Changelog

What shipped on reranker.uk — demo improvements, new guides, and site infrastructure.

15 Oct 2026 (evening) — Light or dark, as your system is set

  • First visit follows your system theme — the site used to open in the dark theme for everyone. Until you use the theme switch, it now follows your operating system’s light or dark setting, and changes with it; once you’ve picked one with the switch, your choice stays
  • Site search — a search button in the header, or press / on any page, finds guides and model pages by title, section headings and description. It runs in your browser from a small index built with the site, the Chinese pages search the Chinese pages, and it works with the keyboard alone: arrow keys move through the results and Esc closes it

15 Oct 2026 — Guides in reading order; text no longer touches the screen edge on phones

  • A reading order for the guides — the guides page now starts with the five guides meant to be read in sequence, numbered, then lists the pages for specific jobs: choosing a model, reranking from mainland China, costs, troubleshooting and the rest
  • Margins on phones, everywhere — on every guide and model page, the text ran right up to the edges of a phone screen: a stylesheet rule meant to set top and bottom spacing was also removing the side margins. Fixed, and a new test checks every page at phone width

14 Oct 2026 (late night) — The models table keeps its first column on a phone

  • Model names stay in view — on a phone the models table scrolls sideways, and the model name used to scroll away with it, leaving scores with no row label. The name column and the column headings now stay put while you scroll; on a desktop the table is unchanged
  • Side margins on phones — the homepage, changelog and demo sections had lost their side margins on phones, so the new pipeline diagram touched the screen edges. They’re back

14 Oct 2026 (night) — Homepage diagram corrected; shortcuts for the job in front of you

  • The pipeline diagram was labelling the wrong arrows — “top 50–100” sat on the arrow into the retriever, and “top 5–10” sat on the arrow into the reranker. Those counts are what each stage returns, not what it receives. The diagram now shows the retriever returning 50–100 candidates and the reranker keeping the top 5, and on a phone it stacks instead of shrinking the labels until they cannot be read
  • Four shortcuts — choose a model by scenario, rerank from mainland China, the cost calculator, and what to check when reranking didn't help. They were already on the site; the homepage didn't point at them

14 Oct 2026 (later) — Every page has its own preview card

  • Share previews that say what the page is — a link to any page used to show the same site-wide cover in WeChat, X or Slack. Each page now has its own card with its title, description and section, in English or Chinese to match the page, and CI checks the cards stay in step with the pages

14 Oct 2026 — New guide: reranking in mainland China

  • Alibaba Cloud, SiliconFlow or your own server — a guide for readers whose servers are in mainland China: what each route offers, prices in yuan, Qwen’s Chinese retrieval scores (CMTEB-R) for the open models, working code for Model Studio and SiliconFlow, and how to download weights through hf-mirror.com or ModelScope. Every price and score on it is checked daily against the vendor’s own page

12 Oct 2026 (later) — Every benchmark figure is now checked daily; Mixedbread v1 figures corrected

  • About 70 more figures under the daily source check — the benchmark tables on the Jina and Qwen pages, and the third-party figures quoted in the models table’s notes, were typed into the pages, so a vendor revising a number would have gone unnoticed. Each one is now read from the vendor’s own model card and checked against it every day, like the rest of the site’s numbers. The gte-multilingual row also gains the scores Qwen published for it
  • Mixedbread v1, corrected — the Mixedbread page showed xsmall-v1 and base-v1 at ~55.5 and ~59.8 BEIR, and an MS MARCO column from ~38.0 to ~42.3, none of which we could trace. Mixedbread’s own model cards give 43.9 and 46.9 on their 11-dataset BEIR run, with Accuracy@3, and no MS MARCO figures; the table now shows those

