Skip to content

AI turns documents or topics into real, native PowerPoint decks—with native shapes, transitions and animations, data-backed charts and tables on demand, audio narration from speaker notes, and support for your own .pptx templates. · by Hugo He

docs(roadmap): record web entity-safety gate and native image size cap

hugohe3148b326be6ed
diff --git a/docs/roadmap.md b/docs/roadmap.md
index 238f4b62..3b46ef3d 100644
--- a/docs/roadmap.md
+++ b/docs/roadmap.md
@@ -72,6 +72,10 @@ The past two months' structural capability growth. Single flags / incremental po
- **Web image sourcing → best-match + reviewable + manual replace** — web rows no longer silently download a candidate pool; the default downloads **only the best match**, with the pool demoted to an opt-in `--save-candidates` escalation (default 4). Every download writes a ≤1024px review copy under `images/.review/` (placed / promoted asset stays full-resolution). Suitability review is **model-agnostic**: a multimodal model reads the review copy, a non-multimodal one hands the `source_page_url` to a human — no assumption of vision. New `image_search.py --from-url <url>` downloads any human-chosen image URL into the target (recorded `license_tier: manual`, inheriting page context) as the universal manual-replacement path; `--promote` now recomputes the credit from the chosen candidate (never the replaced image's). It all runs inside Step 5 and never halts the pipeline (a poor fit becomes `Needs-Manual` + placeholder). Web search is positioned as a placeholder-grade fallback that does **not** guarantee quality — high quality comes from AI generation or picking an image by hand and swapping it in
+- **Web image entity-safety gate (exact subjects no longer lose to a "hi-res wrong image")** — building on the row above: web candidates now pass a `required_terms` entity gate that blocks metadata-relevant-but-wrong-subject images (a polished Rome monument winning a "Chongqing landmark" row). Groups are ANDed, `A|B` gives intra-group aliases (cross-language `Chongqing|重庆`), and matching is lowercased / separator-normalized / space-compacted so multi-word and CJK names match; an entity hit counts as the relevance signal (a zero query-token overlap no longer rejects, so a CJK-titled landmark passes under an English query — the old reject still applies when no `required_terms` are set). Pixel area drops from a dominant score (cap 5000) to a tie-breaker (cap 1500) plus a title-hit boost, so entity accuracy and relevance outrank raw resolution. CLI `--require-terms` (repeatable, comma / `|`), batch `required_terms`, and `--from-url` all inherit it, recorded in `image_sources.json` for audit; a toothless gate (weak `required_terms`) emits a warning. It is a **metadata gate** paired with the `.review` visual check, not a visual classifier
+
+- **Native PPTX export caps image media size** — keep generated decks editable without embedding huge source images: new native image sizing modes — `cap` (default) limits only oversized source images to a max dimension, while `display` sizes from the rendered SVG box for more aggressive compression; native export preserves full embedded pixels while SVG/PPT display cropping stays as editable picture-crop metadata; `finalize_svg` keeps its existing slice/meet behavior and adds default rendered-size downsampling for compact SVG snapshots. Docs land the `cap` / `display` modes and the `--no-image-optimize` escape hatch
+
---
## In progress / Next
diff --git a/docs/zh/roadmap.md b/docs/zh/roadmap.md
index 9fde0de7..028c4270 100644
--- a/docs/zh/roadmap.md
+++ b/docs/zh/roadmap.md
@@ -72,6 +72,10 @@
- **网络配图改为「最佳图 + 可复核 + 人工换图」** — web 图来源不再默认静默下载一池候选,而是**默认只下最佳匹配图**,候选池退化成 `--save-candidates` 的显式升级路径(默认 4 张);每张下载图生成 ≤1024px review 副本(`images/.review/`,放置 / promote 仍全分辨率)。合适性复核做成 **model-agnostic**:多模态模型读 review 副本自查,非多模态则把 `source_page_url` 交人工判断——不假设模型有视觉。新增 `image_search.py --from-url <链接>`:把人找到的任意图片 URL 下载并替换目标(记 `license_tier: manual`、继承页面上下文),作为通用人工换图通道;`--promote` 改为从被选候选重算署名(不沿用旧图 credit)。全程在 Step 5 内、不合适转 `Needs-Manual` + 占位,**不阻塞主流程**。定位上 web 搜索是「兜底取图、不保证质量」,真要高质量靠 AI 生图或自己手动挑图换入
+- **Web 配图实体安全门(精确主体不再被「高清错图」赢走)** — 承上条 web 配图:给 web 候选加 `required_terms` 实体门控,挡住「元数据相关但主体错误」的图(一张精修的罗马纪念碑赢下「重庆地标」行)。`required_terms` 各组之间 AND、组内 `A|B` 给别名(跨语言 `Chongqing|重庆`),匹配做小写 / 分隔符归一 / 空白压缩以兼容多词与 CJK 名;命中实体即视作相关信号(零 query 词重叠不再否决,CJK 标题地标可在英文 query 下通过,无 `required_terms` 时旧的否决逻辑照旧)。像素面积从主导分(cap 5000)降级为 tie-breaker(cap 1500)+ 标题命中加权,让实体准确性与相关性压过纯分辨率(避免高清错主体赢)。CLI `--require-terms`(可重复,逗号 / `|`)、批量 `required_terms`、`--from-url` 均继承,记入 `image_sources.json` 备审;门控形同虚设(弱 `required_terms`)会发 warning。定位是与 `.review` 视觉复核配对的**元数据门**,不是视觉分类器
+
+- **原生 PPTX 导出图片媒体大小封顶** — 保持生成 deck 可编辑、不嵌入巨幅源图:新增原生图片尺寸模式——`cap`(默认)只对超大源图限制最大边长,`display` 按渲染 SVG 框尺寸做更激进压缩;原生导出保留完整嵌入像素,SVG/PPT 显示裁剪仍走可编辑的 picture-crop 元数据;`finalize_svg` 保留原有 slice/meet 行为,另加默认按渲染尺寸下采样以产出紧凑 SVG 快照。文档落地 `cap` / `display` 两模式与 `--no-image-optimize` 逃生舱
+
---
## 进行中 / 下一步