🎓 yt-summariser

Turn 2-hour YouTube lectures into focused Markdown research notes — free, local, no API keys, built to last.

yt-summariser downloads a video's transcript, scores every segment across four independent signals, detects topic boundaries, and produces a structured Markdown report with a TL;DR, chapter summaries, timestamped highlights, cross-chapter insights, and actionable takeaways. Unlike cloud summarisers it requires no API key, runs entirely on your machine, grounds every claim in the actual transcript, and queries Wikipedia on-demand only for terms the speaker left unexplained. ok.

Designed for daily use over 2–3 years. See the Maintenance Guide for the exact update checklist.

✨ What you get (per video)

Output	Description
`<video_id>.md`	Structured Markdown: TL;DR · chapters · highlights · insights · takeaways
Highlight Segments	Timestamped table of top moments with source badges (📄 / 🎙️ / 🖼️)
Chapter Detection	Automatic topic boundaries with YouTube deep-links
Wikipedia Context	On-demand definitions for concepts the speaker left unexplained
`<video_id>_highlights.mp4`	Clipped highlight reel (skip with `--no-video`)

🏗️ How it works

Transcript tiers (best available wins, cheapest tried first)

Tier	Source	Accuracy	Cost
0	YouTube manual captions (creator-uploaded)	~98 %	Free, instant
1	YouTube auto-generated captions	~85–95 %	Free, instant
2	Whisper ASR local fallback	~80–90 %	Free, ~1–3 min/hour
3	Frame OCR via easyocr (weak segments only)	~88–92 % on slides	Free, ~2 s/frame
4	moondream2 visual caption (when OCR finds nothing)	~70–80 %	Free, needs `transformers`

A segment is weak (triggers Tier 3/4) when it has fewer than 5 words, less than 0.5 words per second, or text like [Music] / [Applause].

Synthesis stack

Stage	Tool	Fallback
Scoring	TF-IDF · MiniLM cosine · word density · audio energy	TF-IDF + density
Chapter detection	Cosine similarity shifts on MiniLM embeddings	Single chapter
Chapter titles	BART `facebook/bart-large-cnn` `max_length=10`	TF-IDF top-4 terms
LLM synthesis	Mistral-7B-Instruct Q4_K_M via `llama-cpp-python`	Sumy LSA + KeyBERT
Wikipedia	REST API, ~3–8 calls/video, `[[LOOKUP: term]]` only	Silently skipped

Wikipedia is a dictionary, not an editor. It is consulted only when Mistral emits a [[LOOKUP: term]] marker, meaning the transcript itself left a concept unexplained. It never adds missing topics or criticises the speaker.

🚀 Quick Start

# 1. Install core dependencies (fast, always required)
pip install yt-dlp scikit-learn

# 2. Install recommended dependencies (one-time, models downloaded on first run)
pip install -r requirements.txt

# 3. Markdown report only (fastest, no models needed)
python -m summarise "https://youtube.com/watch?v=VIDEO_ID" \
    --no-llm --no-enrich --no-video

# 4. Full pipeline with highlight video (default)
python -m summarise "https://youtube.com/watch?v=VIDEO_ID"

# 5. Multiple videos in parallel
python -m summarise \
    "https://youtube.com/watch?v=VIDEO_A" \
    "https://youtube.com/watch?v=VIDEO_B" \
    --workers 2

# 6. Entire playlist
python -m summarise "https://youtube.com/playlist?list=PLxxxxxxxxxxxxxxxx"

# 7. Custom output length (12 minutes of highlights)
python -m summarise "https://youtube.com/watch?v=VIDEO_ID" \
    --target-mins 12

# 8. Save to a specific folder
python -m summarise "https://youtube.com/watch?v=VIDEO_ID" \
    --output-dir ~/research/ml-lectures

Reports are written to ./output/<video_id>.md by default. Re-running the same URL is a no-op: if <video_id>.md already exists the video is skipped.

