demo env · onlineatlas · live demo

Ask the corpus.
Watch it answer_

Atlas answers only from a fictional corpus — twenty chunks you can read below. Every claim is cited inline as [Doc Title], and when retrieval comes back empty it says so instead of improvising.

20: chunks
4: retrieved
0: new deps

atlas -- retrieval mode

I am Atlas, a retrieval-augmented assistant.

I can only answer from the demo corpus — ask me about the eval numbers, the changelog, or anything in those documents.

01The corpus

6 docs · everything the assistant knows

Atlas Design Notes

×4

tone, citation rules, refusals

Atlas Runbook

×3

cold start, escalation paths

Atlas Eval Report

×3

48 golden questions, rubric

Atlas Changelog

×3

v0.1 → v0.3, title bonus

Atlas API Reference

×4

search_knowledge contract

Atlas FAQ

×3

what to try, what to expect

6 document types · 20 chunks total · the assistant reads all of them

02The spec

no embeddings, no vector store, no new dependencies

corpus20 chunksfictional, internally consistent

retrievallexicalterm frequency + title bonus

k4hard cap, no reranking

tools1search_knowledge(query)

dependencies added0no embeddings, no vector db

citationsinlineevery claim as [Doc Title]

on missadmit itno improvising from priors

swap point1 filelib/atlas/search.ts

the swap point

Retrieval lives in one file. When the corpus outgrows lexical scoring, that file is replaced and nothing else in the demo changes — the tool, the prompt, the panel, the eval set.

Ask the corpus.Watch it answer_