20:"$Sreact.view_transition" 21:I[9754,["/_next/static/chunks/2t36f5p4pjzlb.js","/_next/static/chunks/1gyqeidjm5eoa.js","/_next/static/chunks/0_fjnxc4tt_l_.js","/_next/static/chunks/1ymt1shyhmu-4.js","/_next/static/chunks/3kv2zf0i7s2wc.js","/_next/static/chunks/1vs_3g_g75zmk.js","/_next/static/chunks/2g4dh802s8oz3.js","/_next/static/chunks/1bc3f0shukf1j.js","/_next/static/chunks/192kgs9v_t14j.js","/_next/static/chunks/0z6pw0v5c_lx5.js","/_next/static/chunks/08r60e6icmtkc.js","/_next/static/chunks/3trx6c0i7_x-3.js","/_next/static/chunks/40sro5bzp9bi_.js","/_next/static/chunks/1wgf3tk38y6cf.js"],"ProjectDetail"] 22:T3ef9, vibe · 现代嵌入向量的可扩展基准测试

Extensible benchmark for embeddings

现代嵌入向量的可扩展基准测试

评估向量索引性能的基准测试框架。每一个查询落入密集的候选池，只有真正落入真值邻域半径的向量才被召回——紧贴邻域的检索，才是这个基准的全部意义。

召回率 · recall@k有多少真值重新回来了

吞吐 · qps回到真值域半径的速度有多快

邻域 · k被检集合与真值邻域的贴合程度

机制：一个查询落入候选池

VIBE 用于评估近似最近邻搜索方法（向量索引）在现代嵌入数据集上的性能。一个查询向量落入密集的候选池，真值的 k 个最近邻定义了不可见的半径；每一种索引都在尝试只把这个半径内的点拉回来。

被检集合越贴合真实邻域，召回就越高；完成一次召回的速度，就是吞吐。紧贴邻域、快速响应，二者构成每一次测量都围绕的核心。

测量与权衡

质量与吞吐始终在同一把尺子上彼此换算：召回回来越多，单位时间能回答的查询就越少。以下数值以表格对齐的等宽数字读出，正如一张结果表。

度量	表述
recall@k	召回 —— 被检集合与真值邻域的重合程度
qps	吞吐 —— 每秒回答的查询数
k	邻域 —— 真值最近邻的半径

数据覆盖

包含分布内与分布外两类设置的数据集，以及一组全面的最先进向量搜索算法。分类账式的两列，界线分明。

分布内 · in-distribution

检索发生在与训练数据同分布的嵌入之上
反映索引熟悉域上的常规表现

分布外 · out-of-distribution

检索发生在与训练数据不同的分布之上
检验索引在未见环境中的稳健性

精度与模式

支持以 8 位整数与二进制精度量化的数据集，同时支持 GPU 算法。选择的精度，决定了拉回真实邻域时所用的表示。

量化精度 · quantized

8-bit 整数与二进制精度

算力 · compute

GPU 算法支持

可扩展：把新方法接进来

作为可扩展的基准，VIBE 允许沿着两个接缝生长：新的索引方法，与新的数据集。两者都能加入这一张比对表，而不必改写整体。

新增方法

把一种新的向量索引挂进基准，让它在同样的一组数据集上被测量。

新增数据集

为基准接入新的嵌入数据，让所有已登记的方法在同一分布上同台比较。

硬件与运行环境

基准在真实硬件上跑向同一份真值：CPU 车道与 GPU 车道各自承载，最终收敛到同一个答案集。运行环境方面，已覆盖 HPC 环境，包括 Slurm 与 NUMA。

CPUIntel Xeon Gold 6230 · 在 CPU 车道上运行基准

GPUNVIDIA V100 · 在 GPU 车道上运行算法

环境LinuxApptainer / SingularitySlurmNUMAAVX-512

结果与可视化

结果可在官方网站查看，并提供多种工具与可视化，让被检集合与真值邻域之间的贴合程度一目了然。

vector-index-bench.github.io →

主语言 Python · MIT 许可 · 52 stars

23:T3eaa, vibe · Extensible benchmark for embeddings

Vector Index Benchmark for Embeddings

Extensible benchmark for embeddings

Benchmark framework for vector index performance. Each query drops into a dense candidate pool, and only the vectors inside the true-k radius should come back — how tightly the retrieved set hugs the neighborhood is the entire point.

recall@khow much of the truth comes back

qpshow fast the true-radius set is answered

khow tightly retrieved hugs the true neighborhood

Mechanism: a query drops into the pool

VIBE evaluates approximate nearest neighbor search methods — vector indexes — on modern embedding datasets. A query vector falls into a dense candidate pool; the true k nearest neighbors define an invisible radius, and each index tries to pull back only the points inside that radius.

The tighter the retrieved set hugs the true neighborhood, the higher the recall; the speed of completing that pull is the throughput. Tight retrieval and fast response frame every measurement the benchmark takes.

Measurement and the tradeoff

Quality and throughput trade against one another on the same scale: the more of the truth you bring back, the fewer queries you answer per second. The figures below read as measurement, aligned like a results table.

Measure	Reads as
recall@k	overlap between retrieved set and the true neighborhood
qps	queries answered per second
k	radius of the true-k nearest neighbors

Data coverage

Datasets cover both in-distribution and out-of-distribution settings, alongside a comprehensive collection of state-of-the-art vector search algorithms. Two ledger books, sharply divided.

In-distribution

retrieval runs over embeddings from the same distribution as the training data
captures typical performance on familiar territory

Out-of-distribution

retrieval runs over embeddings unlike the training distribution
stresses how an index behaves on unseen settings

Precision and modes

Quantized datasets are supported in 8-bit integer and binary precision, and GPU algorithms are supported. The precision you choose decides the representation used to pull back the true neighborhood.

Precision · quantized

8-bit integer and binary

Compute · compute

GPU algorithm support

Extensible: plug a new method in

As an extensible benchmark, VIBE grows along two seams: new index methods, and new datasets. Either can join the comparison table without rewriting the whole.

Add a method

Attach a new vector index to the benchmark so it is measured across the same set of datasets.

Add a dataset

Bring new embedding data into the benchmark so every registered method competes on the same distribution.

Hardware and the run

Benchmarks race across real hardware toward the same ground truth: CPU lanes and GPU lanes each carry work, and every method converges on the same answer set. For the run, HPC environments are supported, including Slurm and NUMA.

