cd /news/ai-tools/show-hn-ferrovec-retrace-semantic-se… · home › topics › ai-tools › article
[ARTICLE · art-171473] src=github.com ↗ pub= topic=ai-tools verified=true sentiment=↑ positive

Show HN: Ferrovec Retrace – semantic search over pages you've read, all local

Developer singhpratech released Ferrovec Retrace 0.1.1, a Chrome Manifest V3 extension that stores the readable text of pages a user visits, embeds it with a bundled MiniLM model, and enables local semantic search with no account, server, or API key. The extension requires Chrome 116 or newer, ships as a 36 MiB unpacked build (22 MB model, 13 MB onnxruntime wasm), skips pages under about 400 characters, password fields, blocklisted hosts, and incognito windows, and calls ferrovec's compact() on a forget so an end-to-end test finds no trace of the forgotten text, passage id, URL, or title in its on-disk store files.

key takeaway

Ferrovec Retrace 0.1.1 bundles a MiniLM embedding model and onnxruntime wasm into a 36 MiB Chrome MV3 extension that indexes up to 12 passages of about 200 words per page locally, and calls ferrovec's compact() on forget so the rebuilt index and write-ahead log contain no trace of the removed text.

read7 min views4 publishedNov 18, 2026
Show HN: Ferrovec Retrace – semantic search over pages you've read, all local
Image: Michielbdejong (auto-discovered)

Semantic search over every page you have read. Stored only in your browser. No account, no server, no API key.

Ferrovec Retrace is a Chrome (Manifest V3) extension built on ferrovec. It saves the readable text of the pages you visit, turns it into vectors with a small embedding model that ships inside the extension, and lets you search it by meaning. Search for "that article about why the build got slow" and it finds the page even if those words never appear in it.

Status: version 0.1.1, on the Chrome Web Store. You can also load it unpacked (see below).

There are three ways.

Install it from the Chrome Web Store. It needs Chrome 116 or newer.

  1. Download ferrovec-retrace-0.1.1-chromium.zip from theFerrovec Retrace 0.1.1 release .
  2. Unzip it.
  3. Open chrome://extensions , turn onDeveloper mode , clickLoad unpacked , and choose the unzipped folder. It needs Chrome 116 or newer.
cd js && npm install && npm run build

cd ../extensions/retrace
npm install
npm run fetch-model   # downloads the MiniLM model files once, to bundle them
npm run build         # writes extensions/retrace/dist

Then open chrome://extensions, turn on Developer mode, click Load unpacked, and choose extensions/retrace/dist. It needs Chrome 116 or newer.

The unpacked extension is 36 MiB (22 MB of that is the model, 13 MB the onnxruntime wasm).

  • Type rt , a space, and what you remember into the address bar. For example:rt tutorial about sqlite wal mode .
  • Or click the toolbar icon to open the search page.
  • Results are grouped by page. Strong matches are shown first; weaker matches sit behind a disclosure.
  • The privacy controls (below) are on the search page and in the extension's options.

Retrace only knows about pages you open after installing it. It does not import your browser history.

Retrace runs a content script on http and https pages. For each page it:

  • extracts the readable text with Mozilla Readability;
  • skips pages with fewer than about 400 characters of text;
  • skips any page that has a password field;
  • skips hosts on the blocklist;
  • never runs in incognito windows ("incognito": "not_allowed" );
  • splits the text into passages of about 200 words, at most 12 per page, and stores each passage with { url, title, host, visitedAt, chunk, of } .

Not captured: incognito windows, file: and other non-http pages, and single-page-app navigations after the first load (the content script runs once per document load).

  • ** and resume.** While d the toolbar badge shows "off" and nothing is indexed.
  • Blocklist. Editable, one pattern per line. The defaults cover banking and payments, webmail and messaging, health, and sign-in hosts. The list is deliberately broad; remove any entry you want indexed.
  • Forget this site ,Forget the last hour andDelete everything .
  • Stats on what is stored, andJSON export .

Removing a document from a search index usually only hides it. After a forget, Retrace calls ferrovec's compact(), which rebuilds the index and rewrites both snapshot slots and the write-ahead log. The end-to-end test then searches Retrace's on-disk store files and finds no trace of the forgotten text, passage id, URL or title.