⚙️ All Options

Flag	Type	Default	Description
`urls`	positional (1+)	—	YouTube video or playlist URLs
`--target-pct`	float	`0.10`	Highlight duration as fraction of video length
`--target-mins`	float	`None`	Override `--target-pct` with exact minutes
`--no-video`	flag	off	Skip ffmpeg clipping, produce `.md` only
`--no-llm`	flag	off	Skip Mistral-7B; use Sumy/KeyBERT extractive fallback
`--no-enrich`	flag	off	Skip frame OCR and visual captioning
`--no-captions`	flag	off	Run OCR but skip moondream2 visual captioning
`--output-dir`	str	`output`	Directory for `.md` files and `.mp4` clips
`--whisper-model`	str	`base`	Whisper model: `tiny` `base` `small` `medium` `large`
`--workers`	int	`1`	Parallel threads for batch processing (max 15 videos)
`--loglevel`	str	`warning`	Verbosity: `debug` `info` `warning`

📄 Example Output

Snippet from a hypothetical video on Attention Is All You Need (Vaswani et al.):

# Attention Is All You Need — Explained

| | |
|---|---|
| **URL** | [https://youtu.be/iDulhoQ2pro](https://youtu.be/iDulhoQ2pro) |
| **Duration** | 1h 26m 12s |
| **Processed** | 2024-06-15T14:32:07 |
| **Transcript** | 📄 auto — 1847 segments total |
| **Highlights** | 9 clips · 8m 43s |

---

## 🎯 TL;DR

The Transformer replaces recurrence with self-attention [00:04:11], achieving
state-of-the-art translation while training fully in parallel.  Multi-head
attention lets the model attend to different representation subspaces jointly
[00:09:33]. Positional encodings using sine/cosine substitute for the
sequential inductive bias previously provided by RNNs [00:21:58].

---

## 🏷️ Key Concepts

`self-attention` `multi-head attention` `positional encoding`
`scaled dot-product` `encoder-decoder` `layer normalisation` `BLEU score`

---

## 📖 Chapters

### Chapter 1 · Motivation & RNN Limitations · [[00:00:00]](https://youtu.be/iDulhoQ2pro?t=0)
The speaker explains why LSTMs struggle with long-range dependencies and
prevent efficient parallelisation during training [00:02:44].

### Chapter 2 · Scaled Dot-Product Attention · [[00:07:18]](https://youtu.be/iDulhoQ2pro?t=438)
Attention is a weighted sum of values where weights are computed from query-key
compatibility [00:08:55]. The 1/√d_k scaling prevents gradients from vanishing
in high-dimensional spaces [00:11:02].

---

## ✂️ Highlight Segments

| # | Time | Score | Source | Excerpt |
|---|---|---|---|---|
| 1 | [00:08:55](https://youtu.be/iDulhoQ2pro?t=535) | 0.91 | 📄 auto | Attention is a weighted sum of the values where the weight is a compatibility function of the query with the key... |
| 2 | [00:19:47](https://youtu.be/iDulhoQ2pro?t=1187) | 0.87 | 📄 auto | Instead of a single attention function we found it beneficial to project queries, keys and values h times... |
| 3 | [00:43:12](https://youtu.be/iDulhoQ2pro?t=2592) | 0.79 | 🖼️ ocr | [ON SCREEN: BLEU Score Table — Transformer (big): 41.0 EN-DE, 43.9 EN-FR] |

---

## 💡 Actionable Takeaways

1. Apply 1/√d_k scaling before softmax to avoid gradient saturation.
2. Start with h=8 attention heads; ablations show degradation below h=4.
3. Add positional encodings before the first encoder layer.

---

## 📚 References

**self-attention** — a mechanism that relates different positions of a single
sequence to compute a representation of that sequence.
· [Wikipedia](https://en.wikipedia.org/wiki/Attention_(machine_learning))

🧩 Architecture

YouTube URL
    │
    ├─ Tier 0/1: yt-dlp captions ──────────► Segments        [transcript.py]
    ├─ Tier 2:   Whisper (fallback) ─────────►
    │
    ├─ Tier 3/4: Frame OCR + Caption ────────► EnrichedSegs   [enricher.py]
    │            (weak segments only)
    │            easyocr · moondream2
    │
    ├─ Scoring ───────────────────────────────────────────    [scorer.py]
    │   TF-IDF · MiniLM cosine · word density · ffmpeg energy
    │
    ├─ Chapter Detection ─────────────────────────────────    [segmenter.py]
    │   cosine similarity shifts · BART titles
    │
    ├─ Greedy Selection ──────────────────────────────────    [selector.py]
    │   fill target_duration · merge adjacent clips
    │   └─► SelectedSegments
    │
    ├─ Mistral-7B Synthesis ──────────────────────────────    [synthesiser.py]
    │   llama-cpp-python · grounded · timestamp-cited
    │   Fallback: Sumy LSA + KeyBERT
    │   Emits [[LOOKUP: X]] markers for unknown terms
    │
    ├─ Wikipedia ─────────────────────────────────────────    [references.py]
    │   on-demand REST API · ~3–8 calls/video
    │   resolves [[LOOKUP: X]] markers ONLY
    │
    ├─ ffmpeg Clipping ───────────────────────────────────    [clipper.py]
    │   yt-dlp stream URL · -c copy · concat demuxer
    │   └─► <video_id>_highlights.mp4
    │
    └─ Markdown Reporter ─────────────────────────────────    [reporter.py]
        deep-links · source badges · injected references
        └─► <video_id>.md

