LinguFlow 1.3.1
Release version: v1.3.1
Status: Superseded by v1.4.0
GitHub Release: https://github.com/newtc22222/lingu-flow/releases/tag/v1.3.1
Previous: v1.3.0 · v1.2.1
Production cut of staging since v1.3.0. AI generation gains TOEIC part6 (text completion) and part7 (reading comprehension) passage sets, a generated set in the Question Bank can be edited as a whole, and passages/explanations keep their paragraph breaks — with part7 able to produce a real markdown table for schedules, price lists, and invoices.
Version strings are 1.3.1 on the package, GET /api/health, OpenAPI, and
the arcade footer.
Product highlights
| Area | What's in 1.3.1 |
|---|---|
| AI generation | TOEIC part6/part7 passage sets (not just flat part5 questions); sharper part5 distractor/explanation quality; per-part topic suggestions |
| Question Bank | A set renders as one grouped card with inline [N] blank tags; "Edit document" rewrites the shared passage and every stem in one request |
| Passage rendering | Multi-paragraph passages/explanations keep their line breaks; part7 can render a real markdown table instead of prose |
| Live sittings | AI hints (H) are no longer available mid-exam — rejected server-side before any vendor call |
| Admin | Sortable tables and UI cleanup on the user/feature-flag panels |
New since v1.3.0
TOEIC part6/part7 AI generation
POST /api/ai/generate-questionsnow produces linked passage sets fortoeic/part6andtoeic/part7, reusing the existing shared-passage storage (passage_group) — no schema change.- Part 6 is always exactly 4 stems: three word/phrase blanks
(
blankType: "word") testing grammar, word form, or vocabulary, plus one sentence-insertion blank (blankType: "sentence") whose options are full candidate sentences. Structural validation rejects a candidate whose blank markers aren't ordered____(1)____..____(4)____or whoseblankTypemix doesn't match. - Part 7 is 2–4 stems with
questionType(detail / inference / vocabulary / purpose / not-except) — a set needs at least two distinct types, and wrong options must be a plausible misreading of the passage rather than something irrelevant to it. - The generate modal's "how many sets" is capped at 5 for these parts (flat generation stays capped at 10 — one set is heavier output than one question). Topic suggestions are now scoped per exam part: part6/part7 offer passage-genre topics (office-memo, business-email, invoice-billing, travel-itinerary, and 10 more) instead of falling back to generic grammar/vocabulary topics.
Sharper part5 prompt quality
- Distractors must share the same word family (word-form items) or be plausible near-synonyms (vocabulary items) — never eliminable by meaning alone.
- Explanations address every option by letter, not just the key.
- Sentences stay single-topic in a workplace register (memos, meetings, invoices, schedules); a batch is required to vary its grammar point and its correct-answer letter rather than repeating either.
easy/medium/hardnow have concrete definitions (e.g. "hard" means a subtle collocation or a referent separated from the blank by another clause) instead of being left to the model's own judgment.
Editable passage sets
- A generated or manually-authored Part 6/7 set now renders in the
Question Bank as one card: the shared passage (with the same
[N]blank tags the exam view uses) followed by its stems, instead of N disconnected flat rows. - New
PUT /api/questions/sets/{passageGroup}rewrites the shared passage and every stem's own content in one transaction. Stem count/ids must match exactly — this is a content edit, not a restructuring tool. Per stem, it reuses the existing invariant: once a stem has been answered, its options/correct answer/answer key are frozen (409), but its wording and the shared passage always stay editable.
Paragraph- and table-aware passages
- Multi-paragraph passages and explanations no longer collapse into one
run-on block —
white-space: pre-wrapnow applies everywhere they're shown (exam-taking view, Question Bank preview, post-exam results review), matching what the exam-taking passage pane already did. MarkdownRenderer.vue(the existingmarked+DOMPurifycomponent already used for flashcards and AI hints) gained a left-align mode, and any passage document with no blank markers — every part7 passage, since blanks only ever occur in part6 — now renders through it instead of plain text. Part6's interactive blank-button rendering is unchanged.- The generation prompt can now ask for a real markdown table (schedules, price lists, invoices, itineraries) instead of folding structured data into a sentence — verified against a live provider, including catching and fixing an initial version that produced table rows joined on one line with no newlines between them.
Live sittings
- AI hints (
H) are no longer available during an in-progress exam.POST /api/ai/hintrejects an in-progress session before any vendor call, and the hint control is unmounted from the sitting UI. The control and spoiler-filter helpers stay in the codebase for a future practice mode.
Admin
- Sortable tables and a UI cleanup on the user and feature-flag admin panels.
Deploy notes (production)
No migration. Nothing new under backend/alembic/versions/ since
v1.3.0 — the passage-set update path reuses the existing Question table.
Seeder: no seed-version bumps since v1.3.0. No template is rewritten and no questions are archived.
Env: no new variables since v1.3.0.
Confirm GET /api/health reports "version": "1.3.1".
Not in this release
- Audio and photography for the TOEIC L&R sample
- Grading of written responses — essays are still captured and excluded from the denominator, with no reviewer flow (#20)
- Speaking — no mic capture, no sitter-side upload, no audio playback in results (#79, #89, #90, #91)
- Table/chart rendering for part6 — blanks-only prose by design; tables are a part7-only capability
- The Question Bank set-card's per-stem stem text is still a single-line ellipsis truncation, not wrapped
.xlsxupload, Google Sheets URL fetch, Redis cache-aside- Scaled and band scoring — scores remain a raw percentage
Shipped PRs
- #110 this release (staging → main)
Issues closed
None tracked; this work came from iterative design and prompt-quality review rather than filed issues.
Tag: v1.3.1 on main.