CPUIntel Xeon Gold 6230 · runs the benchmark on the CPU lane

GPUNVIDIA V100 · runs algorithms on the GPU lane

EnvironmentLinuxApptainer / SingularitySlurmNUMAAVX-512

Results and the visual site

Results can be viewed on the official website, which also features several tools and visualizations that make the overlap between retrieved set and true neighborhood legible at a glance.

vector-index-bench.github.io →

Written in Python · MIT license · 52 stars

24:T42c,retrieval-pool most directly expresses VIBE's actual mechanism: ANN search is the act of recovering the k true nearest neighbors from a dense vector pool, so a query-token ring enclosing the true-k radius, with a second retrieved arc showing the quality gap, is the one verified relationship that needs no invented numbers and no screenshot. Its ring geometry converts three evidence-backed facts (retrieval around a true neighborhood, a recall@k-vs-throughput tradeoff, extensibility across datasets and methods) into one coherent, measurable figure, mirrored by table-aligned monospaced figures. hardware-lanes and tradeoff-curve capture the throughput and measurement dimensions well, but retrieval-pool owns the core 'nearest-neighbor recovery' idea, is the least template-like, and degrades cleanly to a genuine planar cut at 390px rather than column stacking. It also directly avoids every forbidden default (no device shell, no metric-card triad, no decorative glow/grid/particles) since its shared visual is a precise vector ring diagram, not interface chrome.a:[["$","$20",null,{"enter":{"secondary-forward":"secondary-forward","secondary-back":"secondary-back-enter","default":"none"},"exit":{"secondary-forward":"secondary-forward","secondary-back":"secondary-back-exit","default":"none"},"update":{"secondary-forward":"secondary-forward","secondary-back":"secondary-back-update","default":"none"},"default":"none","children":["$","$L21","cmt2zh3l601us6271b0ynfsxv:zh-CN:beautified",{"project":{"id":"cmt2zh3l601us6271b0ynfsxv","slug":"vector-index-bench-vibe","name":"vibe","description":"评估向量索引性能的基准测试框架","authorName":"vector-index-bench","authorVerification":{"verified":false,"verifiedAt":null},"category":"development_tools","subcategory":"developer_resources","experienceUrl":"https://vector-index-bench.github.io/","qualityScore":91,"featured":false,"officialSite":null,"views":{"original":{"available":true},"official":{"available":false,"reason":"disabled_by_policy"},"beautified":{"available":true,"reason":null},"defaultView":"beautified"},"beautifiedView":{"revisionId":"cmt2zh3lk01v462715kmlgzxo","kind":"artifact","artifact":{"schemaVersion":"localized-static-html-v4","kind":"static_html","locales":["zh-CN","en"],"defaultLocale":"zh-CN","documents":{"zh-CN":"$22","en":"$23"},"designDirections":{"schemaVersion":"dodot-design-directions-v1","directions":[{"id":"retrieval-pool","name":"Retrieval Pool (Neighborhood Radius)","projectMetaphor":"A single query drops into a dense pool of candidate vectors; the ground-truth k nearest neighbors define an invisible radius, and each vector index tries to pull only the points inside that radius back out. How tightly the retrieved set hugs the true neighborhood is the entire point — recall is how much of the truth comes back, throughput is how fast.","signatureVisual":"A derived ring diagram, not a screenshot: a small query token at the center, a circle marking the true-k neighborhood boundary, candidate points just inside and outside the boundary, and a second, partial arc tracing the surviving retrieved set. The gap between the true arc and the retrieved arc visually encodes the quality loss.","heroGeometry":"Concentric rings anchored at a single query origin — the inner true-k radius, the outer candidate pool, and a broken partial arc for the returned set. Depth is implied only by ring separation, never by glow or gradient.","sectionSequence":["mechanism","measurementTradeoff","datasets","precisionModes","extensibility","hardwareAndRun","resultsSite"],"surfaceModel":"Paper-and-ink reference table under a single light source: one page per ring layer, no floating cards. Dividers are hairline rules; space behaves as physical page gutter. Precise vector lines only, matte treatment, no translucent panels, no shadows lifted off the page.","palette":"Ink-navy base, bone/paper background, a restrained amber for the true-k ring, muted slate for candidate points, dark red only for the surviving retrieved arc. Any color is semantic and reused only for its one meaning.","typography":"Monospaced numerals for every empirical value (recall, qps, k, recall@k) so figures read as measurement, paired with a humanist sans for prose. Strong tabular alignment that mirrors a results table.","desktopComposition":"Left-weighted: the hero ring sits in the upper-left, the value proposition to its right, and the empirical claim (recall@k vs qps) aligned as a table row beneath. Every section reuses the single ring vocabulary on its own page-like layer.","mobileTransformation":"The concentric ring collapses to a horizontal 'slice' cut through the pool — candidate line above, truth segment below — a genuine planar cut, not column stacking. Table rows become stacked ledger lines.","mediaStrategy":"No screenshot media. The ring diagram is the only visual and it is a derived illustration of the verified retrieval relationship. No remote runtime images, no placeholder frames.","forbiddenPatterns":["no terminal, browser, editor, or dashboard shell","no three separate metric cards in a row","no glow, gradient, particle field, or background grid as decoration","no fake chart or fabricated recall/qps numbers presented as results","no photos or logos substituted for the ring diagram"],"compositionFingerprint":"singleQueryPool:concentricRings:tableAlignedFigures"},{"id":"hardware-lanes","name":"Hardware Lanes (The Race to Ground Truth)","projectMetaphor":"Every method is racing across real hardware toward the same fixed ground truth. Some run on the Xeon CPU lanes, some on V100 GPU lanes, but all must converge on the same answer set — the race makes speed and correctness visible as distance and arrival order.","signatureVisual":"A horizontal lane diagram with a fixed ground-truth 'finish checkpoint' on the right; each lane (CPU lane, GPU lane) carries method tokens travelling left to right, those missing the truth shown trailing a