Construct a model of
anything — through conversation.
Ask a question. The system recursively decomposes it, investigates each piece with real reasoning, and builds a live graph of what it explored — not a finished answer handed to you, a trace of how it actually got there.
An investigation graph, not a knowledge base
Most tools that answer questions give you a paragraph and hide the work. Discovery.Ai does the opposite: every question gets recursively broken into sub-questions, each one investigated on its own, and the whole process is rendered as a graph you can navigate.
Zoom into any node to see what's actually known about it. Go deeper to trigger a fresh investigation. Switch lenses — economic, technical, historical — and watch the same entity get re-explored through a different frame entirely.
It's deliberately honest about what it is: a trace of how the system explored a subject, not a claim that the subject is now fully modeled. The graph grows exactly as far as you push it.
TRACE
Real sources, scored honestly — even when they're wrong
Every answer can pull from six real, named retrievers: Tavily, Semantic Scholar, arXiv, Open Library, the YouTube Data API, and Wikipedia. Each source becomes a claim with its own confidence score — not "how sure the model is," but how well that specific source actually answers the question.
A source that misses the point isn't hidden — it's attached and scored low, so you can see exactly what the system considered and correctly discounted.
What the system actually does
Recursive investigation
Every question is answered, decomposed one narrower sub-question at a time, or honestly flagged as a boundary — never planned as a batch upfront.
Four-outcome identity resolution
Discovering "Router" doesn't just reuse-or-create by name — REUSE, CREATE, AMBIGUOUS, and CONFLICT are kept distinct so the graph never silently guesses.
Typed relationships
Not everything is "decomposes_into" — routes_to, regulates, depends_on and more are extracted as their own decision, then canonicalized for consistency.
Click any node to inspect it
Every entity opens a real panel: the question that investigated it, why, and every source behind it — not a black box you have to trust.
Bring your own API key
Add a free Groq, Gemini, or Cerebras key in Settings and your investigations run on your own quota — with live status on whether the shared pool is up.
Switchable lenses
Re-investigate the same entity through an economic, technical, or any named dimension — the system re-frames, not just re-reads.
Session history
Every conversation is kept, switchable from the side rail, deletable when you're done with it — nothing is silently overwritten.
Google sign-in
Your investigations are yours — sign in with Google and every graph and session is scoped privately to your account.
Honest provenance
Ask "why is this here" on any entity and see exactly which question surfaced it — the graph never hides its own reasoning.
From question to graph
Ask
Type any question — as broad or specific as you like.
Decide
The agent answers directly, decomposes into one narrower sub-question, or flags a real boundary.
Investigate
Each sub-question is resolved on its own before the next decision is made, informed by what's now known.
Gather evidence
Real sources are retrieved and scored — honestly, including when they don't fit.
Resolve identity
New entities and relationships are only written when the system is genuinely sure they're new — or genuinely sure they're not.
Navigate
Zoom, go deeper, switch lenses, or compare — the graph is yours to explore.
What this doesn't promise
Being upfront about limits is part of the design. Full technical detail is in the documentation — the short version:
A trace of one investigation, backed by real evidence — not a claim that any subject is now fully or correctly modeled.
Shared providers can run dry at busy times, slowing or failing an investigation for reasons unrelated to your question. Your own API key fixes this.
Keyword search sometimes returns off-topic sources. They're scored honestly low, not hidden — the confidence score is the signal to read.
Genuinely competing explanations can get presented as complementary. Read the underlying claims yourself on contested topics.
Start investigating.
Sign in with Google and ask your first question.