guy-hartstein/company-research-agent

Company Research Agent – What it is

A small web app that lets you type a company name and get a structured, source‑backed briefing (business overview, industry context, recent news, etc.). The heavy lifting is done by a FastAPI backend that orchestrates a research graph of LLM‑driven steps and a Tavily web‑search/relevance service. The React + Vite front‑end streams progress to the browser and can export the final markdown report as a PDF.

---

Core capabilities

Feature	How it works
Automated research	Sends queries to Tavily, scores relevance, and collects URLs.
LLM‑driven synthesis	Uses any model reachable via OpenRouter (or direct OpenAI/Gemini keys) to generate search prompts, summarize sources, and write the final briefing.
Live streaming	The UI receives server‑sent events from GET /research/{job_id}/stream so you see the job progress in real time.
PDF export	The backend converts the markdown report to PDF via POST /generate-pdf.
Optional persistence	If a MongoDB URI is supplied, jobs and reports are stored for later retrieval.
Model flexibility	Pick a model in the Settings UI; the same model is used for every step of the pipeline.

---

Architecture at a glance

React + Vite (frontend)  ←→  FastAPI (backend)
        │                     │
        └─ POST /research ──►  Research graph
                               ├─ analyzers
                               ├─ collector (Tavily)
                               ├─ curator (relevance scoring)
                               ├─ briefing (LLM synthesis)
                               └─ editor (final markdown)

The backend talks to:

• Tavily – web search and relevance ranking.
• LLM – any model on OpenRouter, or OpenAI/Gemini directly, for prompt generation, summarisation, and editing.
• MongoDB (optional) – job/report storage.

---

Getting started (local development)

1. Backend – Python 3.11+, create a venv, install requirements.txt, copy .env.example → .env and fill in your OpenRouter/Tavily keys (MongoDB optional).
2. Frontend – copy ui/.env.development.example → ui/.env.development.local, set VITE_API_URL to http://localhost:8000, install npm deps.
3. Run –
   • uvicorn application:app --reload --port 8000 (API)
   • npm run dev inside ui (UI) → open http://localhost:5174.
4. Docker – after the same .env files exist, docker compose up --build brings up both services.

---

Typical workflow

curl -X POST http://localhost:8000/research \
  -H 'Content-Type: application/json' \
  -d '{"company":"Tavily","company_url":"https://tavily.com","industry":"AI search","hq_location":"New York, USA","model":"anthropic/claude-sonnet-5.5"}'

• The response gives a job_id.
• The UI (or a GET …/stream request) streams progress events.
• When finished, GET /research/{job_id}/report returns the markdown; POST /generate-pdf creates a downloadable PDF.

---

Who might use it?

• Investors or analysts needing quick, sourced overviews of target companies.
• Business development teams scouting partners or competitors.
• Students or journalists who want a reproducible briefing without manual web‑search.
• Developers looking for a template to build other LLM‑orchestrated research agents.

---

Limitations & considerations

• API‑key dependent – you must supply a working OpenRouter (or OpenAI/Gemini) key and a Tavily key; without them the service cannot run.
• Source quality – the relevance scoring is heuristic; the final report reflects whatever the LLM synthesises from the retrieved pages.
• Persistence optional – without MongoDB each job lives only in memory for the current server run.
• Model cost – using large models via OpenRouter can be expensive; choose a model that balances price and quality.

---

License

Apache License 2.0 – free to use, modify, and distribute.