corrected slash. A real distance = latency, so shorter lane runs read as faster.","heroGeometry":"A horizontal sweep of parallel lanes converging into a single right-hand checkpoint node. Flow is left-to-right and never vertical-stacked; the checkpoint is a single point, not a sequence of cards.","sectionSequence":["raceOverview","hardwareLanes","quantizationModes","extensionSeams","hpcOrchestration","labConditions","siteAndTools"],"surfaceModel":"Blueprint/trace line-plane: thin construction lines on a flat field, deliberate crosshair registration ticks, no depth. Hardware lanes are the structure; nothing floats above the plane.","palette":"Cool blue-grey field, one warm lane for CPU and one cool-accent lane for GPU, black checkpoint node, all color bound to the hardware-meaning it names.","typography":"Monospaced labels for methods and hardware (as run identifiers), columnar alignment tracking the lane x-positions; tight sparing all-caps ticks for lane headers.","desktopComposition":"Full-width race track dominates: the converging lanes are the horizontal spine of the page and each lane carries its empirical label. Description content sits as compact margins above and below the spine.","mobileTransformation":"The racing lanes rotate 90 degrees into a vertical descent — the finish checkpoint becomes the bottom — so 'down the track' is the mobile gesture instead of stacking columns.","mediaStrategy":"No screenshots. The lane diagram is a derived illustration of the verified search-and-hardware relationship. Remote runtime images and placeholder shells are forbidden.","forbiddenPatterns":["no browser, phone, dashboard, terminal, or editor shell","no generic feature-card wall","no three stat cards in a row","no artificial glow, grid background, or particle field","no invented benchmark results or timing numbers"],"compositionFingerprint":"hardwareRace:convergingLanes:singleCheckpoint"},{"id":"tradeoff-curve","name":"Tradeoff Curve (Quality vs Throughput Spectrum)","projectMetaphor":"Nearest-neighbor search is one continuous spectrum: you can trade how much of the truth you recover against how many queries-per-second you can answer. Every index sits somewhere on the same axis; the page is a walk along that axis, not a row of features.","signatureVisual":"A single derived family of curves mapping recovery-quality against throughput, sweeping from the fast-but-lossy corner to the exact-but-slow corner, with an operating marker showing a current compromise. The curve is an abstraction of a verified tradeoff, drawn structurally, not quoting results.","heroGeometry":"One dominant corner-to-corner diagonal sweep across the full canvas, cut by a single orthogonal operating line; the whole composition is this one trace and its axis labels.","sectionSequence":["openQuestion","theTradeoff","precisionKnobs","dataCoverage","growingTheBench","findResults"],"surfaceModel":"Analytical register: aligned axis grid used only to position the tradeoff trace, ruled horizontally, absolutely no decorative grid elsewhere. Sections hang off the single axis as dated margin entries.","palette":"Very low-chroma ink palette: a warm trace on near-white, a cool orthogonal operating line, axis ticks in grey — three notes maximum, all bound to the graph's meaning.","typography":"A single durable serif for narrative voice plus monospaced for every number, so prose and measurement are clearly different instruments on the same page.","desktopComposition":"The corner-to-corner trace is the fixed horizontal rail the whole page sits under; content sits above or below the rail as marginalia, never as equal cards. One continuous reading line across viewport height.","mobileTransformation":"The full-canvas rail reorganized as a single tall vertical trace that the user scrolls along — scrolling is the act of moving along the tradeoff — instead of collapsing columns.","mediaStrategy":"No product screenshots; the abstract tradeoff trace is the sole visual and it is a structural illustration of a verified relationship. No runtime-remote media, no placeholder frame.","forbiddenPatterns":["no browser, terminal, app, or dashboard shell","no metric-card triad or feature-wall","no decorative grid, gradient, particles, or glow","no screenshot or fabricated result plot passed as real data","no separate hero section with slogan-only then unrelated sections"],"compositionFingerprint":"singleTradeoff:cornerDiagonal:axisRegister"}],"selectedDirectionId":"retrieval-pool","selectionRationale":"$24"},"designBrief":{"schemaVersion":"dodot-design-brief-v1","selectedDirectionId":"retrieval-pool","projectMetaphor":"A single query drops into a dense pool of candidate vectors; the ground-truth k nearest neighbors define an invisible radius, and each vector index tries to pull only the points inside that radius back out. How tightly the retrieved set hugs the true neighborhood is the entire point — recall is how much of the truth comes back, throughput is how fast.","signatureVisual":"A derived ring diagram, not a screenshot: a small query token at the center, a circle marking the true-k neighborhood boundary, candidate points just inside and outside the boundary, and a second, partial arc tracing the surviving retrieved set. The gap between the true arc and the retrieved arc visually encodes the quality loss.","compositionFingerprint":"singleQueryPool:concentricRings:tableAlignedFigures","heroGeometry":"Concentric rings anchored at a single query origin — the inner true-k radius, the outer candidate pool, and a broken partial arc for the returned set. Depth is implied only by ring separation, never by glow or gradient.","sectionSequence":["mechanism","measurementTradeoff","datasets","precisionModes","extensibility","hardwareAndRun","resultsSite"],"surfaceModel":"Paper-and-ink reference table under a single light source: one page per ring layer, no floating cards. Dividers are hairline rules; space behaves as physical page gutter. Precise vector lines only, matte treatment, no translucent panels, no shadows lifted off the page.","palette":"Ink-navy base, bone/paper background, a restrained amber for the true-k ring, muted slate for candidate points, dark red only for the surviving retrieved arc. Any color is semantic and reused only for its one meaning.","typography":"Monospaced numerals for every empirical value (recall, qps, k, recall@k) so