3,072-dimension embeddings: the maths under GoGee's product search

Why 3,072 dimensions, stored as halfvec

The catalogue started on a 1,536-dimension vector column and was migrated to 3,072 dimensions using the half-precision vector type. Higher dimensionality preserves more nuance between products that differ in one attribute, a 5-litre versus 10-litre version of the same item, while half-precision storage keeps the index small enough to stay fast.

The index is HNSW with cosine operators, which is what allows an approximate nearest-neighbour lookup across thousands of SKUs to return inside a page load rather than a spinner.

What actually gets embedded

A product is embedded from a composed text blob rather than its name alone: product name, AI-enhanced name, short and long descriptions, supplier attributes, colours, materials and styles, category names, custom fields, variant labels and SKU. Where product photos have been captioned by the vision model, that caption is included as a visual description line.

This is the single biggest lever on search quality. Thin product data produces thin vectors, and no model choice compensates for a catalogue where half the products have a name and nothing else. It is also why the catalogue-content tooling and the search tooling are sold as one platform rather than two.

The similarity function is a database function

Matching happens in Postgres through a similarity RPC that accepts the query vector, a match count, a minimum similarity, and optional category, brand and price-range filters. It returns products with a similarity value computed as one minus cosine distance.

Because filtering happens inside the same query as the vector match, a price bound or category restriction narrows the candidate set at the database level instead of being applied to an already-truncated result page, which is the usual reason vector search appears to "ignore" filters.

• Threshold is per-call: looser for photo search, stricter for feature-specific queries
• Refinement passes can drop the floor to zero and re-rank within a known set
• Category, brand and price filters are arguments to the match function itself

Embeddings are generated in controlled batches

Vectors are produced by an admin-triggered batch job that selects candidate products, builds the text blob, embeds them and writes the vector plus a timestamp back to the product row. Because the timestamp is stored, re-embedding can target only products whose content changed since they were last embedded.

That is deliberate rather than lazy: a database trigger firing an AI call on every product edit is how clients end up with unpredictable AI bills during a bulk import of 40,000 rows.