Snag — AI Summarize

Turn a YouTube video — or a whole playlist / channel — into a clean, structured Markdown article. The Summarize page (the document-with-sparkle icon in the top bar) runs a three-step pipeline:

1. Transcript — fetch the subtitles/captions with youtube-transcript-api (a single video, or every video in a playlist / channel, combined into one document).
2. Review — read the raw transcript in the app, or download it as .md.
3. Summarize — send it to an OpenAI-compatible endpoint, which returns a full article following a system prompt you can edit.

Everything runs server-side as plain subprocesses. There is no hardcoded endpoint, port, or API key anywhere in the code — all of it comes from the settings you save.

Only local AI is encouraged. You can still point the endpoint at an OpenAI-compatible cloud URL if you want, but the feature is designed around a local, self-hosted model server.

First run

After you create your account, the home page shows a one-time, optional popup asking for your AI endpoint, model, and (optionally) an API key. It is skippable — click set up later and the app is fully usable without AI. The decision is remembered, so it never appears again; you can configure AI any time from Settings → AI / summarize.

How it works

Two plain Python scripts live in the yt_summarize/ folder and are invoked by the backend with subprocess — they are not imported into snag.py:

Script	Job
yt_summarize/yt-transcribe.py	YouTube URL → transcript.md (via youtube-transcript-api)
yt_summarize/summarize.py	transcript + system prompt → AI article (via a POST to /chat/completions)

summarize.py always receives --endpoint (required), --system-prompt (the prompt text, passed in from the database), and an optional --api-key. When a key is set it is sent as an Authorization: Bearer <key> header; when it's blank the request goes out with no auth — which is what a local server expects.

System prompt

The system prompt is stored in the database, not in a file. On first start the app seeds it with a sensible default (title, summary, key points, a sectioned breakdown, and a takeaway). You can edit it any time from Settings → AI / summarize → system prompt to steer the article's structure and tone.

Playlists and channels

The transcript step accepts a type — auto-detect, video, playlist, or channel / user. For a collection the app enumerates the videos with yt-dlp -J --flat-playlist, fetches each transcript, and combines them into a single Markdown document (one ## section per video, with a header showing the source, how many videos were included, and how many were skipped for lacking captions).

• How many videos is up to you: leave the videos field blank to fetch all, or enter a number to fetch only the first N. There is no per-video picker yet.
• Fetches run in parallel (4 workers) so large channels are much faster than a one-at-a-time loop. Videos with no captions in the requested languages — or that time out / fail — are skipped and counted, and the rest still come back, so one bad video never blocks the collection.

Configuration

Settings are saved from Settings → AI / summarize (or from the one-time popup on the home page) and stored in the SQLite database.

Setting	Notes
Endpoint	Any OpenAI-compatible base URL, e.g. http://localhost:8080/v1. Required — there is no default.
Model	The model id. Leave blank to pick from the loaded model list on the Summarize page.
API key	Optional. Blank = no auth (local). Set it for a cloud endpoint that needs a Bearer token.
Temperature	Sampling temperature (default 0.4).
Timeout	Seconds to wait for the summarize call (default 300).
Transcript languages	Priority-ordered language codes (default en).
Include timestamps	Default for the transcript's timestamped section (on by default).
System prompt	The instructions given to the model. Seeded with a default on first run; fully editable.

Pointing at a local model server

Any server that speaks the OpenAI chat-completions API works, for example:

• Ollama — http://localhost:11434/v1
• llama.cpp (llama-server) — http://localhost:8080/v1
• vLLM — http://localhost:8000/v1

Set the endpoint and model, then click load models to confirm the server is reachable and to see what's available.

Using the page

1. Open the Summarize page (document-with-sparkle icon).
2. Paste a YouTube URL (video, playlist, or channel) and set the options: type, how many videos, languages, and whether to include timestamps.
3. Click get transcript. Review it, or click download transcript .md.
4. Pick a model and click summarize.
5. The article renders in the app. Use copy markdown or download .md to take it with you.

Your work survives a refresh. The URL, transcript, article, and options are kept in your browser's localStorage, so reloading the page (or navigating away and back) does not lose anything. Click clear to start over.

Transcripts and summaries are in-app only — they are not written to the downloads folder and do not appear on the Downloads page.

API

Method	Path	Purpose
GET	/api/ai/settings	Read the saved AI settings (plus has_endpoint and setup_done).
POST	/api/ai/settings	Save AI settings. Body: the setting keys.
POST	/api/ai/setup-done	Mark the one-time AI setup prompt as handled.
GET	/api/ai/models	List models from the configured endpoint.
POST	/api/ai/transcript	Fetch a transcript. Body: {url, type?, count?, languages?, timestamps?}.
POST	/api/ai/summarize	Summarize a transcript. Body: {transcript, model?, temperature?, system_prompt?}.

Troubleshooting

• "No endpoint configured" — set the endpoint in Settings first.
• "No model selected" — pick a model on the page or set ai_model.
• Model list is empty — the endpoint is unreachable or not an OpenAI-compatible /models endpoint. Check the URL and that the server is up.
• Timeout — large transcripts (especially whole channels) on a slow local model can take a while; raise the timeout in Settings.
• Transcript not found — the video has no captions in the requested languages. Try adding more language codes to the transcript languages.
• Collection returns "No videos found" — the URL didn't resolve to a playlist/channel, or it's empty. Check the link.
• Many videos skipped — those videos have no captions in the requested languages. Add more language codes or accept the partial result.