# Search-box.ai

**Don't search the web. Search the answer space.** Search-box.ai is a multi-step, agentic web research engine. Given a hard question it autonomously forms hypotheses, decomposes uncertainty, runs purposeful searches, opens sources, extracts verbatim evidence, surfaces contradictions, updates beliefs — and streams a fully sourced answer while you watch every step live. Built on the Anthropic Claude API (reasoning & orchestration) and Firecrawl (web infrastructure). PostgreSQL holds the durable research state; every UI update corresponds to a real backend event. --- ## Screenshots *A completed research session — the question blooms into objectives and claims, every node driven by a real backend event:* Live research session: bloom graph, stats bar, timeline and streamed answer

Live research session: bloom graph, stats bar, timeline and streamed answer

_{Home — ask a question that deserves real research}

_{Mobile-first, one-handed}

--- ## Validated on a real research problem Benchmark question: *"Can transformer KV cache be compressed by an order of magnitude without seriously harming model quality?"* — one autonomous session, no human steering:

The engine assigned differentiated confidence per claim (40% / 85% / 90%), and caught a genuine contradiction: a vendor blog presenting 8× compression as "comparable" quality for a method whose peer-reviewed paper only claims lossless behavior up to 4×. ## How it works ``` question → understand objective → form hypotheses → decompose uncertainty → decide next action → use web tools → inspect evidence → update beliefs → identify gaps/contradictions → repeat → synthesize evidence-backed answer ``` - **No fixed pipeline** — Claude decides research strategy through tool use; the app enforces safety: hard budgets, zod-validated tool contracts, SSRF guard, prompt-injection defense on all scraped content. - **Claims & evidence are first-class objects** — verbatim quotes tied to falsifiable claims with stances (`supports` / `contradicts` / `context`) and probabilistic confidence. - **Contradictions are a research success**, not an error state. - **Provenance is never lost** — every answer sentence traces mechanically: `sentence → [n] marker → citation index → source → evidence → claim`. - **Event-sourced UI** — the client replays the persisted event stream (SSE with `Last-Event-ID` recovery); nothing on screen is a fake progress animation. ## Quick start ```bash pnpm install cp .env.example .env # ANTHROPIC_API_KEY, FIRECRAWL_API_KEY, DATABASE_URL pnpm migrate # apply PostgreSQL schema pnpm dev # web app on http://localhost:3000 ``` Run a session from the terminal: ```bash pnpm research "Your hard question here" ``` ## Repository structure ``` apps/web Next.js app (UI + API routes + SSE) packages/agent orchestrator loop, prompts, tools, session runner packages/research ResearchState service (claims, evidence, sources, contradictions) packages/firecrawl Firecrawl v2 adapter (search, scrape) + SSRF guard packages/anthropic Anthropic Messages API adapter (streaming) packages/events typed research event protocol packages/db PostgreSQL schema, migrations, repositories packages/shared shared types, zod schemas, budgets, ids deploy/ ngrok config + start script (node m3u96a) docs/ architecture, agent loop, research state, event protocol ``` Full design notes in [`docs/architecture.md`](docs/architecture.md). ## Deployment Production runs on node `m3u96a` behind an ngrok tunnel: **https://www.search-box.ai** — see [`deploy/`](deploy/). --- ## Author **Simon-Pierre Boucher** 📫 [contact@spboucher.ai](mailto:contact@spboucher.ai)