📦 Dependencies

Python packages

Package	Purpose	Required?
`yt-dlp`	Subtitle download, stream URL, playlist expansion, audio extraction	✅ Required
`scikit-learn`	TF-IDF scoring, chapter title fallback, concept extraction	✅ Required
`sentence-transformers`	MiniLM embeddings: semantic scoring + chapter detection	⭐ Recommended
`transformers`	BART chapter titles, moondream2 visual captioning	🔵 Optional
`torch`	Required by `transformers` and `sentence-transformers`	🔵 Optional
`accelerate`	Speeds up `transformers` inference	🔵 Optional
`llama-cpp-python`	Mistral-7B-Instruct Q4_K_M local inference	⭐ Recommended
`huggingface-hub`	Auto-downloads GGUF model on first use	🔵 Optional
`openai-whisper`	Local ASR when no YouTube captions exist	🔵 Optional
`easyocr`	OCR on video frames for weak segments	🔵 Optional
`sumy`	LSA extractive summariser (LLM fallback)	🔵 Optional
`keybert`	Key concept extraction (LLM fallback)	🔵 Optional

System tools (must be on `PATH`)

Tool	Purpose	Install
`ffmpeg`	Audio analysis, frame extraction, clipping, concat	`brew install ffmpeg` / `apt install ffmpeg`
`yt-dlp` binary	Subtitle + video download	`pip install yt-dlp` or yt-dlp releases

💡 Design Principles

1. Transcript is the only authority The LLM cannot add information not present in the transcript. Every factual claim must end with a timestamp citation ([00:14:22]). The prompt explicitly forbids hallucination. The transcript is capped at 6 000 chars, prioritising the highest-scored segments.

2. Wikipedia is a dictionary Wikipedia is consulted on-demand and only when the LLM emits [[LOOKUP: term]] — meaning the transcript itself left a concept unexplained. It never adds topics the speaker did not discuss, never corrects the speaker, never signals gaps. Disambiguation pages and no-extract articles are silently discarded.

3. Graceful degradation Every heavy dependency is optional and lazily imported. Missing packages emit a warnings.warn and the pipeline continues with the next-best strategy. Even with only yt-dlp + scikit-learn, the pipeline always produces a report.

🔧 Maintenance Guide — 2–3 Years

This section documents exactly what will need attention over the lifetime of this tool, in order of likelihood. Nothing here requires code changes unless explicitly noted.

🔴 HIGH — Will definitely need action

1. `yt-dlp` — update every 2–4 months

Why it breaks: YouTube changes its internal API, webpage structure, and signature algorithms regularly. yt-dlp is the only component that directly talks to YouTube. When it breaks, subtitle downloads and stream URL resolution both fail.

Symptom:

ERROR: [youtube] VIDEO_ID: Sign in to confirm you're not a bot
ERROR: Unable to extract video data

Fix (30 seconds):

pip install -U yt-dlp

Or if using the binary directly:

yt-dlp -U

How to check without running a video:

yt-dlp --version
yt-dlp "https://youtu.be/dQw4w9WgXcQ" --get-title --no-warnings

In the desktop app: Settings → [Check for yt-dlp update] button does this automatically.

Files affected: utils.py, transcript.py, clipper.py Code change needed: Almost never — yt-dlp's CLI interface is stable. Only needed if yt-dlp removes or renames a flag used in these files.

2. YouTube subtitle format — watch for VTT changes

Why it breaks: The VTT parser in utils.py:_parse_vtt() expects a specific cue block structure. YouTube has changed this format before (adding extra metadata lines, changing timing precision, adding new tag types).

Symptom: Reports are generated but all segments have empty text, or TF-IDF scores are uniformly 0.0.

