I build Siftlume, an open-source (MIT) Chrome extension that turns a page into a visual summary: key figures, charts, timelines, checklists. You bring your own API key or a local model, and there is no backend.

When I shared it on daily.dev, a commenter asked a question I could not answer well:

Do the charts and key figures get citations too, or only the follow-up answers? A number the model gets wrong looks more trustworthy in a chart than it would in a paragraph.

That is the whole problem with visual summaries. A chart is a claim wearing a suit. At the time, only the source-quote block could jump back to the page. Charts and key figures had no link at all.

A few releases later, here is what I changed, what a second commenter made me add, and what is still not solved.

1. Every figure carries its source passage

In v0.7.0, key figures, bar lists and charts each carry the passage they came from. Under a chart, a "Where these numbers come from" list shows the passage behind every value, with the figure marked inside it. A "Show in page" button scrolls to that passage in the original page, plays a video from that moment, or opens the right tab.

The model has to give a passage for each number, and that passage is checked against the source text.

2. "Unverified" for numbers that do not check out

A figure is marked Unverified when its passage is not in the source word for word, or when the passage does not contain the figure. A wrong number no longer looks as trustworthy as a right one.

Matching numbers is harder than it sounds. +186.000, +186,000 and 186K are the same figure, so the check has to accept all of them. And 1% must not match 10%. In v0.8.0, extra zeros are accepted only when a scale word is present (186K, 186 nghìn, 1.2 million).

3. PDFs broke the check in boring ways

Checking against web pages worked. Checking against PDFs gave false "Unverified" labels, which is worse than no label, because people learn to ignore it. The causes were mostly text-extraction details:

  • full-width digits (19:00)
  • ligatures
  • soft hyphens and words broken across lines
  • spaces a PDF puts between Japanese or Chinese characters

The text is normalized (Unicode NFKC plus removing invisible characters) before matching. For PDFs, files, pasted text and Google Docs, Sheets and Slides, "Show in source" opens the extracted text pane and highlights the passage.

4. "Unverified" says why

A bare label teaches nothing. Since v0.8.0 it says which case it is:

  • no passage was given
  • the passage is too short to check
  • the passage is not in the source word for word
  • the passage is in the source but does not contain the figure

5. The gap a reader found: numbers the model computes

After v0.7.0, another commenter pointed at the weak spot. When the model computes a number (a total, a percentage change), no source passage contains it, so it showed as Unverified, exactly like a number the model made up. A reader could not tell "possibly wrong" from "computed from correct inputs". And computed figures are the ones most likely to be wrong.

I filed it as an issue that evening and the fix is in v0.8.1:

  • A figure the model works out from others carries its inputs: the source figures it was computed from, each with its own verbatim quote.
  • If every input is in the source and the arithmetic holds, the figure is marked Derived, and it opens its inputs, each with "Show in page". The arithmetic is checked in the browser, with no request to the model. It covers sums, averages, differences, ratios, shares, percentage changes, products and rounding.
  • If an input is not in the source, or the result does not follow from the inputs, it stays Unverified and says which.
  • A figure the source states in words ("no fee" for 0) is marked Inferred and opens those words.

v0.8.1 is merged and heading to the Chrome Web Store, so the store build may be a version behind for a few days.

What is still not solved

  • "Verified" and "Derived" mean the passage is in the source and the arithmetic holds. They do not mean the model understood the passage correctly.
  • A count the model makes itself ("7 competitors") has no inputs to check, so it stays Unverified.

Also in 0.8.0, briefly

  • Guide mode turns how-to pages into steps, and Auto (now the default) picks how to read each source in the browser, with no extra request.
  • Share as an image draws the TL;DR and up to three key figures as a 1200×630 card in your browser, nothing uploaded.
  • What changed since your last read compares a page you summarized before, locally, with no request to the model. Handy for pricing pages and release notes.

Try it, and tell me where it fails

It is MIT licensed: github.com/tamnt1707/siftlume. The "show as text" switch, the per-figure sources and the derived label all came from comments on my first post, so if you try it on a long PDF or a page full of numbers, I would like to hear which figures it got wrong.