12 Oct 2026 — Older guides brought up to date; framework code re-run

  • Choosing by scenario — the guide recommended a “Voyage code” and a “legal-tuned” Voyage reranker that don’t exist, and Cohere v3.5 for multilingual work. It now points at current models, with new sections for Chinese and multilingual search (including Alibaba Cloud and SiliconFlow in mainland China) and for scanned PDFs and screenshots (Qwen3-VL-Reranker, LightOn-rerank)
  • Code that runs on today’s libraries — we re-ran the RAG guide’s framework examples. LangChain 1.0 moved the reranking retriever to langchain-classic, and Haystack 3 no longer ships a cross-encoder ranker in its core package, so both snippets failed with an import error; they are fixed and were run on LangChain 1.4, LlamaIndex 0.14 and Haystack 3.3. The examples now use bge-reranker-v2-m3 instead of the 2023 bge-reranker-base, and the Jina API example uses jina-reranker-v3.5
  • Self-hosting — the model table starts from bge-reranker-v2-m3, drops an unsourced “2026 open SOTA” label, adds gte-multilingual-reranker-base, points to ready-made servers (Text Embeddings Inference, Infinity), and works out memory from parameter counts instead of a guess
  • Instruction-following — the guide adds Alibaba Cloud’s instruct parameter and Mixedbread’s listwise model. The ColBERT guide’s “updated” date now shows its 9 October change

9 Oct 2026 (night) — The cost calculator covers Google, Alibaba Cloud and SiliconFlow

  • Five more prices in the cost calculator — Google Vertex AI ranking (per query of up to 100 documents) and Alibaba Cloud’s international qwen3-rerank (per token) join Cohere and Voyage in the dollar table. A new yuan table compares Alibaba Cloud’s Beijing-region qwen3-rerank / qwen3.7-text-rerank with SiliconFlow’s bge-reranker-v2-m3, and warns when your passages are longer than qwen3-rerank accepts. Every price is the same daily-checked figure the models table shows

9 Oct 2026 (evening) — Alibaba Cloud, SiliconFlow and Google’s ranking API in the models table

  • Hosted rerankers you can buy in China, and Google’s — the models table now lists Alibaba Cloud Model Studio (qwen3-rerank at ¥0.5 per million tokens in China or $0.1 internationally, the 201-language qwen3.7-text-rerank, and qwen3-vl-rerank for images and video), SiliconFlow (bge-reranker-v2-m3 free, Qwen3-Reranker 0.6B / 4B / 8B on the same endpoint) and Google Vertex AI ranking ($1.00 per 1,000 queries). Prices come from each vendor’s price page and are checked daily
  • The release watch now catches qwen3.7-text-rerank — Alibaba’s newest reranker didn’t match the watch’s pattern for Alibaba model names; it does now

9 Oct 2026 (later) — Hosted rerankers watched too; Mixedbread’s new listwise model

  • Eleven more vendor pages watched — the daily release check now also reads the model lists of ZeroEntropy, Mixedbread, Alibaba Cloud Model Studio, SiliconFlow, Pinecone, Contextual AI, Amazon Bedrock, Google Vertex AI and NVIDIA NIM, so hosted-only rerankers are noticed too, including ones sold mainly in China
  • Mixedbread’s hosted reranker has moved on — Mixedbread now lists v1 and v2 as legacy models and reranks inside Mixedbread Search with mxbai-rerank-v3.1-listwise, a listwise, instruction-following model whose weights aren’t published. The Mixedbread page says so, with the instruction-following figures from Mixedbread’s announcement, checked daily
  • zerank-2’s licence — the weights on Hugging Face were relicensed to Apache 2.0 in July 2026, while ZeroEntropy’s own docs still call them non-commercial; the models table now notes the conflict. NVIDIA’s smaller llama-nemotron-rerank-500m-v2 is mentioned beside the 1B

9 Oct 2026 — Three rerankers the release watch had missed; ColBERT through sentence-transformers

  • The release watch was looking in too few places — it only checked 14 organisations on Hugging Face, and only models with “rerank” in the name. It now checks every model Hugging Face tags as a reranker, whoever published it, plus popular ones the site has never mentioned, and new releases of the libraries the demo and guides use
  • Qwen3-VL-Reranker, zerank-2 and gte-multilingual-reranker-base in the models table — Qwen’s 2B / 8B reranker for text, images, screenshots and video; ZeroEntropy’s Apache 2.0 4B model, also sold as an API; and Alibaba’s 306M multilingual model, which supports 70+ languages. Each row shows the vendor’s own figures with how they were measured, checked daily against the model card
  • CLM-v0.1-8B, explained rather than added — the most-liked reranker-tagged model on Hugging Face scores actions for agents, not passages for RAG; the models page now says so
  • ColBERT code that runs — the late-interaction guide’s snippet searched a ColBERT index it never built. It now rescores candidates with sentence-transformers 6’s MultiVectorEncoder, which we ran on version 6.1.0

