Recommended Free Tools
A filtered vector query can request 10 results and return only two without an error. Amazon S3 Vectors’ new ENHANCED mode is designed to address that kind of underfill by applying the metadata filter before similarity search. In a small synthetic benchmark, it returned complete, exact results for filters matching five or fewer vectors—but it did not consistently return the exact nearest neighbors when filters matched 50 or 135 vectors.
That distinction matters: fewer results than requested is an underfill problem; returning the wrong eligible neighbors is a ranking problem. Ajay Kumar’s October 2026 DEV Community test measured both against locally computed exact neighbors. It did not measure real CLASSIC mode, so it cannot establish a direct mode-versus-mode win or verify AWS’s separate “up to 5x” claim.
As an Amazon Associate I earn from qualifying purchases.
What pre-filtering changes in a filtered vector search
A vector search ranks items by similarity to a query vector. A metadata filter narrows the eligible items—for example, to one document type, tenant, or path prefix. With a selective filter, only a small fraction of the index may qualify.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →AWS describes two S3 Vectors modes. CLASSIC evaluates filter conditions while searching candidates; when few vectors match, it can return fewer than the requested top K. ENHANCED evaluates the metadata filter first, then searches only vectors that match. AWS calls this pre-filtering. The distinction is about which vectors are considered during search, not a promise that every result will match a locally calculated exact ranking.
#1 Best Overall
In Kumar’s benchmark, the exact reference was calculated locally: first retain vectors matching the filter, then compute exact cosine distances for all of them and take the nearest K. That is the “exact ground truth” against which the service’s approximate search results were compared.
How the benchmark measured accuracy and underfill
Kumar reports loading 50,000 synthetic vectors with 384 dimensions and cosine distance into one disposable ENHANCED index in us-east-1 on October 5, 2026. The test used 50 query vectors generated from the same 64 clusters as the indexed data, but from a separate random stream. Metadata labels were independent of vector position. Filters ranged from broad—matching 25,000 vectors—to empty, with selective cases matching 135, 50, 5, 3, or 1 vector.
The author reports 6,000 filtered ENHANCED logical requests, 600 unfiltered ANN references, and 300 client-side post-filter requests. For that baseline, the client requested unfiltered results with budgets of 100, 1,000, or 10,000, then applied the filter locally. This is not a measurement of S3 Vectors CLASSIC mode.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Recall@K is not the same as completeness
Recall@K measures how many returned items overlap with the exact top-K eligible neighbors, divided in this study by the smaller of K or the total number of matching vectors. A low score means the returned ranking missed exact neighbors. Completeness measures how many of the available result slots were actually filled. A query can have high recall among the items it returned yet still underfill; it can also return K items but rank them poorly.
The benchmark’s figures therefore need both measures. When a filter matches fewer vectors than K, the maximum possible result count is the number of matches—not K. When it matches more than K, a full response can still contain the wrong nearest neighbors.
What ENHANCED returned for selective filters
For filters matching five or fewer vectors, Kumar reports ENHANCED Recall@K of 1.000 at every tested K, with complete results. In this dataset, it found all eligible items and matched the locally computed exact neighbors.
The client-side baseline could miss eligible vectors even after requesting 10,000 unfiltered results. At K=10, it found the one FONE match for 42% of query vectors and found an average of 24% of the five F001 matches. These figures describe this benchmark and its query-budget approach, not a guarantee about other indexes or workloads.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Where the exact-neighbor result weakened
ENHANCED was not exact across all selective cases. In Kumar’s single-index results, recall declined as K increased for filters matching 50 and 135 vectors:
| Filter | Matching vectors | Recall@5 | Recall@10 | Recall@20 | Recall@50 |
|---|---|---|---|---|---|
| F01 | 50 | 0.948 | 0.898 | 0.811 | 0.610 |
| FAND2 | 135 | 0.952 | 0.932 | 0.883 | 0.724 |
At F01/K=50, all 50 vectors matched the filter, but the service returned an average of 30.52 of those 50 available items. The author reports identical result sets across three repeats and says the cause was unknown. This case shows why an exact-ground-truth comparison should track completeness separately from ranking recall.
