arizuko › products › Taxonomy chat (Marble)

Taxonomy chat (Marble)

“What does a child need to learn before fractions?” — the agent traces the prerequisite graph of an open education dataset and answers in plain words. One folder, one bundled dataset, a public chat widget anyone can open.

system shape

This is a read-only knowledge bot over a fixed dataset. Where the supermarket agent searches the live web each turn, this one queries a dataset that ships inside the folder — no outbound fetch for data, no per-user state. The six primitives at their leanest:

• Event. A message from the public web-chat widget lands as a row in the folder’s inbox. No platform adapter, no login — a /chat/<token>/ link is the whole front door.
• Routing. One rule maps the widget’s JID to the single marble folder. Every visitor shares it.
• Agent (folder-as-data). The folder holds a CLAUDE.md (who the agent is, how to answer, how to cite) and a taxonomy/ directory of JSON — topics.json, dependencies.json, clusters.json. The agent answers by running jq over those files: trace a prerequisite chain, list what a topic unlocks, filter by subject or age.
• Authorization. A single public chat token. Anyone with the link can ask; no per-user ACL rows.
• Turn. One ephemeral container per question. The agent reads the JSON, walks the graph, and replies with topic names cited. Nothing is written back.
• State. The shared DB (per-turn message rows) and the bundled dataset on disk. No per-user sub-folders, no memory files. The dataset is reference, not scratch — nothing accumulates between visitors.

example

web chat · public link

  visitor  what does a child need before fractions?
  marble   Before fractions, these foundations come first:
           whole-number place value, equal sharing / partitioning,
           multiplication as equal groups, and division as sharing.
           Each is a prerequisite edge into "Introduce fractions".

  visitor  what's the deepest chain in the taxonomy?
  marble   Understanding Earth's atmosphere (for 10-11-year-olds)
           requires learning 9 other topics first - the longest
           prerequisite chain in the graph.

The taxonomy behind this example is the Marble Skill Taxonomy v1: 1,590 K–5 micro-topics wired into a 3,221-edge prerequisite graph across 8 subjects. It is the open-data twin of the interactive curriculum graph on withmarble.com/curriculum — same dataset, an agent you can talk to instead of a browser you click.

where this sits — a bundled-dataset bot

Like the supermarket agent, this is the minimal end of the spectrum:

• No per-user memory. Every visitor starts fresh.
• No folder tree. One level, one folder.
• No grants table. The chat token is the fence.
• No web-search egress. The dataset is on disk; the box never leaves for data, only for the model call.

The one thing it adds over the supermarket shape is a bundled knowledge base and a deliberately cheap model. A public bot answering lookup questions doesn’t need a frontier model — pin it to Haiku and trim its skills, and each answer costs little and lands fast. The setup guide walks both tuning steps.

data & trust

• Your server. The dataset and message history live under /srv/data/arizuko_<name>/. Backup is one tar.
• One outbound path. Anthropic, for model calls. There is no web-search egress — the taxonomy is bundled, so the agent answers from the folder, not the internet.
• No user data collected. No per-user memory, no linked identities, no email. The only persistent state is the message log.
• Licensing. The Marble Skill Taxonomy is © Marble (withmarble.com), released under CC BY-SA 4.0. Keep the attribution in the agent’s CLAUDE.md and cite it in replies.

setup

arizuko group <instance> add web:<folder> <folder>   # register the public group
cp taxonomy/*.json /srv/data/arizuko_<instance>/groups/<folder>/taxonomy/
# pin the model to Haiku + trim skills at /dash/groups/<folder>/settings
arizuko token <instance> issue chat <folder>    # mint the public chat link

Full step-by-step — register the group, drop in a dataset, write the brain, pin the model, trim skills, mint the link — is in the operator setup guide.

go deeper

• primitives — the minimal recomposition: one folder (Agent) with a bundled dataset, one route (Routing), a jq-over-JSON Turn, no per-user State
• web chat tokens — mint a public widget URL without a platform adapter
• supermarket agent — the same minimal shape, but searching the live web instead of a bundled dataset
• chat SDK — embed the widget in any page
• deployment guide — first instance, first channel, first token