5 Oct 2026 — Voyage rerank-3, what changed; LightOn-rerank added

  • Voyage rerank-3, from Voyage’s announcement — the site said Voyage’s docs didn’t say whether rerank-3 takes natural-language instructions. Voyage’s 30 September announcement does: the Rerank 3 series keeps them. The Voyage page now covers what changed — a drop-in upgrade at the same price, scores calibrated to rerank-2.5, the largest gains on long documents and code, and Voyage’s own comparison against Cohere and Qwen3 — with each figure checked daily against the announcement
  • LightOn-rerank in the models table — LightOn’s Apache 2.0 family (0.8B / 2B / 4B, pointwise or listwise) scores text passages and document-page images with one model. The table shows LightOn’s own BEIR figure with its protocol, which differs from the other rows
  • A daily watch for new releases — every morning the site now checks Hugging Face and the Cohere, Voyage and Jina model pages for rerankers it doesn’t cover yet, and adds them after reading each vendor’s own announcement

3 Oct 2026 — The RAG guide’s model picks, corrected; the Chinese pages, finished

  • “Which reranker to pick” was out of date — the RAG guide called bge-reranker-v2-m3 English-only (it’s BAAI’s multilingual model), recommended Cohere Rerank v3.5 for “best multilingual quality” and Voyage Rerank 2. It now points at Cohere Rerank 4 and Voyage rerank-3, describes each by how it bills and what it supports rather than with an unsourced superlative, and its code sample uses Cohere’s current client and model
  • English left on the Chinese pages — five guides’ tables of contents, the RAG pipeline diagram, every source note on the model table’s BEIR column, the demo’s input hints and screen-reader labels, and several badges were still in English. Tables of contents now take their wording from the translated headings, so they can’t fall behind again

2 Oct 2026 — Structured data that matches the page

  • Hidden FAQs removed — five pages carried FAQ markup for questions the page never shows, and it had gone stale: it still said Jina v3 beats Qwen3-Reranker-4B and quoted a Hit@1 tie, both withdrawn from the visible pages. Search engines and AI assistants read that markup, so it’s gone (Google stopped showing FAQ results for sites like this in 2023 anyway)
  • Chinese pages describe themselves in Chinese — the /zh/ pages shipped the English structured data verbatim, some declaring their language as English. Headlines, descriptions, language and breadcrumbs now come from each page’s own title and description at build time, so they can’t drift from the page again; a CI check enforces it

1 Oct 2026 — The demo recovers when the mirror can’t serve

  • Automatic fallback between model hosts — the demo picks whichever of huggingface.co and hf-mirror.com answers first. Answering isn’t serving: outside mainland China, hf-mirror.com redirects every file to huggingface.co without the CORS header a browser needs, so if it won the race the model never loaded. A load that fails at the network level is now retried once on the other host
  • Tested against the real hosts — the daily real-network check now also loads each model with hf-mirror.com winning the race, and confirms the fallback ranks correctly. Every page is also checked for CSP violations on every pull request, not just the demo

1 Oct 2026 — The demo scores about twice as fast

  • Multi-threaded inference in the demo — the demo pages are now cross-origin isolated, which lets ONNX Runtime Web score on several threads. Measured A/B on the same 4-vCPU runner: about 1.9× faster for all three models (Jina v1 tiny 424 → 229 ms for 10 passages). Browsers that don't support the isolation header simply stay single-threaded
  • Only where it's needed — isolation and a CSP that also allows blob: scripts (the threaded runtime loads that way) apply to the demo pages alone; every other page keeps the stricter policy. A new daily check confirms both on the live site

1 Oct 2026 — Latency figures measured, or sourced, or gone

  • The browser demo, measured — a new script runs the live demo in Chromium on a 4-vCPU GitHub Actions runner. Scoring 10 passages takes about 430 ms with Jina v1 tiny, 515 ms with ms-marco MiniLM and 1.4 s with mxbai xsmall. The site had claimed 30–80 ms and 50–150 ms
  • Demo model sizes corrected — the picker listed mxbai xsmall at 70 MB and ms-marco MiniLM at 90 MB; the demo actually downloads 92 MB and 23 MB. So ms-marco was labelled "larger" when it's the smallest, and mxbai "balanced" when it's the largest and slowest. The "try a smaller model" button now picks the actual smallest
  • Sourced where vendors publish — mixedbread's own A100 latencies on the mxbai page, and Jina's API latency figures on the Jina page
  • Removed where nobody measured — the CPU/GPU millisecond ranges for bge and mxbai in the models table, Cohere's "~100–200 ms", and the self-hosting guide's T4 figure had no source; they're now described without numbers, and the RAG guide's latency table says which rows are estimates

