"Ask the rules": a model answers a rules question from the loaded packs, with citations (later release) #1011

New issue

Open

opened 2026-10-01 20:38:10 +00:00 by claude-bot · 1 comment

claude-bot commented

2026-10-01 20:38:10 +00:00

Collaborator

The owner's answer to Q138 (G2), 2026-10-01: (b): keep "ask the rules" as its own issue for a later release, and close #553. To be recorded as D266.

#553's lookup half is built: packs are imported by someone who affirms they may, and the rules are searchable from the app bar, the Prep drawer, the wiki's Rules & bestiary chapter and Table mode's Find, with a hit shown as a stat block or a cited rule that can go on the shelf.

This issue is the part that was never scoped: a box where someone asks a rules question in words and the configured model answers from the retrieved rules text, citing the entries it used.

To scope before building (not for v4.5.0):

Retrieval: which packs a campaign's question may draw on, and how many entries fit the model's context window.
The risk the question named: a confident wrong ruling. Every answer cites its entries, says when the loaded rules don't cover the question, and never answers from the model's own memory of a ruleset.
Who may ask (players, or GMs only), where the box sits, and whether an answer can be kept on the shelf.
AI use: counted in the campaign's usage and credits like every other call.
Licence: only text the server already holds from an imported pack is sent, and the Privacy page's "what goes to a model" paragraph (D269) covers it.

Not scheduled.

**The owner's answer to Q138 (G2), 2026-10-01: (b)**: keep "ask the rules" as its own issue for a later release, and close #553. To be recorded as D266. #553's lookup half is built: packs are imported by someone who affirms they may, and the rules are searchable from the app bar, the Prep drawer, the wiki's Rules & bestiary chapter and Table mode's Find, with a hit shown as a stat block or a cited rule that can go on the shelf. **This issue is the part that was never scoped:** a box where someone asks a rules question in words and the configured model answers from the retrieved rules text, citing the entries it used. **To scope before building (not for v4.5.0):** - Retrieval: which packs a campaign's question may draw on, and how many entries fit the model's context window. - The risk the question named: a confident wrong ruling. Every answer cites its entries, says when the loaded rules don't cover the question, and never answers from the model's own memory of a ruleset. - Who may ask (players, or GMs only), where the box sits, and whether an answer can be kept on the shelf. - AI use: counted in the campaign's usage and credits like every other call. - Licence: only text the server already holds from an imported pack is sent, and the Privacy page's "what goes to a model" paragraph (D269) covers it. Not scheduled.

claude-bot referenced this issue

2026-10-01 20:40:37 +00:00

[Feature] Rules and bestiary lookup at the table and in Prep #553

claude-bot referenced this issue

2026-10-01 20:48:26 +00:00

Record D254–D269, the owner's answers of 1 October (#737) #1012

claude-bot added this to the v4.8.0 - A Living World Fills In milestone

2026-10-27 17:53:40 +00:00

claude-bot modified the milestone from v4.8.0 - A Living World Fills In to Backlog - Speculative / Long-Term

2026-11-06 01:59:33 +00:00

claude-bot commented

2026-11-06 01:59:44 +00:00

Author

Collaborator

Moved to the backlog (owner, 2026-11-06), out of v4.8.0. A proposed design was written up first; it's recorded here so nothing is lost.

Proposed shape

An Ask option next to Search in Find at the table, Prep's Rules & bestiary drawer and the rules library.
Answers carry numbered citations that open the entries in the existing preview, and can go on the shelf with their citations.
Retrieval: full-text search over the campaign's visible packs, with the same visibility rules as browsing (no GM-only entries for players, no disabled packs, no other campaigns' packs). Take the top matches that fit the provider's declared context window.
Answer only from those entries, then validate (the #1280 pattern):
- every citation must be an entry the model was given;
- quoted wording must appear in that entry;
- an answer with no valid citation is never shown.
Not covered: "The loaded rules don't cover this", plus the closest entries. Never an answer from the model's memory of a ruleset.
The campaign's own model, counted in usage.

Open questions for the owner

Players as well as GMs (with a campaign switch), or GMs only?
Where first: the table and Prep, or the library as well?
One question at a time, or follow-ups?
Before building screens, test the local model against a set of real questions with known answers. How many entries fit, and is the quality good enough?

**Moved to the backlog (owner, 2026-11-06),** out of v4.8.0. A proposed design was written up first; it's recorded here so nothing is lost. **Proposed shape** - An **Ask** option next to Search in Find at the table, Prep's Rules & bestiary drawer and the rules library. - Answers carry numbered citations that open the entries in the existing preview, and can go on the shelf with their citations. - **Retrieval:** full-text search over the campaign's visible packs, with the same visibility rules as browsing (no GM-only entries for players, no disabled packs, no other campaigns' packs). Take the top matches that fit the provider's declared context window. - **Answer only from those entries, then validate** (the #1280 pattern): - every citation must be an entry the model was given; - quoted wording must appear in that entry; - an answer with no valid citation is never shown. - **Not covered:** "The loaded rules don't cover this", plus the closest entries. Never an answer from the model's memory of a ruleset. - The campaign's own model, counted in usage. **Open questions for the owner** 1. Players as well as GMs (with a campaign switch), or GMs only? 2. Where first: the table and Prep, or the library as well? 3. One question at a time, or follow-ups? 4. Before building screens, test the local model against a set of real questions with known answers. How many entries fit, and is the quality good enough?