If you've got a vector file sitting next to your tabular feature set, the whole value of it collapses to one question: does row 4,812,009 in the embedding file actually correspond to the same parcel as row 4,812,009 in your model's feature table. Everything else, embedding dimension, what the vector encodes, how often it refreshes, is downstream of getting that join right. Get it wrong and the model trains on a mismatched vector for that row with nothing in the logs to flag it.
Picking the join key
Most AVM pipelines already have a canonical parcel identifier somewhere, usually a FIPS code concatenated with an assessor's parcel number (APN), sometimes a proprietary internal ID that was itself derived from APN at ingestion. Whatever you use internally, the embedding vendor needs the same key, in the same format, not a "close enough" version of it.
The failure mode here is almost never a missing parcel. It's formatting drift. APNs get stored with dashes in one system and without them in another. Leading zeros get stripped by a spreadsheet import somewhere upstream and never recovered. County FIPS codes get truncated to four digits when they should be five.
None of these show up as errors. They show up as a join that returns fewer rows than you expected, and a model that trains on a subset of parcels that doesn't match your actual portfolio. The row count looks fine at a glance; it's only off by the handful of counties where someone's export script drops a leading zero.
Before you run the real join, run a reconciliation pass: count distinct keys on both sides, check for duplicate keys on either side (parcel splits and re-plats are the usual culprit), and spot-check a sample of a few hundred records against the underlying address or lat/long. If you're joining at county or MSA scale, do this per-county, because formatting conventions aren't consistent even within a single state.
If your record ID isn't the assessor's APN, the vendor should still be able to key off it, but that means you're providing them with a crosswalk, and that crosswalk is now part of your pipeline's dependency graph. Version it. If your internal ID scheme changes (it will, eventually, usually during a core system migration), the crosswalk breaks and the embedding join thins out with no error thrown, just fewer matched rows each run until someone goes looking.
Wiring it into the feature pipeline
Once the key problem is solved, the embedding is just another feature block to merge in, structurally no different from bringing in a new tax-assessment extract. A few things specific to vectors are worth planning for up front:
- Treat it as a fixed-width block, not individual named features. You'll typically bring in the embedding as a single array column or as N numbered columns (
emb_0throughemb_N), not as hand-labeled features, because nobody, including the vendor, can tell you what dimension 47 "means" on its own. - Decide your null-handling policy before you need it. New construction and freshly re-platted parcels won't have a vector yet if they postdate the last annual refresh. Your pipeline needs an explicit fallback, usually zero-imputation plus a flag feature so the model can learn to downweight the missing-vector case rather than treating a zero vector as a real signal.
- Keep the embedding's vintage as a feature, not an assumption. A vector generated from imagery a year old behaves differently from one refreshed last month, especially after a parcel has had visible work done on it. A
vector_as_of_datecolumn costs you nothing and saves you a debugging session six months from now when valuations on recently renovated parcels look off. - Join once, upstream, into your feature store. Don't re-join the raw vector file at scoring time inside the model service. Materialize the join once into your feature store or batch table, same as any other sourced feature, so the join logic lives in one place and gets tested like the rest of your ETL.
The join itself is mechanically simple, a left join on a cleaned key, usually ten lines of code. The work that takes time is the key reconciliation before it and the null-handling policy after it. Budget for those, not for the join.
If you're at the stage of sourcing a vector file keyed to your own record IDs rather than building the embedding pipeline from raw imagery yourself, that's worth a look before you commit engineering time to the harder problem.