1 Oct 2026 — Qwen3, Contextual AI, GTE and NVIDIA rows checked against their model cards

  • Contextual AI Rerank v2 was described wrongly — the table called it an English-only hosted API with prices on request. It has shipped open weights on Hugging Face since August 2025 (1B, 2B and 6B), covers 100+ languages with a 32K context, and is licensed CC BY-NC-SA — non-commercial
  • Two unsourced scores removed — the NVIDIA row's "Hit@1 83.0" and the GTE row's "ties nemotron on Hit@1" appear in neither model card. NVIDIA report Recall@5 for their own embed + rerank pipeline; GTE now shows Alibaba's own 56.19 BEIR
  • Qwen3 numbers now sourced — the table shows Qwen's own MTEB-R (65.80, 69.76, 69.02). "The 4B edges the 8B by ~0.48 on BEIR" is corrected to 0.74 on MTEB-R, with the 8B ahead on Chinese, multilingual, long-document and code. The Qwen page had still been comparing against "bge ~60" and "mxbai ~62", both withdrawn earlier; it now shows Qwen's full table
  • Qwen3 takes instructions — its page said "maybe not" for instruction-shaped relevance, but the Qwen rerankers are instruction-aware. The quick start now uses the model card's own sentence-transformers code, and the instruction-following guide lists three rerankers that take an instruction, with real usage

1 Oct 2026 — Cohere's 500-token rule, and Voyage rerank-3

  • A correction to our own correction — in September this site called it a myth that Cohere splits documents over 500 tokens into extra billable chunks. Cohere's own pricing FAQ says it does: a document over 500 tokens, query included, counts once per 500-token chunk toward a search's 100 documents. The cost calculator now models that, and the "long passages favour Cohere" rule it used to give is gone
  • Voyage rerank-3 is current, not a preview — Voyage now list rerank-3 and -lite as their recommended rerankers and rerank-2.5 among older models. The Voyage page, models table and calculator lead with rerank-3. The 200M free tokens apply to rerank-3; the 33% Batch API discount and natural-language instructions are documented only for rerank-2.5
  • Prices are sourced and checked daily — Cohere's per-search and Voyage's per-token prices now live in data/models.json; the calculator reads them from the page, and the daily source check confirms them against cohere.com/pricing and Voyage's docs
  • Smaller fixes — Cohere trial keys (1,000 calls a month) may not be used in production, so the scenario guide no longer suggests them for customer support; Cohere's context is 32,768 tokens, not 32,000; the RAG guide's API cost was more than ten times too low; rerank-2-lite's context is 8,000 tokens, not 16,000; Voyage does have a TypeScript library

1 Oct 2026 — jina-reranker-v3.5, and Jina's licences spelled out

  • jina-reranker-v3.5 added — Jina's August 2026 flagship: 0.6B, listwise, a drop-in replacement for v3 and 1.22–1.56× faster. 63.20 BEIR nDCG@10 in Jina's own test. It's in the models table, and the Jina page is rewritten around it
  • Licences stated — the v2, v3, v3.5 and m0 weights are CC BY-NC 4.0, non-commercial. This site called them "open weights" without saying so. The Jina page now has a licence section, and the self-hosting guide's model table a licence column
  • "Jina v3 beats Qwen3-Reranker-4B" withdrawn — true in the v3 paper's 2025 run (61.94 vs 61.16), but in Jina's own 2026 re-run v3 scores 62.10 and the 4B 62.28. The Jina rows in the models table now use that 2026 run, the same one as v3.5
  • Unsourced Jina prices removed — "1M tokens/month free" and "~$0.018 / 1M tokens" had no source we could find. Jina's own page says new keys get free trial tokens and then token packages; we now list their published rate limits instead of a price
  • Smaller fixes — parameter counts written as MB on the Jina and mxbai pages (mxbai-rerank-large-v1 is 435M, not the 1.5B its own README states); jina-reranker-v1-base-en, no longer on Hugging Face, removed

