Library & Assistant

ApocOS gives you two ways to ask, split by intent — both fully offline, both answering only from local sources (your indexed code and projects, the entity graph, this machine, and the installed knowledge corpora: man pages, RFCs, DevDocs, Wikipedia). Neither ever reaches the Internet to answer.

• Library (CLI: aios ask) — retrieval-first. For looking something up. It searches the offline corpora first, then answers grounded in the passages it found, with the source shown prominently beneath (and an Open full article link where an illustrated copy exists). If nothing in the library matches, it says so and answers from general knowledge, clearly labelled unverified — so you always know whether an answer is backed by a real source. This makes it trustworthy even on a small model, where the corpus, not the model, is the authority.
• Assistant (CLI: aios chat) — model-first. For conversation, reasoning, and doing things: follow-up questions, explaining code, planning a route, querying the map or the machine. It has the full tool loop, streams live with the deterministic breakdown of how it reached an answer, and can act (run commands / write files) when you tick actions. Under each answer, 📚 source passages lazily expands any corpus text it drew on.

Use the Library when you want a sourced fact; use the Assistant when you want a conversation or an action. (Manage the corpora themselves in Settings → Knowledge corpora; see The offline library.)

Looking something up — the Library

aios ask "the human heart"              # grounded lookup: answer + the source passages it cited
aios ask "how does TCP work?"
aios ask                                # a small read-only REPL if you give no question

On the dashboard, the Library tab is a search box: type a query, read the grounded answer, and the source passages appear beneath it — click Open full article ↗ to read the full illustrated Wikipedia/how-to article in the reader.

Conversation and actions — the Assistant

aios chat                               # a rich interactive REPL (history, completion, streaming)
aios do "add a healthcheck to the compose file"

The dashboard Assistant tab is the same pipeline, streamed live. It is read-only by default; tick actions (or use aios do) to enable action tools — it can then run commands and write files, but each action is gated by your approval before it happens. Use it for tasks, not just questions.

”Where is the nearest…?”

Questions of the form “where is the nearest / closest X” — water, a hospital, fuel, shelter, food, an ATM — are answered directly from the map, without the model choosing anything. ApocOS reads the kind of place you asked for, runs the offline search itself, and writes the answer from the result; if you also ask how long it takes to get there (“…and how long to walk?”), it routes over the real road and footpath network and reports that time instead of estimating one.

This matters on a small model. The map answer is the one your life might depend on, so it is produced by code, not inference — which means it is identical on every ApocOS unit regardless of the model installed, it arrives in about a second, and it cannot invent a place that is not there. The ranked list of the other nearby matches, and the route, are shown beneath the answer. Anything more open-ended (“the best place to camp near here”) still goes to the model and its full tool loop.

Keeping the graph current

Both surfaces reason over an entity graph built from your files, code, and git history:

aios reindex [dir]     # (re)index a directory into the graph
aios watch [dir]       # keep the graph live — reindex on file changes (Ctrl-C to stop)

Good questions to try: “which projects use MQTT?”, “explain this codebase”, “find every SQL migration touching invoices”, “why does this service exist?”.