Slide Autosummarize

An Obsidian plugin that turns a PDF slide deck into study-ready Markdown class notes: concise notes per slide, LaTeX equations, and only the relevant figures — picked by the LLM with bounding boxes, cropped out of the slide, and embedded.

Works with Google Gemini (default), OpenAI, Anthropic Claude, and Ollama (fully local). Providers are modular — see Adding a provider.

---

Install

Requires Obsidian 1.4.0 or later.

Option A — Community plugins (once accepted)

1. Obsidian → Settings → Community plugins → Browse.
2. Search Slide Autosummarize → Install → Enable.

Not in the directory yet (submission in progress). Until then, use one of the options below. You can follow the listing at https://community.obsidian.md/plugins.

Option B — BRAT (easiest from GitHub, auto-updates)

BRAT installs plugins straight from a GitHub repo and keeps them updated.

1. Install BRAT from Settings → Community plugins → Browse and enable it.
2. Run Ctrl/Cmd+P → BRAT: Add a beta plugin for testing.
3. Paste: https://github.com/BartolomeoZisa/obsidian-slide-autosummarize
4. Click Add Plugin → enable Slide Autosummarize in Settings → Community plugins.

BRAT tracks the latest GitHub release, so future versions arrive without manual copying.

Option C — Manual install from GitHub (no BRAT)

1. Open the releases page and pick the latest release.

2. Download the three assets: main.js, manifest.json, styles.css (under Assets; do not download the "Source code" archives).

3. Create this folder in your vault:

   <your-vault>/.obsidian/plugins/slide-autosummarize/
   

   • Windows: C:\Users\<you>\Documents\<vault>\.obsidian\plugins\slide-autosummarize\
   • macOS: ~/Documents/<vault>/.obsidian/plugins/slide-autosummarize/
   • Linux: ~/Documents/<vault>/.obsidian/plugins/slide-autosummarize/

   The folder must be named exactly slide-autosummarize and the three files must sit directly inside it (not in a subfolder).

4. In Obsidian: Settings → Community plugins → turn off Restricted mode if needed → click the refresh icon next to Installed plugins → enable Slide Autosummarize.

5. If it does not appear, restart Obsidian.

Direct download links (once a release is published):

• main.js
• manifest.json
• styles.css

Or from a terminal:

VAULT="$HOME/Documents/MyVault"   # <- your vault path
DEST="$VAULT/.obsidian/plugins/slide-autosummarize"
mkdir -p "$DEST"
BASE=https://github.com/BartolomeoZisa/obsidian-slide-autosummarize/releases/latest/download
for f in main.js manifest.json styles.css; do curl -fL "$BASE/$f" -o "$DEST/$f"; done

Windows PowerShell equivalent:

$vault = "C:\Users\you\Documents\MyVault"
$dest  = Join-Path $vault ".obsidian\plugins\slide-autosummarize"
New-Item -ItemType Directory -Force -Path $dest | Out-Null
$base  = "https://github.com/BartolomeoZisa/obsidian-slide-autosummarize/releases/latest/download"
foreach ($f in "main.js","manifest.json","styles.css") {
  Invoke-WebRequest "$base/$f" -OutFile (Join-Path $dest $f)
}

Option D — Build from source

git clone https://github.com/BartolomeoZisa/obsidian-slide-autosummarize.git
cd obsidian-slide-autosummarize
npm install
npm run build        # typecheck + production bundle -> main.js

Then copy main.js, manifest.json, styles.css into <vault>/.obsidian/plugins/slide-autosummarize/ and enable the plugin.

Development: npm run dev watch-builds main.js on save.

Updating

• BRAT: updates automatically, or run Ctrl/Cmd+P → BRAT: Check for updates to all beta plugins.
• Manual: download the new main.js from the latest release and replace the old one (keep your data.json — that file holds your settings and API keys).
• Community directory (after acceptance): Settings → Community plugins → Check for updates.

---

Set up a provider

Open Settings → Slide Autosummarize and pick a provider.

Google Gemini (default)

1. Create a free API key at https://aistudio.google.com/apikey.
2. Provider: Google Gemini → paste the key.
3. Model: pick from the dropdown (gemini-3.5-flash-lite is the default) or choose Custom… and type any model id.

OpenAI

1. Provider: OpenAI → paste your sk-… key.
2. Pick a model (gpt-4o-mini, gpt-4o, …) or Custom… for another id.

Anthropic Claude

1. Provider: Anthropic Claude → paste your sk-ant-… key.
2. Pick a model or Custom….

Ollama (local, no API key)

1. Install Ollama and pull a vision model, e.g.:

   ollama pull qwen2.5vl
   

   Other options: llava, gemma3. Text-only models work too — turn off Send slide images to the LLM first.

