Skip to content

Forma 1.4.0

Choose a tag to compare

@waqarsyd waqarsyd released this 10 Nov 14:12
· 102 commits to main since this release

Sharper reading, a free fallback, and a chat that knows your report (10 November 2026)

Forma reads a picture of a page with PaddleOCR, which finds far more of its text than before, writes reports with Google’s Gemini 3.x Flash models, moves to the next free AI you set up when one reaches its limit, and answers questions about your report and about Forma from the report and Forma’s own documentation.

New

  • Report design. When the AI you chose reaches its limit, the next one you have set up answers on its own (OpenRouter, then a local AI, after Gemini), for a report and in the chat alike, and the reply says which answered. A request to a local AI is never sent anywhere else, and when every AI gives out, Forma’s own draft is your report as before.
  • Workspace. Settings suggests a pair for Ollama on your computer, a model that writes reports and reads pictures and a page reader, glm-ocr, sized for 8 GB or 16 GB of memory, with the commands that fetch them and one button that fills them in.
  • OCR. A page reader on Ollama or LM Studio, such as glm-ocr, can read a picture’s text beside Forma’s own reading; the AI is told to check against the picture any word the two read differently.
  • Chat. Ask about the open report (which fields it prints, its totals, groups, bands, parameters or paper) and Forma answers from the file itself, with no AI and no request.
  • Chat. Ask how Forma works and the chat answers from Forma’s own documentation, naming the article, instead of guessing at a button that may not exist.
  • Chat. Ask for the report in another language ("translate it into German") and Forma offers a button that translates every typed caption and title in one request, leaving the layout as it is, as a revision beside the original.
  • Chat. Ask about the data (which row is highest, what the amounts add up to) and Forma first asks whether the rows the Preview prints may go to the AI, saying how many and which fields; nothing goes until you say so, and the reply says when they went.
  • Preview. Right-click a control and choose Describe for the AI to say in a sentence what it is for; the AI reads it with the report, so a request can name a control by what it does. The sentence is the control’s accessible description, which DevExpress’s own checks ask for.
  • Chat. Three files or more, or nine pages, get a plan first: tick the files, choose one report or one for each, see what it will cost, then start. Nothing is read or sent before.

Improved

  • Report design. Forma uses Gemini 3.x Flash (and Flash-Lite for quick answers) when it picks a model for your key, as Google retires Gemini 2.5 Flash on 1 December. It reads as well as Gemini 3 Pro on documents, on the free tier.
  • OCR. An English picture or scan is read by PaddleOCR in your browser: on real report pages it found 97% of the printed characters where the reader before found 61%, so Forma’s draft and the AI start from far more of the page. Other languages are read as before, and nothing you upload leaves your browser.
  • Report design. A picture of a page is read at high resolution, so small print and thin lines come through. A PDF is not, since its text is exact already.
  • Report design. After a report from a picture, Forma checks it kept the captions and titles it read on the page, asks the AI once more, naming them, when several are missing, and the REPX pane lists any still left out.
  • OCR. The words the new reader is unsure of are flagged for review at its own threshold, so the review lists the ones most likely to be wrong.
  • Report design. Forma’s own draft of a picture is read by a box reader trained again, on every layout of the company’s and on pages DevExpress printed: on real reports it finds footer totals three times as often, group headers and footers far more often, and builds a banded draft for more pages. Page numbers at the foot of a page are kept as page numbers.
  • Workspace. Settings says what a free OpenRouter model’s provider may do with what you send, and where your OpenRouter account decides which providers it may use; when those settings rule out a model, Forma says which setting to change instead of that the model does not exist.
  • Chat. The AI is told the report’s counts and names rather than counting them, and Forma will not bind or group a field the report’s data does not have; a reply that names one says it would print nothing.
  • Workspace. A batch is kept in your browser as it runs, so after a limit, a reload or the next day it carries on from where it stopped; a limit pauses it rather than failing file after file, and Clear removes it.
  • Report design. Forma’s own draft of a picture, the report you get when every AI has reached its limit, now draws a records region as a table, keeps the captions of totals lines such as Subtotal and Grand Total, and sizes its text as the page printed it rather than up to twice as large.

Fixed

  • Report design. A repair request that hits a limit keeps the report the AI already wrote, instead of falling back to Forma’s draft.
  • Report design. Forma no longer spends a request trying Gemini 2.0 Flash, which Google has shut down, when it first picks a model for your key.

Every release, with a feed: https://openforma.web.app/changelog