Question Answering

Ask a wiki page a question¶

Every wiki page carries an inline Q&A box. Ask a question, pick a model, and a coordinating agent answers from the same graph that backs the pages. Every claim carries a citation back to its file.

How the Agentic Wiki builds answers¶

flowchart LR
    Q([Question]) --> C["Coordinating agent"]
    C --> P1["Probe agent"]
    C --> P2["Probe agent"]
    C --> PN["..."]
    P1 --> G[("Knowledge graph")]
    P2 --> G
    PN --> G
    G --> A["One cited answer"]

Each probe agent takes one angle of the question. It enters the knowledge graph at the symbols and notes an ANN lookup ranked highest, then follows the edges out.

The fan-out is bounded. When the step budget runs out, the coordinator answers from what was found rather than running on. Questions that span modules gain the most, because one probe agent's evidence sits beside another's before a word is written.

Grounded, cited answers¶

[path:line-range] chips sit in the answer text. Click one and a source card expands with the code excerpt. Below the answer, Cited Sources and Retrieval details name every wiki page and source file the run read, plus the model that wrote it.

An Agentic Wiki answer to 'What is this project for?' generated with claude-sonnet-4-6: a left rail with a Summary card, an expandable Cited Sources list (project overview, agentic search engine, source capability graph, channels & integrations, API server), and a Retrieval details panel listing the pages accessed and the model used; the answer on the right describes the project's core capabilities and repository structure with inline wiki citation chips, and a follow-up question box sits at the bottom

Answers that teach the wiki¶

A Q&A run writes back. A durable fact from an answer is deposited as an anchored note, off the critical path. → The memory grows as the wiki is used

Ask from your agent fleet¶

Claude Code, Codex, Cursor or another Mewbo gets the same cited answer over the MCP server, without the console.

ask_wiki returns a cited answer for an indexed project. A long question returns an answer_id with status: "running".
get_wiki_answer replays an answer by its answer_id once it settles.
read_wiki_structure / read_wiki_page / list_wiki_projects browse the structure and pages directly.
submit_insight teaches the wiki a durable fact your agents found. → The memory grows as the wiki is used

The same operations sit on the REST API under /v1/wiki/*. The MCP server page has the full tool list and the REST API reference has the HTTP surface.