figures read as measurement, paired with a humanist sans for prose. Strong tabular alignment that mirrors a results table.","desktopComposition":"Left-weighted: the hero ring sits in the upper-left, the value proposition to its right, and the empirical claim (recall@k vs qps) aligned as a table row beneath. Every section reuses the single ring vocabulary on its own page-like layer.","mobileTransformation":"The concentric ring collapses to a horizontal 'slice' cut through the pool — candidate line above, truth segment below — a genuine planar cut, not column stacking. Table rows become stacked ledger lines.","mediaStrategy":"No screenshot media. The ring diagram is the only visual and it is a derived illustration of the verified retrieval relationship. No remote runtime images, no placeholder frames.","evidencePlan":["VIBE is an extensible benchmark for approximate nearest neighbor search methods (vector indexes) using modern embedding datasets (material.json.summary.evidence[0]; material.json.projection.description).","Includes in-distribution and out-of-distribution datasets, and a comprehensive collection of state-of-the-art vector search algorithms (material.json.summary.evidence[1]; readme headings 'In-distribution datasets' and 'Out-of-distribution datasets').","Supports quantized datasets in 8-bit integer and binary precision, and GPU algorithms (material.json.summary.evidence[2]).","Supports HPC environments including Slurm and NUMA, and benchmarks run on Intel Xeon Gold 6230 CPUs while GPU algorithms use NVIDIA V100 (material.json.summary.evidence[3] and material.json.summary.evidence[5]).","Results can be viewed on the official website at https://vector-index-bench.github.io/, which also features several tools and visualizations (material.json.summary.evidence[4]).","Primary language is Python and the license is MIT (material.json.extraction.facts 'primary_language' and 'license')."],"localeContract":{"zh-CN":"中文与英文共享同一方案与章节顺序。中文文案保持自然密度：使用 slogan '现代嵌入向量的可扩展基准测试'，oneLine '评估向量索引性能的基准测试框架'，overview 段落与 highlights/bestFor 列表仅呈现 material.json.localizations['zh-CN'] 已提供的事实，不添加任何扩展数据或指标。技术术语（recall@k、qps、ANN、Slurm、NUMA、Avx）保留英文。","en":"Chinese and English share the same direction and section order. English copy keeps natural density: use slogan 'Extensible benchmark for embeddings', oneLine 'Benchmark framework for vector index performance', and present the overview, highlights and bestFor lists only from facts provided in material.json.localizations.en — no added metrics, quotes, or invented results."},"forbiddenPatterns":["no terminal, browser, editor, dashboard, or phone shell with no supporting evidence","no three separate metric cards in a row as a default landing grouping","no repeated feature-card wall that any project's copy could fill","no decorative glow, gradient, particle field, or background grid standing in for a real visual idea","no fake product screenshot, fake chart, or fabricated recall/qps/quantization results","no same radius/shadow/blur/spacing scale used across every section","no desktop columns merely stacked on mobile; mobile must recompose via a planar cut"],"historicalReuse":null,"mediaStrategyNote":"assetManifest is empty and antiTemplateContext is empty; no local asset or historical seed exists, so no historicalReuse block is emitted."},"visualReview":{"schemaVersion":"dodot-automated-design-review-v1","status":"approved","summary":"The pages faithfully implement the retrieval-pool direction. The derived ring diagram (query token, amber true-k ring, slate candidate points, dashed dark-red retrieved arc) is present in both locales and is the sole, non-screenshot visual; every prose claim, hardware figure, and capability traces to material.json; the paper-and-ink ledger surface model, ink-navy/bone/amber/slate/dark-red palette, monospaced tabular-nums figures and monospaced-plus-sans pairing all follow the brief; and the seven sections match the retrieval-pool sequence exactly. The mobile ≤760px CSS recomposes the ring into a genuine horizontal planar slice (candidate line above, truth segment below) rather than column stacking, and the two locales share equivalent, natural copy with matching section order. A single non-blocking locale-naturalness note is recorded.","checks":{"projectFit":{"passed":true,"evidence":"Section sequence mechanism->measurementTradeoff->datasets->precisionModes->extensibility->hardwareAndRun->resultsSite matches retrieval-pool brief exactly. Paper-and-ink ledger surface via hairline rules and no floating cards; ink-navy/--bone paper/amber true-k/slate candidates/dark-red retrieved palette; monospaced tabular-nums numerals for recall@k, qps, k paired with a humanist sans; left-weighted hero with ring upper-left, value prop to its right, and an aligned ledger (recall/qps/k) beneath. "},"evidenceBound":{"passed":true,"evidence":"zh slogan ''现代嵌入向量的可扩展基准测试'' and oneLine ''评估向量索引性能的基准测试框架'' match material.json.localizations.zh-CN; en h1 ''Extensible benchmark for embeddings'' and oneLine ''Benchmark framework for vector index performance'' match localizations.en. In/out-of-distribution datasets plus SOTA algorithms (summary.evidence[1]); 8-bit integer/binary quantization and GPU (evidence[2]); HPC with Slurm and NUMA, Intel Xeon Gold 6230 CPU, NVIDIA V100 GPU (evidence[3],[5]); results URL vector-index-bench.github.io (evidence[4]); Python/MIT and 52 stars (extraction facts / repository.stars). No invented recall/qps/quantization numbers appear. "},"signatureVisual":{"passed":true,"evidence":"A derived inline SVG ring diagram (not a screenshot) exists in both locales and locales: central query token, inner amber true-k ring (r=104), outer candidate-pool ring, slate candidate points inside and outside the boundary, and a dashed dark-red partial arc for the retrieved set whose gap from the true arc encodes quality loss, exactly as the signatureVisual/heeroGeometry specify. No image or screenshot media. "},"mobileTransformation":{"passed":true,"evidence":"At max-width:760px the CSS hides .ring and displays a dedicated .ring-slice SVG, a horizontal planar cut through the pool with a line of candidate points above and a paired truth/retrieved segment below, a genuine recomposition rather than column stacking. Surrounding ledgers (.row) and tables reflow as stacked/scrollable ledger lines. "},"localeParity":{"passed":true,"evidence":"zh-CN