Two caveats. The small per-URL bookkeeping record (URL and title) is deleted from IndexedDB, but its bytes can remain in IndexedDB's LevelDB log file until Chromium compacts it. And Chromium's own history and cache still know the page; that is outside Retrace.

page -> content script (Readability, skip rules)
     -> service worker (queue, settings, omnibox)
     -> offscreen document: ferrovec worker + OPFS + wasm HNSW + MiniLM
     -> search page (grouped results)
  • The content script extracts text and sends it to the service worker.
  • The service worker applies the and blocklist rules, chunks the text, and queues the work. The queue holds 200 pages.
  • The store lives in an offscreen document, because a service worker cannot host the ferrovec worker. It holds a ferrovec store (OPFS, crash-safe snapshots and write-ahead log, HNSW in wasm). It closes itself after 5 idle minutes and reopens on demand.
  • The model (Xenova/all-MiniLM-L6-v2 , q8) and the onnxruntime wasm are bundled in the extension. Nothing is fetched at runtime. The content security policy isscript-src 'self' 'wasm-unsafe-eval' .
  • Permissions:storage ,unlimitedStorage andoffscreen , plus the content script onhttp andhttps pages. There is no host permission beyond the content script and no network permission.

Run in Playwright Chromium (new headless, --load-extension):

Check Result
End-to-end checks 36 of 36 pass
Unit tests 50 of 50 pass
8 paraphrased queries over 8 fixture articles right page first every time
Reopen after a browser restart results persist
Blocked host, password page, d state not indexed
Network every request the browser made went through a logging proxy, cross-checked with Chromium's NetLog: zero requests initiated by the extension
Forget no trace of the forgotten text, id, URL or title in the on-disk store files

The omnibox (rt) could not be driven in headless mode, so the automated test does not cover omnibox suggestions.

One Linux desktop, single-threaded wasm. Your machine will differ.

What Result
Indexing about 110 ms per passage; about 1.1 s for a full 12-passage page (p50 1069 ms, p95 1403 ms); did not slow down as the store grew to 2,851 passages
Query at 2.8k passages p50 19 ms, p95 26 ms
Cold start (store open and model load) 290 to 560 ms
Disk about 3.2 KB per passage per snapshot; steady state about two snapshots plus up to 4 MiB of log, so up to about 6.5 KB per passage; 2,851 passages (320 pages) is an 8.5 MB snapshot
Extension size 36 MiB unpacked
  • Indexing costs about 1 s of CPU per long page, in the background. The queue holds 200 pages and drops the oldest under heavy browsing.
  • Memory use has not been measured.
  • Tested up to about 2.9k passages. Every checkpoint rewrites a full snapshot, so cost grows with the store.
  • Single-page-app navigations are not re-captured.
  • What counts as a close match was measured on 35 real pages (test/RELEVANCE.md ). On queries written after tuning: 84 % of close matches were the right page, 95 % of right pages were shown as close, and 7 of 10 questions about things not in the index showed nothing close. A question about a topic near something you read (the Sphinx, after reading about the Pyramids) can still show that page as close, and a loosely worded query for one of several similar pages (sibling chapters of a book) may show a sibling, or nothing close, with the page in the weaker list.
  • Only the first 12 passages (about 2,600 words) of a page are indexed.
  • Omnibox suggestions are not covered by the automated test.
  • On the Chrome Web Store only. Edge, Firefox and Safari have no store listing.
Browser Status
Chrome 116+ Tested
Edge, Brave Expected to work (same engine, nothing Chrome-only in the manifest). Not tested
Firefox Port planned, not built
Safari Not planned
cd extensions/retrace
npm run typecheck   # tsc --noEmit
npm test            # unit tests
npm run e2e         # end-to-end run in Chromium (needs dist built)
npm run bench       # scale benchmark (needs dist built)

Building Retrace added two things to ferrovec. Both shipped in ferrovec 0.6.0, and the extension builds against the local js/ package.

  • Ferrovec.open(name, { transformers: { env, dtype } }) configures transformers.js inside the store's worker, so the model and wasm load from local files with no remote fetch.
  • db.compact() makes removed or replaced documents leave the store's files, not only its results.

MIT, like ferrovec.

── more in #ai-tools 4 stories · sorted by recency
── more on @ferrovec retrace 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/show-hn-ferrovec-ret…] indexed:0 read:7min 2026-11-18 · —