1 Oct 2026 — bge-reranker's BEIR figure, and a daily check on every source

  • bge-reranker-v2-m3 now reads 55.36 — BAAI's own BEIR figure. The ~60.1 the models table carried could not be traced to anything BAAI or anyone else published
  • The bge-reranker page's benchmark table, rebuilt — all four rows' BEIR and MS MARCO figures were untraceable. It now shows what BAAI did publish: BEIR for the v2 models, C-MTEB reranking for v1. The size column was also showing parameter counts as megabytes
  • Jina v1 tiny and ms-marco MiniLM-L6 get real BEIR figures — 48.54 and 48.64, both from Jina's v1 model card (17 BEIR datasets). The ms-marco figure is Jina's measurement, not its authors', and says so on hover
  • The footnote stops claiming a common protocol — it said the bge and Jina rows used "the classic BEIR 18-dataset average". Neither does: BAAI average 15 datasets, Jina 13, over different first-stage retrievers. The same model lands over a point apart between them (bge-reranker-v2-m3: 55.36 by BAAI, 56.51 by Jina), so every BEIR figure now carries a * and its protocol
  • A daily check that sources still say what we say — every figure in the table that is published as text is re-fetched from its source each day and the run fails if the number has gone. Jina's v3 paper has already been revised once (61.85 in the arXiv full text, 61.94 on the model card) — the kind of drift this exists to catch

1 Oct 2026 — An accessibility pass, and a check to keep it

  • Links in text are underlined — they were told apart from surrounding text by colour alone; in the dark theme that was a 1.47:1 difference, well under the 3:1 a reader without full colour vision needs. Buttons, cards and navigation are unchanged
  • Five text colours fixed — the light theme's teal and amber labels and the Chinese dark-theme page labels were 3.1–4.2:1, under the 4.5:1 minimum for body-size text
  • Code blocks reachable by keyboard — a code sample wider than the screen scrolls sideways, and could only be scrolled with a mouse
  • A check on every page, both themes — every build now runs axe-core's WCAG 2.1 AA rules across all 48 pages in light and dark; it found over 400 issues the first time and finds none now

28 Sep 2026 — transformers.js 4.3.0, and what report-only caught

  • The demo runs on transformers.js 4.3.0 — up from 3.5.1, with ONNX Runtime Web 1.22 → 1.31. Verified against live jsDelivr and HuggingFace on all three demo models before merging, not just checked against the source
  • The report-only CSP earned its keep — its first run against real traffic flagged that HuggingFace redirects model weights to separate CDN hosts the policy didn't list. Enforcing, that would have blocked every model download in the demo. Fixed before it was ever enforced
  • …and caught the upgrade too — transformers.js 4 loads part of ONNX Runtime from a blob: URL by default. The demo now switches that cache off instead of loosening the policy for every page
  • A stricter real-network test — the daily check now runs every model in the demo's picker, not just the default, and fails if an obviously relevant passage doesn't outrank an unrelated one, so a model that loads but ranks wrong no longer passes

27 Sep 2026 — Security headers, report-only for now

  • public/_headers added — X-Content-Type-Options, X-Frame-Options, Referrer-Policy, Permissions-Policy, and a Content-Security-Policy, generated at build time rather than hand-maintained
  • CSP ships report-only — this repo's sandboxed dev environment can't reach real jsDelivr/HuggingFace/hf-mirror.com traffic or real WASM instantiation to confirm a stricter policy wouldn't break the demo in practice, so it logs violations instead of blocking anything until that's been checked against the real thing
  • No hand-copied hash — the CSP's script-src allows the site's one inline script (early theme detection) via a sha256- hash computed from its actual content at build time, not typed in by hand where it could drift
  • Both smoke tests now watch for CSP violations — tests/helpers/csp.mjs, asserted in both the mocked and real-network demo tests; verified the mechanism actually catches something by deliberately breaking style-src locally and watching the test fail

26 Sep 2026 — ms-marco MiniLM's BEIR figure was never published

  • ms-marco MiniLM-L6 now reads "not published" — this table carried it at an unsourced ~55.0; sentence-transformers' own docs report NDCG@10 on TREC DL 19 (74.30) and MRR@10 on MS MARCO Dev (39.01) for this model, not a BEIR average. Found while checking the rest of the table after the mxbai correction below — same failure shape, different model