and en share identical section order, structure, and section bodies, with each h1 using its own brief-provided slogan. Copy is equivalent and natural; only technical terms (recall@k, qps, ANN, Slurm, NUMA, AVX-512) remain in English on the zh page, per the localeContract. A minor kicker redundancy is noted as a warning. "},"antiTemplate":{"passed":true,"evidence":"Composition is the project-specific ring diagram plus page-print ledger rows; no generic hero/card-wall/device-shell; no decorative gradient, glow, particle field, or background grid standing in for a real visual idea; the ring is the only visual and it is a derived illustration, not interface chrome. "},"noFakeUi":{"passed":true,"evidence":"QA reports fakeControlCount=0, emptyLinkCount=0, brokenImageCount=0 for both locales at desktop and mobile; precision modes and hardware items are text-ledger rows and informational pills, not interactive widgets; the results anchor is a real href to vector-index-bench.github.io; no fabricated benchmark figures. "}},"issues":[{"code":"zh-hero-kicker-redundancies","severity":"warning","locale":"zh-CN","description":"In zh-CN the hero kicker shows the English slogan ''Extensible benchmark for embeddings'' while the en page shows the project brand name ''Vector Index Benchmark for Embeddings'' in the same slot; using an English slogan as a decorative kicker on the zh page is slightly redundant because the zh h1 already states the slogan in Chinese. Content is still factual and neither locale is broken. ","repairInstruction":"Optional: align the zh kicker with the brand-name style used on the en page (e.g. ''Vector Index Benchmark for Embeddings''), or drop the kicker, to improve cross-locale symmetry and zh naturalness. "}]},"repairEvidence":[{"attempt":1,"before":{"qaPassed":true,"designReviewApproved":false,"qaEvidence":{"zh-CN":[{"viewport":"desktop","passed":true,"localePassed":true,"metrics":{"language":"zh-CN","h1Count":1,"h2Count":7,"sectionCount":7,"textLength":1418,"cjkCount":828,"latinLetterCount":304,"scrollWidth":1280,"clientWidth":1280,"scrollHeight":3935,"h1Top":133.65625,"bottomInset":282.140625,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"ae35967e95e69fb243f141af0743bcc65ca730001e2ae7cbf9f011b3258bbffb","middle":"c76987a5cbee7c1e845c717d10645b8481341ab4cdab2268db8baf17b6161fa0","bottom":"ad79255958f725aafbf9a814a356952981f26979278d497e967528af2651cd7e"},"consoleErrors":[],"networkRequests":[]},{"viewport":"mobile","passed":true,"localePassed":true,"metrics":{"language":"zh-CN","h1Count":1,"h2Count":7,"sectionCount":7,"textLength":1418,"cjkCount":828,"latinLetterCount":304,"scrollWidth":390,"clientWidth":390,"scrollHeight":5127,"h1Top":499.65625,"bottomInset":280.5625,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"7343d474eff3a77750e4acd393f65631a0558adfaeb9098a13972c9fa66285ef","middle":"31ccf76d10741d56a895b910bddb4efebf9b101e9c1f72f67fa1dcce856e5d86","bottom":"7f95c2b1ca210997b5fb6054d504813d19aea276dc5392ed44395f895c931e60"},"consoleErrors":[],"networkRequests":[]}],"en":[{"viewport":"desktop","passed":true,"localePassed":true,"metrics":{"language":"en","h1Count":1,"h2Count":7,"sectionCount":7,"textLength":3549,"cjkCount":0,"latinLetterCount":2865,"scrollWidth":1280,"clientWidth":1280,"scrollHeight":4271,"h1Top":133.65625,"bottomInset":281.65625,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"d78b111d2a1fc695079614e901260ed8ed09683a899951f4948d9cf91a20de06","middle":"f08ce0a6249d91e5c149ad4198107fdcf2c06d159d63fdae095c4eda9f0f2d19","bottom":"32ecd03fde99011660dde947334417e9cf7ff7f6cf9a6d30a9107a29d3021a4f"},"consoleErrors":[],"networkRequests":[]},{"viewport":"mobile","passed":true,"localePassed":true,"metrics":{"language":"en","h1Count":1,"h2Count":7,"sectionCount":7,"textLength":3549,"cjkCount":0,"latinLetterCount":2865,"scrollWidth":390,"clientWidth":390,"scrollHeight":5860,"h1Top":499.65625,"bottomInset":318.953125,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"38dd0d82c7bdf016a7e3c2af82d70b5db6064e5a5bfd52c9fd4b13d19b9a3028","middle":"116551ee3036ffce5390b588538557e5932abfcce2cfb6c7cdb55033dfef5a82","bottom":"72fb62e42511da541ee47c11d51d0f76ef36c168cc11a262da035feb727c555a"},"consoleErrors":[],"networkRequests":[]}]},"visualReview":{"schemaVersion":"dodot-automated-design-review-v1","status":"revise","summary":"The page faithfully implements the retrieval-pool direction: the signature concentric-ring SVG diagram is present, evidence-bound content is accurate, sections follow the brief's order, and the paper-and-ink ledger treatment avoids template defaults. However, the mobile transformation fails the brief's explicit requirement: at ≤760px the layout only stacks the two-column grids into single columns and scales the ring down, rather than recomposing the ring into the required horizontal planar slice (candidate line above, truth segment below). A secondary locale warning concerns the asymmetric en hero h1.","checks":{"projectFit":{"passed":true,"evidence":"Section sequence mechanism→measurementTradeoff→datasets→precisionModes→extensibility→hardwareAndRun→resultsSite exactly matches the retrieval-pool brief. Paper-and-ink ledger surface model via hairline rules and no floating cards; ink-navy/bone/amber/slate/dark-red palette; monospaced numerals for every empirical value (recall@k, qps, k) paired with humanist sans; left-weighted hero with the ring upper-left and an aligned recall/qps ledger row beneath."},"evidenceBound":{"passed":true,"evidence":"All copy is traceable to material.json: en slogan and zh slogan/oneLine from localizations; datasets in/out-of-distribution and SOTA algorithms (summary.evidence[1]); 8-bit integer and binary quantization plus GPU (evidence[2]); HPC/Slurm/NUMA, Intel Xeon Gold 6230 CPU, NVIDIA V100 GPU (evidence[3],[5]); results site URL vector-index-bench.github.io (evidence[4]); Python/MIT and 52 stars from extraction facts/repository."},"signatureVisual":{"passed":true,"evidence":"A derived inline SVG ring diagram exists in both locales: central query token, inner amber true-k ring, slate candidate points inside/outside the boundary, and a dashed dark-red partial arc for the retrieved set — the gap between true arc and retrieved arc encodes quality loss exactly as specified. No screenshot is used."