Changelog
This page records what shipped in each released version of the site and the tool, what was knowingly left unfinished, and what changes next. There is one entry so far: v0.1, published 17 September 2026. Nothing earlier exists, and no entry here is backfilled to look like a longer history.
Last updated:
v0.1 is the first public version, released on 2026-09-17. It ships a browser-local document question-answering tool: files are parsed, chunked, embedded, indexed and searched inside the tab, the embedding model is a one-time download of 23 to 25 MB, and a larger optional generation model is available behind an explicit confirmation step.
v0.1 — 17 September 2026
What shipped
| Area | What is in v0.1 |
|---|---|
| Import | Magic-byte sniffing before parsing, then PDF, DOCX, TXT, Markdown, CSV, HTML and JSON. PDF page numbers and DOCX heading levels are preserved, so a citation can name a page or a section. |
| Preparation | Normalisation that drops repeated headers and footers, followed by structure-aware chunking at roughly 700 characters with 15% overlap. |
| Embedding | Local vector computation in a Web Worker, on the GPU through WebGPU where the browser exposes it, otherwise on the CPU through the WebAssembly build of onnxruntime. |
| Storage | Text chunks and vectors in IndexedDB for the site, model weights in Cache Storage. Clearing site data removes both, and there is no server-side copy. |
| Retrieval | Hybrid search combining vector similarity with BM25, fused by reciprocal rank fusion at k=60, re-ranked by maximal marginal relevance at lambda 0.7, with the top 6 passages passed to the answering step. |
| Answering | Extractive answers by default, assembled from retrieved sentences with citations pointing back to a document, a page or a section. A strictness setting controls how much interpretation is allowed. |
| Optional generation | Qwen2.5 0.5B or 1.5B, 4-bit quantised, 400 MB to 1.0 GB, downloaded only after a dialog names the size and you confirm it. Roughly 3 to 8 tokens per second on the CPU path. |
| Embedding models | bge-small-zh-v1.5, about 25 MB, 512 dimensions, default for Chinese; all-MiniLM-L6-v2, about 23 MB, 384 dimensions, default for English; multilingual-e5-small, about 120 MB, 384 dimensions, for mixed-language libraries. |
| Limits | 25 MB per file, 40 files, 200 MB and 20,000 chunks per knowledge base, with the file cap lowered to 10 on phones. |
| Offline behaviour | After the model weights are cached, question answering works with no network. The model download is the only outbound request the tool makes, and it carries a file name. |
| Language | Every page is published in English and Simplified Chinese, with the language switch in the header and hreflang links on both versions. |
| Keyboard and screen readers | A skip link, labelled navigation and panels, a visible focus outline, live regions announcing indexing and generation progress, and reduced-motion support. The accessibility page lists what is still missing. |
| Advertising | Google AdSense, up to six labelled slots per page (three in-page banners, two side rails, one collapsible anchor), never inside the chat workspace. Personalized ads load only with consent; elsewhere a refusal narrows requests to non-personalized ones. |
| Taking work out | A session exports from the chat panel as Markdown, JSON or plain text, carrying the model name, the strictness setting and every citation. Answers also copy to the clipboard, and browser printing produces a PDF of any page. No part of that export touches a server. |
Known limitations in v0.1
These are the limits shipped knowingly, not defects discovered later:
- Scanned PDFs with no text layer are rejected. No OCR is bundled, so a scan has to be converted elsewhere first.
- Encrypted PDFs are rejected, with no password prompt, because there is no decryption path in the app.
- Legacy .doc files are not supported, and any single file above 25 MB is refused.
- Tables in PDFs are read as text only: column alignment is lost, while CSV rows keep their structure.
- Generation on the CPU/WASM path runs at roughly 3 to 8 tokens per second, and generation is off by default on phones.
- The model download dialog does not trap keyboard focus or close on Escape, and the mobile Docs/Chat/Settings switcher does not implement the tablist pattern. Both are listed on the accessibility page as outstanding.
- The index lives in one browser profile. There is no cloud sync, so another machine means re-importing the files.
- Accuracy is not guaranteed. Every answer carries citations so it can be checked, and the check is the reader’s.
What is next
The first work is the two accessibility gaps named above: keyboard focus handling in the confirmation dialog and a proper tablist for the mobile panel switcher. Beyond that, priority comes from reports sent to the mailbox, and fixes are published here rather than promised in advance.
No dates are attached to future work and no roadmap is published, because a roadmap with dates that cannot be met is worth less than a record of what actually shipped. This page is that record.
How this page is written
- One entry per released version, headed by the version number and an ISO date in the form 2026-09-17.
- Version numbers move when shipped behaviour changes, not when a page is edited. A typo fixed on the models page is a correction to that page, not a release.
- A limitation that gets fixed is removed from the entry that listed it and named in the entry where the fix shipped, so a reader can see when a gap closed.
- A limitation that turns out to be wrong is corrected in place and noted here, including when the correction makes the project look worse.
- Security fixes are described after they are live, with the mechanism kept vague until users on the current version are no longer exposed.
Reporting something this page gets wrong
If a feature listed above does not behave as described, that is a defect worth reporting rather than a changelog wording problem. Write to guweiicy@gmail.com with your browser and version, the steps you took and what happened, and the entry gets corrected along with the fix.