26 Sep 2026 — mxbai-rerank-v2, a corrected BEIR figure, and a demo smoke test

  • mxbai-rerank-large-v1's BEIR score corrected — this table carried an unsourced ~62.1 for months; mixedbread's own comparison table puts it at 49.32*. A reader flagged the gap between that number and mixedbread's own figures for their newer models, which is what surfaced it
  • mxbai-rerank-v2 added — a newer, Qwen2.5-based, RL-trained generation (base-v2 0.5B, large-v2 1.5B) that adds 100+ languages including Chinese; v1 was English-only, which the table didn't make clear
  • data/models.json — benchmark numbers that have been through a correction like the one above now live in one sourced file instead of a hand-typed literal in every page that mentions them; the build refuses to render one with no source and warns when one hasn't been re-checked in 6 months
  • A smoke test for the live demo — npm test runs the demo's load → score → render pipeline against a fake model on every PR; a separate daily job runs it against the real jsDelivr / HuggingFace chain, which is what would actually catch one of those going dark

18 Sep 2026 — URLs without the .html, and two pricing corrections

  • URLs dropped their .html — every internal link, the sitemap, hreflang tags and JSON-LD now point straight at the extensionless URL Cloudflare was already redirecting to, cutting a 307 round-trip off every internal navigation
  • Voyage's free-tier grant moved on — the 200M free-token allowance is no longer on rerank-2.5; it now ships with the rerank-3 preview, added to the model table and cost calculator alongside rerank-3-lite
  • Cost calculator token formula fixed — Voyage bills the query once per document reranked, not once per query; the calculator undercounted tokens (and cost) on any workload with more than one candidate per query
  • Cohere chunking, clarified — an outside audit suggested Cohere silently splits passages over 500 tokens into extra billable chunks; that isn't how it works (context is 32,768 tokens and chunking is opt-in), so the calculator page now says so directly instead of adopting the wrong model
  • Homepage diagram, now bilingual — the "Reranking in one diagram" figure was inline SVG the translation pass couldn't see; its labels and caption now render in Chinese on /zh/
  • Model table accessibility — the comparison table gets a caption, column/row headers, and a scrollable, keyboard-focusable region for screen readers and narrow viewports

25 Aug 2026 — Two guides for the questions people actually ask

  • Reranking didn't help — a diagnostic walkthrough of the seven reasons a rerank stage shows no lift, starting with the one that explains most of them: retrieval never returned the right document
  • Rerank on a vector database — the retrieve-wide-then-rerank pattern with code for pgvector, Qdrant and Elasticsearch, including the HNSW ef_search trap that makes a wider limit return padding instead of candidates
  • /llms.txt — a generated map of the site for assistants that read one before citing a source, including how to read our benchmark footnotes
  • Feed autodiscovery — the changelog RSS is now advertised from every page, not just the changelog

14 Aug 2026 — Cost calculator, and demo links that land somewhere

  • Rerank cost calculator — Cohere bills per search, Voyage per token, so the cheaper vendor flips with passage length and top-k. Put your own volume in and see where the line sits
  • Scenario deep links — all 25 demo links across the guides and model pages now open the scenario the page is actually about, via a new ?s= parameter
  • Shorter share links — an untouched built-in scenario shares as ?s=rag instead of a 900-character URL
  • Translations survive link edits — the build now retargets links inside a translation instead of dropping it back to English

11 Aug 2026 — Real Chinese URLs, and a 2026 model refresh

  • Chinese lives at /zh/ — every page is now pre-rendered in Chinese at its own URL with a self-referencing canonical, instead of a client-side toggle that left all three hreflang tags pointing at one page
  • Lighter pages — translation moved to build time, so ~200 KB of dictionaries no longer ship to the browser
  • Sitemap is generated — built from the page tree with hreflang alternates and lastmod from git, replacing a hand-maintained file that had drifted
  • Cohere Rerank 4 — rerank-v4.0-pro and rerank-v4.0-fast replace v3.5; 32k context, billed per search
  • Voyage rerank-2.5 — 32k context and instruction following; the index table's per-doc pricing was wrong and is now per token
  • Jina v3 numbers firmed up — 0.6B, 61.94 BEIR nDCG@10, 64 docs in a 131K context
  • Honest Qwen3 rows — dropped an unverifiable "~75+" figure for the 8B; reports put the 4B slightly ahead of it
  • Design refresh — fluid type scale, a real elevation ramp, and one focus-visible treatment site-wide

