How the map builds itself

To hand your agent every fact it needs, and to say when a list is complete, Limma files each fact at an address: who or what it is about, which property, and when. The names those addresses use are the map. You do not write it: Limma builds it from your own documents and records, and keeps it current as they change.

What is automatic today

Facts from tables and text. Every figure Limma reads, in a table or in a sentence, is filed by who or what it is about, which property it is and when, such as Acme Corp · revenue · FY2023, with the exact characters it came from. When two places give different numbers for one address, both are kept and each is marked contested (an address, contested).

Clause types, from your own documents. The clauses of your documents are sorted into types, such as a termination clause or a risk factor, and the types are induced from your workspace's own documents rather than from a list someone wrote. A new clause reuses a type already named when it is close to one, so one kind of clause keeps one name across your documents.

Record exports, mapped from the records themselves. Connect a CSV, Excel or JSON export from a CRM, an ERP or a directory and Limma drafts the mapping and its relations from the records: which fields hold dates, numbers, yes or no, a short list of values, or an id. No model is involved. You read the draft and confirm it once, and later syncs reuse it (a collection nobody wrote a pack for).

Labels in the document's own words. A fact is filed under the words your document prints, unless they are already a name in the map, letter for letter once case, punctuation and plurals are set aside. Every sentence is also given a short topic label that must share a word with its passage, and those labels are how a question about an idea, such as where do we discuss pricing risk?, is answered (ideas).

Names from three sources. The names the map uses come from:

• a built-in core: every country and the regions it belongs to, conventions such as DACH, the Nordics or EMEA, and every language with a two-letter ISO code, so a search for DACH finds a record in Austria (vocabulary);
• your own columns: a connected export's field names become names in the map, and a field that holds a short list of values gets that list from the values your records hold, growing with each sync. A column called revenue and a PDF that prints "Revenue" are one name;
• what Limma learns from your documents: the labels their facts are printed under, their clause types and their topics, as above.

Similar names are never merged without evidence. A label that only resembles a name already in the map is offered as a proposal and kept apart until somebody accepts it. Two records that look alike, "Ferrum Systems AG" and "Ferrum Systems Ltd" or two people with one name, are never made one because of their names. When you connect an export you say which of its fields identify a record (Limma drafts the questions from the field names). Records are then merged only on evidence that a different thing could not plausibly share, such as one registration number. Anything weaker is listed as a possible duplicate, and nothing is merged until somebody confirms it. No record is rewritten, so a merge is undone by one decision (one thing, entered twice).

The map is versioned. Every search and facts reply names the version of the map it read. The version moves whenever a name your documents carry is added, re-spelled or decided, so an agent that kept the map's names knows when they have changed (search).

What is set up with you

Typed rows read from sentences of a specialised kind, such as the roles and their dates in an expert network's profiles, are set up with you during onboarding.

Everything Limma does not do is listed in what it will not do.