The Vector Pulse
Search that understands 'warm winter wear' means the heated jacket.
The incident
Search analytics showed the same brutal pattern weekly: users typed concepts ('warm winter wear'), the catalog had the product ('Heated Jacket Pro'), and WHERE name LIKE '%warm%' connected them zero times. The product existed. The revenue didn't.
Why the obvious approach fails
Keyword matching sees strings, not meaning — synonyms, phrasing, and intent are invisible to it. But the fashionable overcorrection fails too: pure vector search is fuzzy where users are precise. Someone typing an exact SKU or a product name wants that item, not its five nearest semantic neighbors. Production search needs both signals fused — lexical precision when the query is exact, semantic recall when it's conceptual — plus a pipeline that keeps embeddings fresh as the catalog changes, because a stale index is a silent liar.
The standard
- Search is hybrid by default: BM25 keyword scoring fused with vector kNN similarity, weights tuned per corpus.
- Embedding happens on write — create/update events auto-embed through a pipeline; the index is never a batch-job afterthought.
- Exact-match signals (SKUs, codes, titles) short-circuit to lexical results; semantics augment, never override, precision.
- Relevance is measured with a maintained golden query set; ranking changes ship with eval evidence.
- Index freshness is monitored: embedding lag between write and searchability is a tracked SLO.
How Cortex validates it
At plan time
Search blueprints declare the hybrid strategy, embedding pipeline, and freshness SLO.
At validation
assert: search.mode == hybrid(bm25, knn)assert: write_event → embedded within SLOassert: golden_queries.recall >= release_barAssertion syntax is illustrative.
At review
The approver sees the golden-set results: conceptual queries finding exact products, exact queries staying exact.