README.md

metadata

title: Web Research Agent
emoji: 🔎
colorFrom: gray
colorTo: red
sdk: docker
app_port: 7860
pinned: false

Web Research Agent 🔎

Ask a question → get a cited briefing built from live web sources you can verify — then dig deeper into any claim, section, or source, in your own words, and it appends an expandable research branch.

Project #13 of the "30 AI Projects in 15 Days" build-in-public challenge. Focus: multi-step agent orchestration (plan → search → read → reflect → cite).

What makes it different

Most research tools hand you a static answer. This one is explorable: click dig deeper on any citation, highlighted sentence, or source, type what you want investigated, and it spawns a scoped research pass rendered inline as a branch — which you can dig into again. A recursive research tree.

How it works (honest by construction)

• Plan — the question is broken into targeted search queries.
• Search — keyless, multi-source: general web (DuckDuckGo / SearXNG / Bing) + Google News, and for the technical area, arXiv + GitHub. No API keys required.
• Read — top results are fetched and extracted to clean text (redirects resolved, SSRF-guarded).
• Reflect — deep mode critiques its own draft for gaps and runs another round.
• Cite — the writer uses ONLY the sources it actually read, cites every claim inline [n], and says what's missing rather than inventing. Citations render as domain chips; click one for the supporting passage.

Modes & areas

• Modes: quick (search → read → cited answer) · deep (plan → gather → reflect → report).
• Areas: web · opportunity · technical (adds arXiv + GitHub) · competitive · client — each blends the right sources.

Run it

pip install -r requirements.txt
python app.py                    # http://127.0.0.1:7860  (web UI)
python agent.py "your question" --mode deep --area technical   # CLI
# Backend: local Ollama by default (free/private); set RESEARCH_LLM_BACKEND=claude + ANTHROPIC_API_KEY for cloud.
# Optional: RESEARCH_SAVE=1 banks each brief to a memory store so research compounds (local use).

Two tiers

• This hosted demo narrates with a hosted model, keyless search.
• GritAI Studio runs the whole thing on your own local GPU fleet — free per-run, fully private — and can bank every brief to your own knowledge base so research compounds over time.

---

Built by Robert Lucyk · GritAI Solutions · part of the 30-in-15 challenge.