},"mobileTransformation":{"passed":false,"evidence":"At max-width:760px the CSS only collapses the .dat-two/.prec/.flex2/.row grids to single (stacked) columns and reduces the .ring max-width; no recomposition of the concentric ring into the brief's required horizontal planar slice (candidate line above, truth segment below). This is column stacking plus scaling, explicitly the forbidden pattern."},"localeParity":{"passed":true,"evidence":"zh-CN and en share identical section order and structure; section bodies are equivalent and natural. Note: zh-CN h1 uses the provided zh slogan while the en h1 is an invented headline ('Benchmarks that race vector indexes toward the true neighborhood') with the provided en slogan demoted to the kicker — flagged as a warning below."},"antiTemplate":{"passed":true,"evidence":"Composition is the project-specific ring diagram plus page-print ledger rows; no generic hero/card wall/device shell; no decorative gradient, glow, particle field, or background grid standing in for a real visual idea."},"noFakeUi":{"passed":true,"evidence":"QA reports fakeControlCount=0 for all locales/viewports; precision 'modes' and hardware items are text-ledger rows and informational pills (not widgets); the results link is a real href to vector-index-bench.github.io; no fabricated recall/qps/quantization numbers or device-shell chrome."}},"issues":[{"code":"mobile-planar-cut-missing","severity":"blocking","locale":"all","description":"The brief's mobileTransformation requires the concentric ring to collapse into a horizontal planar slice through the pool (candidate line above, truth segment below), a genuine planar cut and not column stacking. The ≤760px CSS only stacks the two-column grids to single columns and scales the ring SVG down via max-width, leaving the ring as a circular diagram — the forbidden default and a direct deviation from the delivery contract.","repairInstruction":"Add a ≤760px media-rule that recomposes the hero ring as a horizontal slice cut: render a line of candidate points above a paired truth segment (e.g. a compact horizontal SVG along the pool), while also stacking the surrounding ledgers, instead of merely reducing the existing circular ring's max-width. Keep table rows as stacked ledger lines as already done."},{"code":"en-hero-h1-not-brief-slogan","severity":"warning","locale":"en","description":"zh-CN uses the brief-provided zh slogan as the h1, but the en page replaces the provided en slogan ('Extensible benchmark for embeddings') with an invented headline ('Benchmarks that race vector indexes toward the true neighborhood') and demotes the provided slogan to the kicker — breaking h1 symmetry between locales.","repairInstruction":"Align the en hero with the zh pattern by using the brief-provided en slogan as the h1 (or promote the brief slogan to h1 in both locales for symmetric hero headers), and reserve the invented phrase for secondary copy if desired."}]}},"after":{"qaPassed":true,"designReviewApproved":true,"qaEvidence":{"zh-CN":[{"viewport":"desktop","passed":true,"localePassed":true,"metrics":{"language":"zh-CN","h1Count":1,"h2Count":7,"sectionCount":7,"textLength":1418,"cjkCount":828,"latinLetterCount":304,"scrollWidth":1280,"clientWidth":1280,"scrollHeight":3935,"h1Top":133.65625,"bottomInset":282.140625,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"ae35967e95e69fb243f141af0743bcc65ca730001e2ae7cbf9f011b3258bbffb","middle":"c76987a5cbee7c1e845c717d10645b8481341ab4cdab2268db8baf17b6161fa0","bottom":"ad79255958f725aafbf9a814a356952981f26979278d497e967528af2651cd7e"},"consoleErrors":[],"networkRequests":[]},{"viewport":"mobile","passed":true,"localePassed":true,"metrics":{"language":"zh-CN","h1Count":1,"h2Count":7,"sectionCount":7,"textLength":1418,"cjkCount":828,"latinLetterCount":305,"scrollWidth":390,"clientWidth":390,"scrollHeight":4945,"h1Top":291.3125,"bottomInset":292.90625,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"197cae91c216604769ee6eae0a056263af7adf245dd8de7401c82840df1ce5fe","middle":"c9cfc86752f647314bf8cdac8d07bbd5a52487706ea266013264ea42e90b8d6c","bottom":"0d8f9b83d248d0bfbf8dd1feda15d22355a20bbb8819ec9e85dddc8792a66e0b"},"consoleErrors":[],"networkRequests":[]}],"en":[{"viewport":"desktop","passed":true,"localePassed":true,"metrics":{"language":"en","h1Count":1,"h2Count":7,"sectionCount":7,"textLength":3520,"cjkCount":0,"latinLetterCount":2841,"scrollWidth":1280,"clientWidth":1280,"scrollHeight":4215,"h1Top":133.65625,"bottomInset":281.96875,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"b82b8b4714705911a04d90106cb421a67ed75b0e077f2c06d44ea90759a80952","middle":"e1cbd6da2cc9da71d5e907a463e49c9d5a01998d4d907c87edf2b7290d44fb35","bottom":"dc438bab9270a0087b02512996c29cf013de2b295662149df0092ba4e01b9ecb"},"consoleErrors":[],"networkRequests":[]},{"viewport":"mobile","passed":true,"localePassed":true,"metrics":{"language":"en","h1Count":1,"h2Count":7,"sectionCount":7,"textLength":3520,"cjkCount":0,"latinLetterCount":2842,"scrollWidth":390,"clientWidth":390,"scrollHeight":5644,"h1Top":291.3125,"bottomInset":330.734375,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"bd9f8744d1059845726b2edbada3f0ac94f302f278b2ba06aba15f4ecd7a1b42","middle":"fd7ba4a3733a82cf9ef4d1373d6832c8d39b5c9310ca3059335489a65428c329","bottom":"3cf9a72edc01ac0921a1e836c4b77672f71246e327aada7bdca4c60a8068ec6e"},"consoleErrors":[],"networkRequests":[]}]},"visualReview":{"schemaVersion":"dodot-automated-design-review-v1","status":"approved","summary":"The pages faithfully implement the retrieval-pool direction. The derived ring diagram (query token, amber true-k ring, slate candidate points, dashed dark-red retrieved arc) is present in both locales and is the sole, non-screenshot visual; every prose claim, hardware figure, and capability traces to material.json; the paper-and-ink ledger surface model, ink-navy/bone/amber/slate/dark-red palette, monospaced tabular-nums figures and monospaced-plus-sans pairing all follow the brief; and the seven sections match the retrieval-pool sequence exactly. The mobile ≤760px CSS recomposes the ring into a genuine horizontal planar slice (candidate line above, truth segment below) rather than column stacking, and the two locales share equivalent, natural copy with matching section order. A single non-blocking locale-naturalness note is recorded.","checks":{"projectFit":{"passed":true,"evidence":"Section sequence mechanism->measurementTradeoff->datasets->precisionModes->extensibility->hardwareAndRun->resultsSite