Use Wikipedia to help your local AI answer questions
If you use a local LLM to research a topic, a model that supports tool calls can search Wikipedia and read a relevant article during your chat. It can use the article to inform its answer. Naidan uses Wikipedia’s official API for searches and page text, so you do not need to set up a Wikipedia API key or access token. Choose one-time approval to review each search and page request before it is sent.
Open Naidan and choose a chat with a model that supports tool calls.
If you still need to set up a model, see browser model setup or the local engine connection guide. For a model to try with tool calls, choose Qwen3.5 9B (Q4_K_M GGUF).
1. Open the chat’s tool settings
In the chat input, select Tools. The tool sheet opens above the input. Wikipedia starts off in a new default setup.
2. Turn on Wikipedia
Before turning on Wikipedia, note that it also enables Shell and its Naidan file mount so the model can read parts of a long article saved in the browser. Current chat limits which chat history text Shell can read. The mount also makes saved files readable, including image attachments from other chats. It does not select a folder on your computer. Turn on Wikipedia, then close the tool sheet. With tool config persistence off, this temporary change lasts until you reload or close this page.
For details about what Shell can read, see how to read chats and saved files through Wesh sysfs.
3. Ask for a Wikipedia lookup
The Curated hosted version permits Wikipedia in its CSP connection allowlist. The Local-only version blocks external Wikipedia requests, even when the tool is enabled. Compare hosted connection policies
Send a specific request, for example: “Search Wikipedia for the Northern Lights, read a relevant page, and explain the cause.” If the model calls the search tool, check the proposed keyword in the approval panel; do not send private text or IDs as a search. Select Allow once when it matches what you meant to share. Allow for this chat and Allow globally can skip later approval for the same search action in their scope. Deny stops that request.
4. Review the page request
Search returns candidate titles and page IDs. Reading a page uses a separate approval action from search. Check the title and page ID when its approval appears, then select Allow once if you want the page text fetched. A prior approval saved for this page action can skip its later prompts.
What the model receives
Search returns titles and page IDs, not article text. The page tool returns a plain-text extract. For a long extract, Naidan saves the text in browser storage so Shell can read the part the model needs. Text returned directly by the page tool, or a portion read by Shell, becomes input to the selected chat model; with a server model, it is sent to its configured endpoint. The model’s answer is based on the result it uses, so check the source article if accuracy matters.
If there is no tool call, confirm Wikipedia is on in this chat and choose a model that supports tool calls. An enabled tool cannot force every model to use it.
Why Wikipedia does not receive Naidan’s site Origin
In the hosted version, the Wikipedia request does not carry the HTTP Origin of the site where you opened Naidan. Naidan calls this request path privacy fetch. Wikipedia can still receive the search words or page ID, your IP address, and browser information such as User-Agent and Accept-Language. This does not make the request anonymous.
The hosted version sends the request from a sandboxed, opaque-origin iframe, so the browser may send Origin: null instead of the site’s address as the HTTP Origin header. The request omits credentials and referrer. Its Wikipedia route permits only supported Wikipedia API URLs. In the standalone file version, the validated request is sent directly rather than through that iframe.