firmmatchr 0.2.0

Multilingual normalization (breaking change)

normalize_company_name() previously applied German conventions only. It now covers German, French, Italian and English, which is what multilingual registers such as the Swiss one require.

• New lang argument on normalize_company_name() and match_companies(), accepting any of "de", "fr", "it", "en". Defaults to all four; narrow it for single-language data to strip fewer tokens. The default therefore strips more than in 0.1.x — pass lang = "de" for the old scope.
• Legal forms and noise words are now organised per language and cover ~109 tokens (sarl, sasu, eurl, societe, srl, snc, scarl, societa, llp, incorporated, succursale, niederlassung, …).
• New company_name_stopwords() returns exactly which tokens are stripped for a given set of languages, so the normalizer can be audited.
• Fixed the processing order: transliteration now runs before stop-word removal. Accented legal forms such as Société and Sàrl were previously never recognised on the first pass, because the patterns were matched against the still-accented string.
• normalize_company_name() is now idempotent — applying it to its own output is a no-op, so one pass always suffices. This follows from the ordering fix above: previously Nestlé Sàrl became nestle sarl on the first pass and only lost sarl on the second, so callers had to re-normalize in a loop until the result stopped changing. Such loops can now be replaced by a single call. A regression test locks the property in.
• Dotted acronyms are now joined rather than split: S.A.R.L. becomes sarl (and is stripped) instead of four stray letters s a r l. Dots after multi-letter tokens still act as separators, so Co. KG is unaffected.
• Apostrophes are now deleted rather than turned into spaces, so elisions and possessives join up: L'Oréal and Loreal both normalize to loreal, and McDonald's gives mcdonalds rather than mcdonald s.
• Identifying words are deliberately preserved: Deutsche Bank still normalizes to deutsche bank, not bank.
• Country names are no longer stripped (0.1.x removed deutschland and germany). They identify a firm as often as they pad it: removing them turned Credit Suisse into credit and Air France into air, either of which can then exact-match an unrelated company. Since a spurious exact match is trusted and never reaches the fuzzy stages or the LLM validator, keeping them is the safer default. Structural branch markers (Niederlassung, Succursale, Filiale, Standort) are still removed.
• New extra_stopwords argument on normalize_company_name() and match_companies() for corpus-specific filler — use it to strip geography anyway, e.g. extra_stopwords = c("deutschland", "schweiz").

Duplicate handling (breaking change)

normalize_company_name() is lossy by design: "Meier GmbH" and "Meier AG" both become "meier". Previously match_companies() aborted whenever normalization produced duplicate dictionary names, which made it unusable on real dictionaries without hand-written pre-cleaning.

• match_companies() now collapses duplicate normalized names instead of aborting. Each group is matched once against a representative row, and no original row is lost:
  • dict_id in the result is the representative (first) id of the group.
  • New dict_id_all column lists every id in the group, separated by "|".
  • New n_dict_ids column reports the group size (1 when unambiguous).
• New dict_crosswalk() returns the mapping from every original dictionary row to its representative, so collapsed entries always remain traceable.
• New expand_matches() expands a result to one row per original dictionary row, recovering the source strings behind a collapsed group.
• New on_duplicate argument. The default "collapse" is the behaviour above; "error" restores the previous strict abort.
• Duplicate query names — and duplicate query ids — are now handled too. Each distinct normalized query name is matched once and the result is fanned back out, so every input row still gets its own output row.
• Rows whose name is empty or NA after normalization are now dropped with a warning. Previously an empty string could “exactly match” every other empty string, producing spurious matches.

Other changes

• New dict_norm and query_norm arguments let you skip normalization on either side when the input is already clean.
• match_companies() now returns query_name and dict_name alongside the ids, so the result is readable (and ready for validate_matches_llm()) without joining the inputs back on. The set of returned columns has changed; code that relied on the previous three-column output may need updating.
• Fixed an error when the exact-match engine found no matches (the match_type column was referenced before it was created). All engines now return a consistently typed empty result.
• The matching engines no longer add helper columns (block, fts_q) to their inputs by reference.
• Removed LazyData from DESCRIPTION (the package ships no data/ directory); added URL and BugReports. Fixed a dead reference URL and retitled the package, which is no longer German-only.

firmmatchr 0.1.3

• New Features:
  • Added standard OpenAI-compatible support to validate_matches_llm enabling it to connect to local LLM instances (like Ollama or LM Studio).
  • Added an engine argument allowing choice between "azure", "openai", or "local".
  • Extended documentation to cover local LLM usage with code examples.

firmmatchr 0.1.2

• Fix CRAN Submission Issues:
  • Use single quotes for software/package names in DESCRIPTION.
  • Add references for ‘zoomerjoin’ and ‘SQLite’/‘FTS5’ in DESCRIPTION.
  • Ensure validate_matches_llm writes to tempdir() by default.
  • Ensure exported function examples are executable.

firmmatchr 0.1.1

• Fix CRAN check NOTEs:
  • Quote technical terms in DESCRIPTION to avoid spell check warnings.
  • Add cran-comments.md to .Rbuildignore.

firmmatchr 0.1.0

• Initial release.
• Features:
  • Exact matching
  • Fuzzy matching (zoomerjoin)
  • FTS5 search (SQLite)
  • Rarity weighted matching