Search-Box
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 forms hypotheses, extracts verbatim evidence, surfaces contradictions, updates beliefs — and streams a fully sourced answer while you watch every step live at www.search-box.ai.
6
links in the provenance chain (sentence → claim)
8
workspace packages in the monorepo
3
evidence stances: supports / contradicts / context
40–90%
differentiated per-claim confidence (benchmark run)
8× vs 4×
real vendor contradiction caught autonomously
0
fake progress animations — every event is real
Overview
Search-box.ai answers hard questions the way a researcher would, not the way a search engine does. Given a question it autonomously forms hypotheses, decomposes its own uncertainty, runs purposeful searches, opens sources, extracts verbatim evidence, surfaces contradictions, and updates beliefs — repeating until the evidence supports a synthesis. There is no fixed pipeline: Claude decides the research strategy through tool use, while the app enforces safety with hard budgets, Zod-validated tool contracts, an SSRF guard, and prompt-injection defense on all scraped content.
Claims and evidence are first-class objects: verbatim quotes tied to falsifiable claims with stances (supports / contradicts / context) and probabilistic confidence. Contradictions are treated as a research success, not an error state, and provenance is never lost — every answer sentence traces mechanically through sentence → [n] marker → citation index → source → evidence → claim.
The UI is event-sourced end to end: PostgreSQL holds the durable research state, the client replays the persisted event stream over SSE with Last-Event-ID recovery, and nothing on screen is a fake progress animation. Built on the Anthropic Claude API for reasoning and orchestration and Firecrawl for web infrastructure, organized as a pnpm monorepo of eight focused packages, and validated end-to-end on a real research problem about transformer KV-cache compression.
Key Features
No fixed pipeline
Claude decides the research strategy through tool use — understand the objective, form hypotheses, decompose uncertainty, pick the next action — instead of a canned query → scrape → summarize flow.
Claims & evidence as first-class objects
Verbatim quotes tied to falsifiable claims with stances (supports / contradicts / context) and probabilistic confidence — beliefs update as evidence arrives.
Contradictions are a success
Conflicting sources are surfaced, not smoothed over — the benchmark run caught a vendor blog claiming 8× "comparable" compression where the peer-reviewed paper only claims lossless behavior up to 4×.
Provenance never lost
Every answer sentence traces mechanically: sentence → [n] marker → citation index → source → evidence → claim. No orphaned citations, no invented references.
Event-sourced UI
The client replays the persisted event stream over SSE with Last-Event-ID recovery — every node on screen corresponds to a real backend event stored in PostgreSQL.
Safety enforced by the app
Hard budgets, Zod-validated tool contracts, an SSRF guard on fetches, and prompt-injection defense on all scraped content — the model proposes, the backend disposes.
How It Works
Understand & hypothesize
The engine restates the objective, forms explicit hypotheses, and decomposes what it does not yet know into addressable uncertainty.
Act on the web
Claude picks the next action — purposeful searches and source opens through the Firecrawl adapter, every tool call validated against a Zod contract and counted against hard budgets.
Extract & update beliefs
Sources yield verbatim evidence tied to falsifiable claims with stances and confidence; beliefs update, and gaps or contradictions drive the next iteration.
Persist every event
The ResearchState service writes claims, evidence, sources, and contradictions to PostgreSQL; a typed event protocol streams each step to the client over SSE.
Synthesize with provenance
The final answer streams live, every sentence carrying a citation that resolves mechanically through the index to a source, its evidence, and the claim it supports.
Tech Stack
Agent
Platform
Deployment
Highlights
- Validated on a real research problem — transformer KV-cache compression — in one autonomous session with no human steering
- Assigned differentiated confidence per claim (40% / 85% / 90%) instead of a single global score
- Caught a genuine contradiction: a vendor blog presenting 8× compression as "comparable" quality where the paper only claims lossless behavior up to 4×
- Every UI update corresponds to a real backend event — the research graph is a replay, not an animation
- Prompt-injection defense treats all scraped content as untrusted data, never as instructions
- Runs from the terminal too: pnpm research "your hard question here"
Explore Search-Box
Don't search the web — search the answer space — the full source is on GitHub.