25 Jun 2026 — Model landscape refresh

  • Qwen3-Reranker — 0.6B / 4B / 8B rows + deep review page
  • Jina v3 — listwise flagship; tiny kept for browser demo
  • Table adds — gte-reranker-modernbert-base, NVIDIA nv-rerankqa / Nemotron
  • Self-host / homepage / chooser — GPU default Qwen3-4B; bge remains CPU path
  • Score footnotes — MTEB-R* vs classic BEIR; next review Oct 2026

25 Jun 2026 — Minimal polish pack

  • Honest benchmarks — emerging rows use ≈ / n/a; next review Sep 2026; pricing lag disclaimer
  • Demo max_length — 256 / 384 / 512 tokens; char warnings follow selection
  • Lazy transformers.js — loaded on first Rerank only
  • Instruction-rerank guide — task-shaped ranking
  • Homepage — beyond five families + links to ColBERT / instruction guides

25 Jun 2026 — Content expansion & polish

  • Models table — architecture column; Qwen3-Reranker, Contextual AI, ColBERTv2, ms-marco browser rows
  • Late-interaction guide — ColBERT & when to skip cross-encoder rerank
  • Demo presets — E-commerce + Multilingual (7 scenarios); ms-marco ?m= on models table
  • JSON-LD — inLanguage follows zh/en toggle
  • Changelog RSS — /changelog.rss feed

25 Jun 2026 — i18n & SEO completion

  • Model page i18n — pills, TOC, meta dates, Pros/Cons, Other models for all five families
  • Changelog + Privacy i18n — full Chinese body on both pages
  • hreflang — en, zh-Hans, x-default on every page (same URL, client-side toggle)
  • og:locale:alternate — swaps with primary locale on language toggle

25 Jun 2026 — Low-priority polish

  • Guide i18n — full Chinese body for self-host, scenario, hybrid retrieval, and evaluate guides
  • Compressed share links — demo ?z= gzip when URLs exceed ~1600 chars
  • Preset mobile layout — 2-column grid on narrow screens
  • og:locale — switches to zh_CN when language toggle is 中文
  • Dual-diff a11y — table caption, row headers, empty state, aria-labelledby

23 Jun 2026 — Medium-priority UX & content

  • Self-host guide — sentence-transformers, serving, ops
  • Scenario guide — RAG vs support vs legal vs code
  • Passage char counts — per-list stats + 512-char truncation warning
  • New presets — Technical docs, Code search
  • CSV export — Copy CSV alongside JSON and Markdown
  • Light theme — toggle in nav, persisted in localStorage

23 Jun 2026 — Demo UX round 2

  • Model loading panel — progress %, ETA, file name, cache status
  • JSON passages — paste a JSON array or { "passages": [...] } object
  • Dual-model diff view — aligned table with score and rank deltas
  • Models table — sort, filter, and jump to demo with ?m=
  • Mobile passage editor — add/remove list on small screens
  • Error hints — classified messages for network, WebGPU, memory, and limits
  • Changelog + Privacy — this page and a short privacy statement
  • Nav — Home and Guides links in the top bar

23 Jun 2026 — Full release (plan items 1–5, 7–10)

  • Build system: src/partials + src/pages → scripts/build.mjs
  • Demo: three-column results, bi-encoder proxy, dual-model compare, WebGPU, URL sharing
  • Guides index, hybrid retrieval, evaluate rerankers
  • Models: Last verified June 2026, chooser cards, mxbai on homepage
  • i18n: data-i18n keys + shared dictionary; EN/中文 toggle
  • Footer GitHub link; sticky nav; aria-live status

21 Jun 2026 — Initial launch

  • Educational guides on rerankers, cross- vs bi-encoder, and RAG
  • Model comparison pages for bge, Cohere, Jina, Voyage, mxbai
  • In-browser cross-encoder demo with transformers.js
  • Deployed on Cloudflare Workers (static assets)

Try the demo →