Ask your campaign a question and get an answer with citations
Grounded question answering over one campaign’s own transcripts, wiki, recaps and relationship graph. If your sessions do not contain the answer, it says so.
It answers from your campaign, not from the internet
Ask a question in plain English and the system retrieves the relevant material from that one campaign — transcripts, wiki entries, recaps, and the relationship graph — then answers from what it retrieved. Every answer carries citations, and each one links to the wiki entry, session or recap it came from.
The citations are the point. An uncited AI answer about your campaign is a guess you cannot check, and checking is the entire difference between a useful tool and a confident liar. When nothing relevant is retrieved, the honest "I could not find this" is returned without calling the model at all — which also means a failed question does not count against your daily allowance.
How it finds things
Retrieval runs on two legs at once. A keyword and graph leg searches your campaign directly, alias-aware, so a question about "Thor" reaches the entry filed under "Thordak Ironforge". A semantic leg matches on meaning rather than wording, which is what catches the question phrased nothing like the transcript.
The semantic leg is additive: if it is unavailable for any reason, the search quietly degrades to keyword-only rather than failing. The index is kept current as you edit the wiki, so entries you rewrote stop being cited in their old form and entries you deleted stop being cited at all.
Conversations, not one-off queries
Threads are saved. You can ask a follow-up that depends on the previous answer, come back tomorrow and pick the thread up, and browse earlier conversations for a campaign. "Ask about this entity" from a wiki entry, and "Ask AI" from search, both hand a question straight into the same view.
Untrusted by design
Retrieved campaign text is passed to the model wrapped as reference data with explicit instructions to ignore any instructions inside it. That matters more than it sounds: a transcript is a recording of people talking, and if a player reads a sentence aloud that looks like a command to an AI, it lands in your transcript. Treating your own campaign data as untrusted input is the correct posture.
What this does not do
It does not know the rules. This is not a rules lookup and it will not tell you how grappling works — it only knows what happened in your sessions. D&D Beyond is the right tool for rules.
It cannot answer about a session that was never recorded or imported. The scope is strictly one campaign’s own data, and it does not read across your other campaigns.
There is a daily question cap on every paid tier below Party. It exists because each question is a real model call, not because we want you to upgrade for the sake of it.
Which plans include it
Included on every plan, with a daily cap: 5 questions a day on the free Adventurer plan, 50 on Dungeon Master, 200 on Guild Master.
The rest of what it does
- Session transcription that already knows your characters’ names
Upload a session recording and get a timestamped transcript. Your wiki’s character and place names are fed to the model, so fantasy nouns survive.
- A campaign wiki that is written by the session, not after it
NPCs, locations, items and factions are extracted from your session transcript into a linked wiki you can edit, share with players, and search.
- Session recaps you can read at the table without editing them first
A "previously on" recap generated from your session transcript, with key events, decisions and cliffhangers. Regenerate, hand-edit, and keep old versions.
- A relationship map of everyone your party has tangled with
See how NPCs, factions and locations connect — built from your session transcripts rather than drawn by hand, and editable where the AI got it wrong.