See how Atomize finds a source sentence and answers from it
Atomize is a news app that answers questions from the articles it collects. Here you run one sample sentence through a miniature of that pipeline and see which choices keep the answer tied to its source.
In Atomize, the stages are: cut articles into pieces, search them by meaning and words together, have a model re-rank the top 25 to 60 matches, then have a model write the answer. A newer search path, in testing, re-ranks the top 10 and also flags numbers it can't trace to a source. This miniature runs the cut and the search in your browser. It describes re-ranking and the number check where they happen, and copies the answer from the source instead of calling a model.
Start with one nine-word sample sentence.
The condition in this sentence decides the answer, so the cut has to keep it with the rule.
Open the full labMore sample sources, an evidence workshop, meaning clusters, and a simulated worker shift.
Full lab
Load a set of sample sources, change how they are cut and searched, and check which source sentences back each answer.
Read a source parcel. Split it into sentences: did the refund condition survive? Compare with the starting setup, then try an exact identifier query.
Remove a source. Turn meaning down to zero. Retrieve one more chunk. Follow what changes across answers, briefs and search.
Order desk
A customer was billed twice. Can your factory find the rule and keep its conditions intact?
Loading sources…
The order could not be loaded.
Start freshSource documents · read or remove
Words → meaning → evidence
Choose an order to start the floor.
Decision bench
Whole documents keep conditions together. Sentences isolate a precise fact but can lose nearby context.
0 = exact word coverage. 1 = vector similarity. Neither score proves a claim.
| Chunk | Baseline score | Current score | Status |
|---|
Model & scoring detail
This teaching model mixes query-word coverage with normalized cosine similarity: weight × (cosine + 1) / 2 + (1 − weight) × coverage. No model weights download to your browser.
Demo model: all-MiniLM-L6-v2, 384 dimensions, with embeddings prepared ahead of time (this Evidence Workshop and the Meaning Garden). The Sentence Machine and Factory Floor use nomic-embed-text, 768 dimensions. Atomize production uses OpenAI text-embedding-3-small, 1536 dimensions.
Here the answer is composed by a rule, and the quote rules are a teaching rule written for these fictional missions. Atomize doesn't use this rule. It has a model re-rank the top 25 to 60 matches and write the answer from the top-ranked articles. A newer search path, in testing, re-ranks the top 10 and also flags numbers it can't trace.
Evidence inspector
Claims to check: a high rank is only a candidate. Open a parcel or citation to read the source words.
Demo composition · no live LLM
Meaning Garden
Whole source passages group by real cosine similarity. Moving a point on screen cannot change its group — the underlying score decides that, not pixels. Authored conflict links are shown separately from similarity.
| Chunk | Document | Source text | Group | Similarity and authored relationship |
|---|
Night Shift
A simulated pipeline: workers pull queued jobs, a broker can go down, and a cache can go stale. Nothing here reaches a real queue, cache or service. Every displayed value is computed by this turn-based simulation, not measured from production.
Scenario capacity is 2 provider slots. Set workers to 4 to see two workers wait while the shared resource stays full.
Raw event fields
| Layer | Teaching distinction |
|---|---|
| Long-running work | Railway hosts long-running Node workers in the reference architecture. |
| Broker | BullMQ uses Redis as its broker; this simulation does not contact either service. |
| Response and site cache | An in-memory cache plus Redis; a newer search path, in testing, keeps answers for 5 minutes. The cache can be fresh, stale, or refreshed independently of broker availability. |
| Scheduled refresh | Vercel cron jobs trigger refresh work, for example an RSS refresh every 15 minutes. |
| Transformer attention | A transformer attention KV cache is a separate model-runtime mechanism and is not simulated here. |