What vector search forbids

You cannot have exact nearest neighbours at scale, arbitrary filters, fresh updates, and a stable index simultaneously.

Exactness costs a full bandwidth-bound scan.
Filters in the middle selectivity band defeat both available strategies.
Freshness is expensive because the structures rebuild rather than merge.
None of it survives an embedding-model change, which invalidates every stored vector at once.

The escapes and their prices

Escape	Buys	Price
Approximation (the foundational one)	speed	recall, which is measurable, a rare gift: measure the recall/QPS curve rather than trust defaults
Quantization	bandwidth and capacity	distance error, recovered by exact reranking of survivors
Partitioning by filter value	filtered search	combinatorial index explosion and loss of generality
Hybrid retrieval	coverage of each retriever's blind spot	double cost and a fusion method that must discard calibration it does not have
Full fan-out	shard-level parallelism	tail latency that gets worse the more you scale

Case, resolvedConstructed case · Amara at Haatly

Similar items: 3 results instead of 10

Brute-force similar items took half a second per query, a filtered query returned 3 results instead of 10, and new vectors from a better model made the strip worse. Where it started: Vector search →

Brute force (0.5 s, 2 QPS). Amara builds an IVF index with PQ codes: 10 million × 96 bytes ≈ 0.96 GB instead of 30 GB, so scans read about 30 times fewer bytes, and IVF reads only a small fraction of cells. She measures the recall-versus-QPS curve against brute-force truth on 1,000 sample queries and picks the point with recall ≈ 0.95 rather than trusting defaults. Exact distances rerank the survivors.

Three results instead of ten. The filter (Pune, in stock, under 500 rupees) selects about 0.5% of the corpus, in the hard middle band. Post-filtering fetches too few; pre-filtering over 50,000 items by brute force is affordable here (50,000 × 3 KB = 150 MB of reads per query, about 3 ms at 50 GB/s). She routes by selectivity: if the filter leaves under about 100,000 items, brute-force them; if it leaves more, use ANN with a large over-fetch.

The new model. Diego's new embeddings and the old ones live in different coordinate systems. They agree to treat the model as a schema: build a second index, dual-write for two weeks, cut over, then delete the old one.

She also adds a hybrid path: for the query "kurta KT-4471" (a product code), embeddings alone miss the exact code, so she fuses BM25 and vector results using reciprocal rank fusion. The strip now works.