FAQ

Models and providers

Does the Node ever use GPT or Claude?

No. The Node — the chat — runs Qwen and only Qwen. Either locally via Ollama at the variant matched to your hardware tier, or via OpenRouter at a larger Qwen variant (up to roughly 35B). Top-tier proprietary models (GPT-5, Claude Sonnet 4.6) are reserved for Forge, the pack-building tool. The Node's contract with the network is Qwen-only.

Why does Forge use top-tier cloud models then?

Building a pack is a one-shot synthesis: you ingest knowledge once and query it locally on Qwen forever. The marginal cost of a stronger model on the ingest side is paid once; the quality benefit accrues to every future query at zero cost. Forge defaults to GPT-5 Mini and offers GPT-5 and Claude Sonnet 4.6 for material that justifies the spend. See Forge for the full picker.

Why Qwen specifically?

Open weights, strong multilingual coverage, unusually capable per parameter (which makes lower-tier hardware genuinely useful), and a multimodal variant available for image and audio inputs. Standardising on one model family gives the network predictable behavior and consistent multilingual support without locking the project to any vendor.

Privacy and offline use

Where do my chats go?

Nowhere except your own machine. They are stored locally in .occ_chats.json inside the OCC installation folder and never sent to any server. Use Settings → Data → Clear all chats to wipe them — this action cannot be undone.

Where are my files stored?

• upload/ — files attached via the paperclip icon.
• workspace/ — files OCC writes or executes code on.
• expert-packs/ — knowledge packs on disk.

All three live inside the OCC installation directory. Nothing in them is uploaded.

How do I keep a pack private?

Build it with Forge but do not submit it to the public catalog. In Settings → Knowledge Source, switch to Local only when querying — the broker is not contacted, and the pack stays on your machine.

How do I run OCC fully offline?

Flip Settings → Knowledge Source to Local only. As long as Ollama and your model are installed locally and you have at least one local pack, OCC will answer without any network access. Web tools (web_search, fetch_url) will of course fail without internet, but the rest of the pipeline runs fine.

Does the broker see my questions?

No. Retrieval sub-queries are translated into pack-language keywords before being sent; the broker sees keywords, never the original question. When a peer Critic runs, the broker routes the encrypted payload — it has no key to decrypt.

Hardware and performance

Do I need a GPU?

No. OCC adapts to whatever hardware is available, including CPU-only laptops. The setup script detects your VRAM and picks the right Qwen variant. CPU-only machines fall into the micro tier and run qwen3.5:2b. See Getting started for the full tier table.

What if Ollama crashes or won't start?

OCC tries to start Ollama automatically on launch. If it cannot, the loading screen shows a download link. You can also run ollama serve manually from a terminal. Once Ollama is running, restart the Node from the desktop shortcut.

Why does a chat sometimes take longer than expected?

One of three things is happening:

• The model is being loaded into VRAM (cold start, several seconds on the first query of a session).
• A peer Critic is being consulted over the network.
• A tool or skill is doing multi-step work (e.g. web_research opening multiple pages).

The status line under the routing badge tells you what is happening at each step.

Routing and answers

What do the routing badges mean?

The five labels:

• chat — conversational or tool-driven, no retrieval performed.
• server + local — full deliberation against the community packs on the broker, Critic ran locally.
• local · private — full deliberation against your local private packs only, broker not consulted.
• offline · no knowledge — broker unreachable, answer from the base model only with no curated grounding.
• network — full deliberation, Critic delegated to a stronger peer Node.

See The Node → Routing badges for the full description.

Why does some answer say "offline · no knowledge"?

OCC could not reach the broker (network down, broker temporarily unavailable). The answer is from the base model only — no curated grounding. Try again later, or switch to local mode if you have private packs.

Why do answers cite "the source does not specify"?

Because the retrieved pages did not contain the specific fact you asked about, and OCC's faithfulness rules forbid inventing specifics. The Critic flagged the gap and the Synthesizer chose accuracy over plausibility. Try a related pack, or expand the pack with Forge if you have authoritative sources for that specific.

Packs and Forge

Can I edit a pack by hand?

Yes. A pack is a folder of normal markdown files — open any file in a text editor and edit it. Forge will respect your edits on the next run, and Lint will flag structural issues if your edits broke any conventions. See Anatomy of a pack.

What happens if my Forge build is interrupted?

Use the Resume interrupted mode in Forge. It keeps the pages already written, skips sources recorded as complete in wiki/log.md, and processes only the remaining ones. Use after a crash, a timeout, or a manual stop.

Can I run Lint without rebuilding?

Yes. The Lint section in the Forge panel can be invoked on any pack at any time, independently of a build run.

Contributing

How do I contribute code?

Open a pull request on github.com/VickyWenSZ/occ. The project is MIT-licensed and all design discussion happens in the open.

How do I propose a new source domain?

Source domains are the list of websites the community accepts as inputs to Forge. Adding one requires a pull request to the source registry — the community reviews and votes. See Governance for the process.

How do I publish a pack?

Build it with Forge, then follow Submitting a pack. The pack enters a community review queue. Approved packs are deployed to the broker and become available to every Node in the network.