Resolve a messy customer file

The same company appears three times in the export: Harbor & Vine Catering, Harbor and Vine Catering LLC, HARBOR & VINE CATERING, each row writing the streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels. a different way. String comparison never joins them, and the standard repair is an abbreviation table that grows every time a new spelling arrives. Meanwhile the customer count in every report is wrong by however far the duplicates run.

What changes

One pipelinestaged pipelineMailwoman's runtime architecture: a sequence of pure-function stages (normalize → query-shape → locale-gate → kind-classifier → phrase-grouper → classifier → decoder) connected by typed handoffs. Each stage is published as its own npm package. runs all three steps. Rows are parsed into components, each address is geocoded to a coordinate, and the matcher blocks on that coordinate before it scores anything. That ordering is what links 11 W 53rd St to 11 West 53rd Street with no rule written for the abbreviation: two rows that resolve to the same point get compared, whatever they call the streetstreetThe named linear feature along which house numbers are ordered. Decomposes into a name plus street affixes; one of the Tier 2 fine labels..

On the twenty-row example file, the three rows above land on one coordinate and the name normalizer has already collapsed their three spellings to one canonical string. With the link threshold read off the measured pair weightsparameterA single learned number inside a model — one weight or bias. Mailwoman's encoder has roughly 30 million of them; training is the search for good values., twenty rows resolve to thirteen entities — the count of distinct organizations in the file.

What you still carry

The defaults are conservative, and on a file that size the result is that nothing links at all: the same twenty rows come back as twenty entities at the shipped threshold. That is the safer direction, because a missed link leaves two entities where one belonged while a false link fuses two real organizations and is harder to catch downstream. It does mean measuring before you trust the output.

Match messy records to one entity each is that measurement: score two pairs whose answer you already know, read the gap, and set the threshold inside it. The same page names the second limit in the same breath — single-file mailwoman registry run cannot reach that setting, so a small file goes through the library instead.

Geocode a CSV of customer addresses is the ingest and geocode half of the pipelinestaged pipelineMailwoman's runtime architecture: a sequence of pure-function stages (normalize → query-shape → locale-gate → kind-classifier → phrase-grouper → classifier → decoder) connected by typed handoffs. Each stage is published as its own npm package.. Each page publishes the CSV it runs against, so every count above is one you can reproduce on your own machine.

Next

• Your first ten minutes — the install and the gazetteergazetteerA geographical index that maps place names and postcodes to real-world coordinates. Mailwoman uses a custom-built Who's On First (WOF) SQLite database as its gazetteer — the 'atlas' half of the grammar/atlas architecture. the pipelinestaged pipelineMailwoman's runtime architecture: a sequence of pure-function stages (normalize → query-shape → locale-gate → kind-classifier → phrase-grouper → classifier → decoder) connected by typed handoffs. Each stage is published as its own npm package. resolves against.
• Pricing — one figure per legal entity, whatever the file size.