What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Vector product quantization (PQ) compresses vectors by replacing groups of coordinates with learned codes, so a search system can rank stored vectors without repeatedly reading every original value. It reduces memory use, but makes distance estimates approximate. In IVF-PQ, a second trade-off is added: the system searches only selected clusters, so it can miss a true neighbor outside them. There is no universal accuracy penalty; the result depends on the data, search settings, metric, and whether the system reranks candidates against original vectors.
How does vector quantization work?
Product quantization divides each vector into m subvectors and learns a codebook for each subspace. Instead of storing all the coordinates, the index stores the codebook entry selected for each subvector. At query time, it estimates distances by looking up and combining query-to-codebook distances. Faiss describes training PQ with k-means and building distance tables over the subquantizer centroids (Faiss ProductQuantizer documentation).
As an Amazon Associate I earn from qualifying purchases.
The number of subvectors and bits used for each code determine the compact code size. In OpenSearch’s description, a code with m subvectors and code_size bits per subvector uses m × code_size bits per vector; with 8-bit codes, that is m bytes before identifiers and other index overhead (OpenSearch vector storage optimization). Codebooks should be trained on vectors representative of the vectors the index will search.
Recommended Free Tools
What IVF adds
In an inverted-file index, or IVF-PQ, a coarse quantizer assigns vectors to clusters called lists. A query searches a selected number of nearby lists, set by n_probes, and scores the PQ codes in those lists. This reduces the number of candidates to inspect, but a neighbor in an unvisited list is unavailable to the search (NVIDIA cuVS IVF-PQ indexing guide).
#1 Best Overall
How much accuracy do you lose with vector quantization?
There is no fixed percentage. PQ approximates vectors and their distances, while IVF-PQ can also omit candidates by not visiting every list. The combined effect varies with the dataset, query workload, metric, code configuration, and search breadth. The implementation documentation describes the trade-off but does not establish a universal recall-loss figure.
Representation error
A PQ code represents a subvector with a learned codebook entry rather than its original coordinates, so its estimated distance can differ from the original-vector distance. Codebook training, the number of subvectors, and bits per subvector all affect the representation. Faiss notes that PQ’s quantization objective minimizes L2 centroid error and that its error is biased toward L2, even though its implementation supports both L2 and inner-product search (Faiss FAQ).
Rank #2
- Used Book in Good Condition
Candidate omission
IVF search sees only the lists it probes. Increasing n_probes usually exposes more candidates and can improve recall, at the cost of more search work. Filtering can also leave eligible vectors inaccessible if they belong to unprobed lists; NVIDIA documents IVF-PQ filtering as applying within the selected lists (NVIDIA cuVS IVF-PQ indexing guide).
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →What reranking can and cannot recover
If original vectors are retained or accessible, a system can retrieve more approximate candidates than it ultimately needs, recompute their distances against the originals, and return the best-ranked results. This can correct ordering among retrieved candidates; it cannot recover a true neighbor that was never retrieved. Reranking also adds computation and depends on access to the original vectors (NVIDIA cuVS IVF-PQ indexing guide).
How much memory does product quantization save?
The payload comparison is straightforward, but it is not a complete index-memory comparison. A float32 vector of dimension d takes 4 × d bytes; an 8-bit PQ code with m subvectors takes m bytes before identifiers, codebooks, and index structures. Faiss lists flat PQ storage at M bytes per vector for nbits=8, and IVF-PQ at M+4 or M+8 bytes per vector depending on ID representation, before broader index structures (Faiss index summary).
OpenSearch provides formula-based estimates for two example configurations. They illustrate one implementation’s estimate, not measured universal memory use or a general comparison between index families.
| OpenSearch example | Estimated memory | Configuration |
|---|---|---|
| HNSW-PQ | Approximately 0.215 GB | 1 million 256-dimensional vectors; hnsw_m=16; pq_m=32; 8-bit codes; 100 segments |
| IVF-PQ | Approximately 0.171 GB | 1 million 256-dimensional vectors; ivf_nlist=512; pq_m=32; 8-bit codes; 100 segments |
Both figures are estimates from OpenSearch documentation accessed in 2026, under the settings shown; actual memory depends on the implementation and complete index configuration (OpenSearch vector storage optimization).
How should you tune IVF-PQ for recall?
Treat code size and search breadth as separate controls. OpenSearch recommends starting with eight bits per subquantizer and tuning m to meet a memory-recall target. In IVF-PQ, n_probes controls how many coarse lists are searched and is a key recall-latency setting (OpenSearch vector storage optimization; NVIDIA cuVS IVF-PQ indexing guide).
- Set a baseline. Measure an exact-search or higher-precision configuration on representative production vectors and queries.
- Choose the quality metric. Hold the query set, ground truth, result count
k, distance metric, and filtering conditions constant. Report recall at the targetk, or define the benchmark’s metric explicitly. - Vary code size and search breadth. Test PQ settings such as
mand, for IVF-PQ,n_probes. Measure recall, memory, and latency together instead of selecting a setting from code bytes alone. - Include the complete cost. Count IDs, codebooks, IVF or graph structures, and retained original vectors. For reranking, include candidate count, original-vector access, and extra computation or I/O.
- Measure the real workload. Keep hardware, concurrency, batch size, and cache conditions consistent; report latency percentiles and throughput where relevant. Evaluate filtered queries separately if the product uses filters.
Training, index construction, and retraining as vector distributions change also consume resources. The reviewed implementation documentation does not establish a portable latency or throughput improvement, so report those results only for the tested hardware, dataset, and settings.
When is PQ a good fit?
PQ is worth evaluating when storing full-precision vectors is a meaningful memory burden and approximate ranking is acceptable. IVF-PQ is a candidate when limiting search to selected lists is also useful and the recall impact can be measured on the target workload. If missed neighbors are costly, compare against a higher-precision baseline and consider reranking—but keep in mind that reranking can reorder retrieved candidates, not restore omitted ones.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
