Fuzzy join¶

Two tables describe the same customers, and neither shares a key with the other. The CRM says Mueller GmbH, the sales sheet says Mueller Gmbh., and Join tables matches neither of them.

Fuzzy join matches rows that are similar rather than identical.

Opening it¶

Data → Fuzzy join..., with at least two tables open. There is no default keyboard shortcut; one can be assigned under Settings → Shortcuts.

Setting it up¶

Pick the left table, whose rows are kept, and the right table, which is searched for a partner. Then say which columns to compare: pick one pair, or several when a single column is not distinctive enough, in which case their scores are averaged.

The measure¶

Measure	Good for
Edit ratio	Typos and small misspellings.
Jaro-Winkler	Names, where a shared beginning counts for more.
Token set	Reordered or differently punctuated words.

These are the same measures Find near-duplicates uses within a single table, not a second implementation of the idea.

Values are compared as normalised text: lowercased, with runs of spaces collapsed and punctuation dropped. That is what lets ACME Ltd. meet acme ltd.

The threshold¶

How similar two values must be to count as a match, from 0 to 1. 0.85 is a sensible start: raise it when you get matches you do not believe, lower it when obvious pairs are being missed.

Blocking¶

Without a blocking column every left row is compared with every right row, which grows with the product of the two row counts and is why a row cap exists. Naming a column that must match exactly, such as a country or a postcode, restricts the comparison to rows that already agree there. It changes which pairs are compared, never which of them match, so it makes a large join possible rather than approximate.

Reading the result¶

The result opens in a new tab with the left columns, the right columns, and two more per step:

match_score_N is how similar the matched pair was. Empty when the row found no partner.
ambiguous_N is true when the runner-up scored nearly as well. Those are the rows to check by hand: the join picked one, but it was a close call.

The status bar reports how many rows matched and how many are ambiguous, and says so when the row cap was reached.

More than two tables¶

Add another table joins the result to a third, and so on, folding left to right. Each step gets its own match_score and ambiguous columns rather than one score for the whole chain, because matching against an already fuzzy result compounds the error, and a single number would hide where it came from.

Ceilings¶

Each left row keeps one partner, the best scorer. A row that genuinely matches two right rows is not expressible.
Values are compared as text. There is no numeric or date tolerance.
The row cap applies per side and is only reached without a blocking column. The status bar says when it bit.
There is no accept/reject review of individual matches yet. Use ambiguous_N to find the ones worth a look, and filter or edit the result tab.

Elsewhere in Octa¶

octa --fuzzy-join crm.csv --fuzzy-join-file sales.csv \
     --fuzzy-on customer=account

See octa --fuzzy-join and the fuzzy_join MCP tool, which the in-app Assistant can also call.