matches retrieval-pool brief exactly. Paper-and-ink ledger surface via hairline rules and no floating cards; ink-navy/--bone paper/amber true-k/slate candidates/dark-red retrieved palette; monospaced tabular-nums numerals for recall@k, qps, k paired with a humanist sans; left-weighted hero with ring upper-left, value prop to its right, and an aligned ledger (recall/qps/k) beneath. "},"evidenceBound":{"passed":true,"evidence":"zh slogan ''现代嵌入向量的可扩展基准测试'' and oneLine ''评估向量索引性能的基准测试框架'' match material.json.localizations.zh-CN; en h1 ''Extensible benchmark for embeddings'' and oneLine ''Benchmark framework for vector index performance'' match localizations.en. In/out-of-distribution datasets plus SOTA algorithms (summary.evidence[1]); 8-bit integer/binary quantization and GPU (evidence[2]); HPC with Slurm and NUMA, Intel Xeon Gold 6230 CPU, NVIDIA V100 GPU (evidence[3],[5]); results URL vector-index-bench.github.io (evidence[4]); Python/MIT and 52 stars (extraction facts / repository.stars). No invented recall/qps/quantization numbers appear. "},"signatureVisual":{"passed":true,"evidence":"A derived inline SVG ring diagram (not a screenshot) exists in both locales and locales: central query token, inner amber true-k ring (r=104), outer candidate-pool ring, slate candidate points inside and outside the boundary, and a dashed dark-red partial arc for the retrieved set whose gap from the true arc encodes quality loss, exactly as the signatureVisual/heeroGeometry specify. No image or screenshot media. "},"mobileTransformation":{"passed":true,"evidence":"At max-width:760px the CSS hides .ring and displays a dedicated .ring-slice SVG, a horizontal planar cut through the pool with a line of candidate points above and a paired truth/retrieved segment below, a genuine recomposition rather than column stacking. Surrounding ledgers (.row) and tables reflow as stacked/scrollable ledger lines. "},"localeParity":{"passed":true,"evidence":"zh-CN and en share identical section order, structure, and section bodies, with each h1 using its own brief-provided slogan. Copy is equivalent and natural; only technical terms (recall@k, qps, ANN, Slurm, NUMA, AVX-512) remain in English on the zh page, per the localeContract. A minor kicker redundancy is noted as a warning. "},"antiTemplate":{"passed":true,"evidence":"Composition is the project-specific ring diagram plus page-print ledger rows; no generic hero/card-wall/device-shell; no decorative gradient, glow, particle field, or background grid standing in for a real visual idea; the ring is the only visual and it is a derived illustration, not interface chrome. "},"noFakeUi":{"passed":true,"evidence":"QA reports fakeControlCount=0, emptyLinkCount=0, brokenImageCount=0 for both locales at desktop and mobile; precision modes and hardware items are text-ledger rows and informational pills, not interactive widgets; the results anchor is a real href to vector-index-bench.github.io; no fabricated benchmark figures. "}},"issues":[{"code":"zh-hero-kicker-redundancies","severity":"warning","locale":"zh-CN","description":"In zh-CN the hero kicker shows the English slogan ''Extensible benchmark for embeddings'' while the en page shows the project brand name ''Vector Index Benchmark for Embeddings'' in the same slot; using an English slogan as a decorative kicker on the zh page is slightly redundant because the zh h1 already states the slogan in Chinese. Content is still factual and neither locale is broken. ","repairInstruction":"Optional: align the zh kicker with the brand-name style used on the en page (e.g. ''Vector Index Benchmark for Embeddings''), or drop the kicker, to improve cross-locale symmetry and zh naturalness. "}]}},"changedFiles":["artifact/zh-CN.html","artifact/en.html"]}],"provenance":{"executor":"opencode","model":"deepseek/deepseek-v4-flash","opencodeVersion":"1.1.58","promptVersion":"dodot-showcase-deepseek-v4","inputHash":"fef938fd779824c128683879dd95f0eede380c4a7029b5b05022ff03a818d43f","artifactHash":"e89f4071fb4824b8ea0309be66a2b3acf5b102ed2d42f41e83fac68e42f2aff4","designSkill":"dodot-showcase-design"},"qaEvidence":{"zh-CN":[{"viewport":"desktop","passed":true,"localePassed":true,"metrics":{"language":"zh-CN","h1Count":1,"h2Count":7,"sectionCount":7,"textLength":1418,"cjkCount":828,"latinLetterCount":304,"scrollWidth":1280,"clientWidth":1280,"scrollHeight":3935,"h1Top":133.65625,"bottomInset":282.140625,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"ae35967e95e69fb243f141af0743bcc65ca730001e2ae7cbf9f011b3258bbffb","middle":"c76987a5cbee7c1e845c717d10645b8481341ab4cdab2268db8baf17b6161fa0","bottom":"ad79255958f725aafbf9a814a356952981f26979278d497e967528af2651cd7e"},"consoleErrors":[],"networkRequests":[]},{"viewport":"mobile","passed":true,"localePassed":true,"metrics":{"language":"zh-CN","h1Count":1,"h2Count":7,"sectionCount":7,"textLength":1418,"cjkCount":828,"latinLetterCount":305,"scrollWidth":390,"clientWidth":390,"scrollHeight":4945,"h1Top":291.3125,"bottomInset":292.90625,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"197cae91c216604769ee6eae0a056263af7adf245dd8de7401c82840df1ce5fe","middle":"c9cfc86752f647314bf8cdac8d07bbd5a52487706ea266013264ea42e90b8d6c","bottom":"0d8f9b83d248d0bfbf8dd1feda15d22355a20bbb8819ec9e85dddc8792a66e0b"},"consoleErrors":[],"networkRequests":[]}],"en":[{"viewport":"desktop","passed":true,"localePassed":true,"metrics":{"language":"en","h1Count":1,"h2Count":7,"sectionCount":7,"textLength":3520,"cjkCount":0,"latinLetterCount":2841,"scrollWidth":1280,"clientWidth":1280,"scrollHeight":4215,"h1Top":133.65625,"bottomInset":281.96875,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"b82b8b4714705911a04d90106cb421a67ed75b0e077f2c06d44ea90759a80952","middle":"e1cbd6da2cc9da71d5e907a463e49c9d5a01998d4d907c87edf2b7290d44fb35","bottom":"dc438bab9270a0087b02512996c29cf013de2b295662149df0092ba4e01b9ecb"},"consoleErrors":[],"networkRequests":[]},{"viewport":"mobile","passed":true,"localePassed":true,"metrics":{"language":"en","h1Count":1,"h2Count":7,"sectionCount":7,"textLength":3520,"cjkCount":0,"latinLetterCount":2842,"scrollWidth":390,"clientWidth":390,"scrollHeight":5644,"h1Top":291.3125,"bottomInset":330.734375,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"bd9f8744d1059845726b2edbada3f0ac94f302f278b2ba06aba15f4ecd7a1b42","middle":"fd7ba4a3733a82cf9ef4d1373d6832c8d39b5c9310ca3059335489a65428c329","bottom":"3cf9a72edc01ac0921a1e836c4b77672f71246e327aada7bdca4c60a8068ec6e"},"consoleErrors":[],"networkRequests":[]}]}},"generatorVersion":"opencode-1.1.58:deepseek/deepseek-v4-flash:dodot-showcase-deepseek-v4","publishedAt":"2026-08-20T09:54:46.338Z"},"rawExtraction":{"projectRevisionId":"cmt2zh3lb01uy6271l5m4li61","projectVersion":1,"sources":[{"canonicalKey":"github:vector-index-bench/vibe","role":"primary","revisionVersion":1,"url":"https://github.com/vector-index-bench/vibe"}],"facts":[{"key":"repository_owner","value":"vector-index-bench","evidence":[{"path":"repository.owner","excerpt":"vector-index-bench"}]},{"key":"repository_name","value":"vibe","evidence":[{"path":"repository.name","excerpt":"vibe"}]},{"key":"description","value":"Vector