Repeated terms
See a term once when several passages or books use it, see one idea under different names together, and say when two of them are not the same.
A group is the same term, with the same meaning, from several passages or books. If three books on sailing each have a card for "tacking", search shows "tacking" once, with how many books and versions it comes from, instead of three results in a row.
Grouping uses no model and costs nothing. It happens at the end of every index run, from the cards and vectors you already have.
Where you see groups#
- Search. A group shows as one result, led by its best match, with a line such as "3 books · 4 versions". Versions lists each version with its own book and passage, and opens its passage.
- The panel. The Groups tab lists the groups, biggest first, and each topic lists its key terms.
- Topic notes. At the next Find topics, a term with several versions in a topic is one line in the note with its versions folded under it. An index run never rewrites topic notes.
Only terms are grouped: the concepts of the nonfiction lens, the ideas of notes and the motifs of fiction. Claims, questions, cases and practices stay as they are.
One name, two groups#
A name is not enough. Two cards named "bearing" are one group only when they mean the same: a bearing as a compass direction and a bearing as a part of a winch stay two groups. The second shows an another sense badge, and each group names the other.
Correct a group#
On the panel's Groups tab, select a group.
- Not the same, on a version, takes it out of the group into a group of its own.
- Join with… joins the group to another one you pick, such as the two senses of "bearing" when you think they are one.
A correction takes effect at once, with no preview. It is two small files, kept in the collection's index, and it shows on the Runs tab, where Undo takes it back. Undo goes newest first. A correction you made before an index run can be undone only after that index run is.
Your answers are kept for every later index. When a card an answer names is gone, that answer is skipped and the rest still apply. The index run's summary says how many were applied and how many no longer apply.
One idea, different names#
Different names can be one idea: "reef" and "reefing", "VMG" and "velocity made good", "jib" and "headsail". After grouping by name, Topic Map compares the groups by meaning, still with no model and at no cost:
- Clear pairs go together on their own. These are a word and its plural or its -ed or -ing form, the same word spelled the British or American way, or initials written in capitals and their full name, when they are also close in meaning. The group then shows both names' versions, and lists the names it was put together from. Keep apart beside a name takes it back out for good (undo it on the Runs tab). Endings that make a new idea, such as "-ism", "-ist" or "-ity", never put two names together on their own.
- Unclear pairs wait for you. These are close in meaning, but nothing in the names settles it, or one name sits inside the other ("rot" and "root rot" may be one idea, or one may be a kind of the other). On the Groups tab, Same idea? shows each pair side by side with what each means. Same idea puts them together; Different keeps them apart for good. Both show on the Runs tab and can be undone there.
- Two senses of one name never join, not even through a third group close to both.
Ask the judge#
Ask the judge… under Same idea? puts the waiting pairs to a model, topic-map.judge in Model Hub. It asks each pair both ways round, because a model can favor whichever comes first, and the judge may say it isn't sure. Before anything is spent you see how many pairs, how many calls, the price and a spend limit. The run stops at that limit.
The judge merges nothing on its own until your own answers show it agrees with you. That takes at least 20 of the pairs it judged that you have also answered, with at least 5 "same" and 5 "different" among them, and 90% agreement over at least 5 of its "same" answers. It also needs a judge model that says how sure it is: one that gives every answer the same confidence can't be calibrated, so it only ever advises. Until then its view shows beside each pair, and What the judge decided lists what it said. A judge that says one is a kind of the other never merges them; that shows as a link. Its answers are kept in the collection's index, so grouping again reuses them without asking again. Undo on its run takes them back.
Show every card#
Setup → Search → Show every card turns grouping in search off: every card shows on its own. The tools and the API never collapse. They give each card its group, so a client sees which cards belong together (see Your collections in Claude).
When groups are out of date#
If the cards or their vectors changed since the last grouping, the Groups tab says "Groups are out of date" and offers Group again. An index made before groups existed offers Group cards. Search carries on as before, with every card on its own, until then. Neither costs anything.