2. Provider: Ollama (local), base URL http://localhost:11434/v1, model qwen2.5vl.

3. No key needed.

Any other OpenAI-compatible endpoint

Provider: Custom OpenAI-compatible → set the base URL (http://host:port/v1), an optional API key, and the model id. Works with vLLM, LM Studio, llama.cpp server, OpenRouter, Mistral, Together, …

---

Use it

1. Put a PDF in your vault.
2. Open it, then either:
   • press Ctrl/Cmd+P → Slide Autosummarize: Summarize active PDF, or
   • right-click the PDF in the file explorer → Summarize PDF slides, or
   • click the ✨ icon in the left ribbon while a PDF is open.
3. Watch the progress notice. When it finishes, the note opens automatically.

Output goes to Summaries/<pdf-name>.md by default, with figures in Attachments/slides/<pdf-name>/.

What the note contains

---
source-pdf: "[[lecture-03.pdf]]"
generated: 2026-10-17T09:12:44.000Z
provider: gemini
model: gemini-3.5-flash-lite
slides: 24
---

# lecture-03 — Slide Summary

> Auto-generated class notes from `[[lecture-03.pdf]]` (24 slides) with Google Gemini
> (`gemini-3.5-flash-lite`) on 2026-10-17.

## Slide 1

Notes for slide 1 …

## Slide 2

Notes, with math: the loss is $L(\theta) = -\sum_i \log p_\theta(y_i \mid x_i)$

$$
\nabla_\theta L = \mathbb{E}[\nabla_\theta \log p_\theta(y \mid x)]
$$

![[Attachments/slides/lecture-03/lecture-03-p02-fig1.png]]
*attention pattern across heads*

• Notes are written like class notes, not slide descriptions.
• Equations are real LaTeX ($…$ inline, $$…$$ display) and render in Reading view.
• Only relevant figures are embedded, cropped to the figure itself. Full slides are never embedded. Figures come from the LLM marking FIGURE [x1,y1,x2,y2] | caption (0–1000 coordinates) and the plugin cropping that region out of the page render.
• Later batches receive the notes written so far, so terminology, notation and definitions stay consistent across the whole deck.

---

Settings reference

Setting	What it does
LLM provider	Gemini / OpenAI / Anthropic / Ollama / Custom.
Model	Curated dropdown per provider, plus Custom… for any model id.
Output folder	Where the Markdown notes go (default Summaries).
Image folder	Where cropped figures (and optional full renders) are saved.
Send slide images to the LLM	Multimodal input. Required for figure detection. Turn off for text-only models / to save tokens.
Embed relevant figures in the note	Crop LLM-picked figures and embed them. Off = text-only note.
Keep full slide renders	Also save every rendered slide PNG (not embedded). Off by default.
Slides per LLM call	Batch size (default 3). Lower it if the model truncates output.
Render scale	Resolution of rendered slides (default 2.0). Higher = sharper crops, more tokens.
Temperature	On = omit the parameter and use the provider's own default.
Max tokens	Response cap per batch.
Auto-retry on overload	Retries per batch on 429/5xx/timeouts, exponential backoff honoring Retry-After (default 5, 0 = fail fast).
Extra instructions	Appended to every prompt, e.g. “Write in Italian, undergraduate level, focus on derivations.”
Auto-summarize new PDFs	Summarize every PDF added to the vault automatically.

---

Troubleshooting

“Could not start the PDF engine” — the pdf.js worker is bundled inside main.js and loaded from a Blob URL, so no network access is needed. If you see this, check the developer console (Ctrl+Shift+I) for the underlying error.

HTTP 503 / “model is experiencing high demand” — transient. The plugin retries automatically with backoff; raise Auto-retry on overload if your deck is long. Switch models if it persists.

Empty or truncated notes — lower Slides per LLM call (e.g. 1–2) and raise Max tokens.

No figures in the note — figures require Send slide images to the LLM and Embed relevant figures. Text-only models cannot detect figures.

Wrong crop regions — the model returned bad coordinates. Lower Slides per LLM call (fewer slides per request = better spatial attention) or use a stronger vision model.

Wrong model id — use Custom… and paste the exact id from the provider's docs.

---

How it works

1. src/pdf.ts — extracts text per page with pdf.js and renders each page to PNG. If rendering fails, the plugin continues text-only.
2. src/summarizer.ts — batches slides, sends text + images + prior notes to the provider, then assembles and writes the Markdown note.
3. src/figures.ts — parses FIGURE lines, converts 0–1000 boxes to pixels, crops the PNGs, and swaps the directives for embeds.
4. src/llm/ — one file per API behind a single LLMProvider interface.

Adding a provider

1. Create src/llm/my-provider.ts implementing LLMProvider (summarizeSlides, validateConfig, defaultModels).
2. Register it in src/llm/factory.ts (allProviders() + ProviderId).
3. Add its settings fields to SummarizerSettings (src/summarizer.ts) and the settings tab (src/settings.ts).

Privacy

The PDF is processed locally. Slide text and rendered page images are sent only to the provider you configure (or to your own Ollama server). API keys are stored in the plugin's data.json inside your vault — use a restricted key.

License

GPL-3.0-or-later — see LICENSE; copyright notice in NOTICE.

This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version. If you distribute a modified version, you must make the source available under the same license.

Third-party components keep their own licenses: pdfjs-dist (Apache-2.0) is bundled into main.js.