Index Benchmark for Embeddings (VIBE) is an extensible benchmark for approximate nearest neighbor search methods, or vector indexes, using modern embedding datasets.","evidence":[{"path":"repository.description","excerpt":"Vector Index Benchmark for Embeddings (VIBE) is an extensible benchmark for approximate nearest neighbor search methods, or vector indexes, using modern embedding datasets."}]},{"key":"primary_language","value":"Python","evidence":[{"path":"repository.language","excerpt":"Python"}]},{"key":"license","value":"MIT","evidence":[{"path":"repository.license","excerpt":"MIT"}]},{"key":"topics","value":"approximate-nearest-neighbor-search, benchmark, embedding-vectors, embeddings, nearest-neighbor-search, python, vector-database, vector-index, vector-search","evidence":[{"path":"repository.topics","excerpt":"approximate-nearest-neighbor-search, benchmark, embedding-vectors, embeddings, nearest-neighbor-search, python, vector-database, vector-index, vector-search"}]},{"key":"source_count","value":"1","evidence":[{"path":"projectRevision.sources","excerpt":"github:vector-index-bench/vibe"}]}],"taxonomy":{"category":"development_tools","subcategory":"developer_resources","confidence":0.65,"evidence":[{"path":"repository.description","excerpt":"Vector Index Benchmark for Embeddings (VIBE) is an extensible benchmark for approximate nearest neighbor search methods, or vector indexes, using modern embedding datasets."}]},"lifecycle":{"stage":"active","confidence":0.75,"evidence":[{"path":"repository.pushedAt","excerpt":"2026-08-10T15:42:37.000Z"}]},"userIntents":[{"slug":"build_software","label":"构建或改进软件","confidence":0.65,"evidence":[{"path":"repository.description","excerpt":"Vector Index Benchmark for Embeddings (VIBE) is an extensible benchmark for approximate nearest neighbor search methods, or vector indexes, using modern embedding datasets."}]},{"slug":"create_content","label":"创作内容或设计","confidence":0.65,"evidence":[{"path":"repository.description","excerpt":"Vector Index Benchmark for Embeddings (VIBE) is an extensible benchmark for approximate nearest neighbor search methods, or vector indexes, using modern embedding datasets."}]},{"slug":"analyze_data","label":"分析与理解数据","confidence":0.65,"evidence":[{"path":"repository.description","excerpt":"Vector Index Benchmark for Embeddings (VIBE) is an extensible benchmark for approximate nearest neighbor search methods, or vector indexes, using modern embedding datasets."}]},{"slug":"learn_explore","label":"学习与探索","confidence":0.65,"evidence":[{"path":"repository.description","excerpt":"Vector Index Benchmark for Embeddings (VIBE) is an extensible benchmark for approximate nearest neighbor search methods, or vector indexes, using modern embedding datasets."}]}],"readme":{"headings":["Overview","Results","Publication","Authors","Alternative Benchmarks","Credits","License","Getting started","Requirements","Building library images","Running benchmarks","Plotting results","Creating datasets from scratch","Adding a new method to the benchmark","Adding a new dataset to the benchmark","Running the website locally","Evaluation","In-distribution datasets","Out-of-distribution datasets","Deprecated datasets"],"commands":[]},"changeSet":[{"field":"repository","beforeHash":"ee3b8b38f35dd2ccf130743e44985b11103e1e6ca533bd1aff4474e2335073de","afterHash":"4bd06e27caa1e8d9a61ea7823198621a5b43a5d9cd9f95d08895490d3d76bd88"}]},"publishedAt":"2026-08-16T12:17:16.465Z","updatedAt":"2026-08-21T13:26:54.604Z","summary":{"id":"cmt2zh3le01v06271etmi8i4w","projectType":"library","oneLine":"评估向量索引性能的基准测试框架","overview":"VIBE 是一个可扩展的基准测试，用于评估近似最近邻搜索方法（向量索引）在现代嵌入数据集上的性能。它包含分布内和分布外数据集，支持量化、GPU 算法以及 HPC 环境，并提供结果网站和可视化工具。适合研究人员和工程师评估和比较向量搜索算法。","highlights":["支持分布内/外数据集","支持量化与GPU算法","支持HPC环境","提供结果网站与可视化"],"bestFor":["评估向量索引性能","比较不同搜索算法","研究向量搜索方法"],"caveats":["需要Linux和Apptainer/Singularity","部分算法需要AVX-512或Intel CPU","基准测试耗时较长"],"categorySuggestion":"development_tools","subcategorySuggestion":"testing_quality","lifecycleStages":["testing_quality"],"audienceIntents":["evaluate","build","learn"],"confidence":0.9,"evidence":[{"claim":"VIBE 是一个可扩展的向量索引基准测试，使用现代嵌入数据集评估近似最近邻搜索方法。","sourceType":"description","locator":"https://github.com/vector-index-bench/vibe"},{"claim":"包含分布内和分布外数据集，以及多种最先进的向量搜索算法。","sourceType":"readme","locator":"https://github.com/vector-index-bench/vibe"},{"claim":"支持 8 位整数和二进制精度的量化数据集，以及 GPU 算法。","sourceType":"readme","locator":"https://github.com/vector-index-bench/vibe"},{"claim":"支持 HPC 环境，包括 Slurm 和 NUMA。","sourceType":"readme","locator":"https://github.com/vector-index-bench/vibe"},{"claim":"结果可在官方网站查看，并提供多种工具和可视化。","sourceType":"readme","locator":"https://vector-index-bench.github.io"},{"claim":"基准测试在 Intel Xeon Gold 6230 CPU 上运行，GPU 算法使用 NVIDIA V100。","sourceType":"readme","locator":"https://github.com/vector-index-bench/vibe"}],"publishedAt":"2026-08-17T13:50:12.315Z"},"github":{"url":"https://github.com/vector-index-bench/vibe","ownerAvatarUrl":"https://avatars.githubusercontent.com/u/211645529?v=4","starsCount":52,"forksCount":12,"language":"Python","syncedAt":"2026-08-17T13:46:51.522Z","license":"MIT","defaultBranch":"main","archived":false,"disabled":false,"lastChangedAt":"2026-08-10T15:42:37.000Z","missing":false},"tags":[],"updates":[]},"entryPoint":"discovery","claim":{"state":"unclaimed","verified":false},"initialView":"beautified","community":{"viewer":null,"state":{"likesCount":0,"creativeCount":0,"followersCount":0,"commentsCount":0,"liked":false,"creative":false,"followed":false},"comments":{"comments":[],"nextCursor":null}}}]}],"$L25"]