/server.py must be running. Browse to http://127.0.0.1:8000, or to http://localhost:5173 if you are using the Vite dev server.

The root path returns build instructions — frontend/dist is missing. Run cd frontend && npm install && npm run build, then restart the server.

"Groq's daily token limit ... has been used up" — the free Groq tier caps tokens per day (100k on on_demand). It resets on a 24-hour rolling window; a full research run costs roughly 10–20k tokens.

A step shows "Retrying" — Groq occasionally returns tool_use_failed when the model emits a malformed function call. Each step retries up to 3 times; quota errors are not retried, since they can't clear.

Notes

• The reader step is treated as optional: if a scrape or tool call fails, the run is marked skipped for that agent and still produces a report from the search results rather than failing outright.
• Runs are held in server memory, so restarting the server clears history replay (the browser then simply re-runs the topic).
• Reports are AI-generated — check the sources before relying on them.