Broad filters looked closer to the unfiltered ANN reference: at K=10, F50 scored 0.984 and F10 scored 0.992, versus 0.994 for the unfiltered reference. With a sufficiently large query budget, the client-side post-filter baseline often matched ENHANCED on broad filters, though it used paginated requests.
What this says—and does not say—about AWS’s claim
AWS announced ENHANCED on September 30, 2026, saying it can return “up to 5x more of the matching vectors” for highly selective filters. That is AWS’s launch claim, not a result established by Kumar’s benchmark. The test had no real CLASSIC measurement: the comparison was against a client that queried broadly and filtered locally. It therefore cannot confirm a fivefold improvement, whether in recall, completeness, latency, or another metric.
The experiment is also bounded to one synthetic index, one AWS region, 50 query vectors, and metadata labels independent of vector position. Kumar notes that real embeddings, correlated metadata, or larger indexes may behave differently. Reported latency includes client and network time, so it is not an AWS server-latency measurement. The results are useful as a demonstration that pre-filtering can help with particular selective cases, not as a transferable performance prediction.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How to evaluate ENHANCED for your workload
Use production-representative vectors, metadata, and queries to determine whether ENHANCED improves the result your application actually needs. AWS likewise recommends representative workload testing: query work can vary with index size, the fraction of vectors matched, and filter complexity, while average recall also depends on the embedding model, dataset dimensions and size, and query distribution.
- Establish an exact reference. For a manageable evaluation set, filter the canonical data first, calculate exact distances among the eligible vectors, and retain the top K. Compare service responses with this reference rather than treating another approximate query as ground truth.
- Record completeness separately from recall. For every filter and K, record how many eligible results were available, how many the query returned, and how many overlap with the exact top K. This distinguishes underfill from an inaccurate ranking.
- Stratify by selectivity and filter complexity. Include broad and highly selective filters, different match counts, and the number and type of constraints used in real requests. Do not assume that a filter’s selectivity alone predicts its behavior.
- Measure latency with a defined boundary. State whether the measurement includes client processing and network time or reflects a server-side metric. The benchmark’s latency included the client and network.
- Check filter-size limits. AWS documents up to 100 filter constraints per ENHANCED query; each evaluated value counts toward the limit. A filter with more than 100 values may need consolidation or multiple queries with merged results.
- Validate the migration path for the bucket. AWS says new indexes in buckets created on or after September 30, 2026, use ENHANCED by default. Older buckets retain CLASSIC unless the bucket default or index mode is changed. Existing indexes can be changed in place with
UpdateIndexMode; AWS says existing vectors do not need re-ingestion and query calls do not need to change.
Filter operators, migration, and cost context
ENHANCED supports metadata operators including equality and inequality, numeric comparisons, set membership, existence checks, and $and/$or. AWS also added $startsWith for string prefixes such as document paths, URLs, or hierarchical keys. The 100-constraint ceiling is important when a query expands a list of values: each evaluated value counts, not just the top-level expression.
AWS says pre-filtering has no additional charge. Standard S3 Vectors charges for storage, PUT requests, and queries still apply, so “no additional charge” does not mean that workload size or query volume is free.
Free tools Windows power users keep installed
One-click scans. No signup required.
Quick Recap
Sources
- Ajay Kumar, “I Tested Amazon S3 Vectors’ New Pre-Filtering Against Exact Ground Truth,” DEV Community, October 6, 2026 — benchmark setup, methods, and reported results.
- Daniel Abib, “Amazon S3 Vectors now supports metadata pre-filtering for higher recall on filtered searches,” AWS News Blog, September 30, 2026 — AWS’s launch description and claim.
- AWS, “Metadata filtering – Amazon Simple Storage Service” — filter semantics, modes, constraints, and performance factors.
- AWS, “Querying vectors – Amazon Simple Storage Service” — query behavior, recall factors, and workload-testing guidance.
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.