Fix: Check the raw .vtt file:

yt-dlp --write-auto-sub --sub-lang en --skip-download \
    --output "/tmp/test_%(id)s" "https://youtu.be/VIDEO_ID"
cat /tmp/test_VIDEO_ID.en.vtt | head -40

Compare the structure against what _parse_vtt() in utils.py expects. The function is self-contained and well-commented — the fix is usually adding a new skip condition for a new header line or a new tag pattern to _VTT_TAG_RE.

Files affected: utils.py (function _parse_vtt, lines ~160–220) Code change: Minor regex or skip-line update, 5–10 lines.

🟡 MEDIUM — Will likely need action within 2 years

3. `llama-cpp-python` — update when Python or LLVM version changes

Why it breaks: llama-cpp-python compiles C++ extensions at install time. A new Python version (e.g. 3.13 → 3.14) or a macOS Xcode update can break the build.

Symptom:

ImportError: llama_cpp.so: cannot open shared object file

or a build failure during pip install llama-cpp-python.

Fix:

pip install -U llama-cpp-python --force-reinstall --no-cache-dir

On Apple Silicon (M1/M2/M3), use the Metal-accelerated build:

CMAKE_ARGS="-DLLAMA_METAL=on" pip install llama-cpp-python \
    --force-reinstall --no-cache-dir

Files affected: synthesiser.py (function _get_llm) Code change needed: Never — the Llama() constructor API has been stable since v0.2. Only change if llama-cpp-python deprecates a parameter (check their CHANGELOG).

4. Mistral-7B model — upgrade to better quantisation or newer version

Why to upgrade: Better models (Mistral v0.3, Mistral Nemo 12B, Llama 3.1 8B) provide higher-quality synthesis with the same or less RAM. Quantisation improvements (Q5_K_M, Q6_K) offer better quality at similar file size.

This requires zero code changes. The GGUF format is stable. Just swap the model file.

How to upgrade:

Download a new .gguf file from HuggingFace:
- TheBloke's GGUF collection
- bartowski's newer quantisations
Point to it:
- Desktop app: Settings → LLM Model Path → Browse
- CLI: set llm_model_path in synthesiser._get_llm() or pass via the model_path argument to synthesise()

Recommended upgrade candidates (2025–2027):

Model	Size	RAM needed	Quality vs current
`mistral-7b-instruct-v0.3.Q4_K_M.gguf`	4.1 GB	6 GB	Marginally better
`Meta-Llama-3.1-8B-Instruct.Q4_K_M.gguf`	4.9 GB	7 GB	Noticeably better
`Mistral-Nemo-Instruct-2407.Q4_K_M.gguf`	7.7 GB	10 GB	Significantly better
`Phi-3.5-mini-instruct.Q4_K_M.gguf`	2.2 GB	4 GB	Smaller, faster, similar

Files affected: synthesiser.py (function _get_llm, lines ~90–120) Code change needed: Only if the new model uses a different prompt format than Mistral instruct ([INST]...[/INST]). Check the model card on HuggingFace. If it uses ChatML format instead, update _build_prompt() accordingly.

5. `sentence-transformers` + `all-MiniLM-L6-v2` — model may be superseded

Why to upgrade: The all-MiniLM-L6-v2 model (80 MB) is from 2021. Better small embedding models now exist.

Recommended replacement (drop-in):

# Replace all-MiniLM-L6-v2 with a 2024 model — same API, better quality

Change this one line in both scorer.py and segmenter.py:

# Before
model = SentenceTransformer("all-MiniLM-L6-v2")

# After (2024 — better quality, same size)
model = SentenceTransformer("sentence-transformers/all-MiniLM-L12-v2")

# Or (2024 — slightly larger but notably better)
model = SentenceTransformer("BAAI/bge-small-en-v1.5")

Files affected: scorer.py (function _score_semantic, ~1 line), segmenter.py (function _encode_texts, ~1 line)

🟢 LOW — Unlikely to need action, but document anyway

6. `easyocr` — update if OCR quality degrades on new slide styles

Symptom: OCR returns garbage or empty strings on slides that look visually clear.

Fix:

pip install -U easyocr

EasyOCR occasionally releases improved models for English. The API (reader.readtext()) has been stable since v1.0.

Files affected: enricher.py (function _run_ocr)

7. `transformers` + moondream2 — update if captioning model is deprecated

The vikhyatk/moondream2 model on HuggingFace is actively maintained as of 2024. If it is deprecated or removed:

Find a replacement on HuggingFace with pipeline("image-to-text"):
- nlpconnect/vit-gpt2-image-captioning (smaller, faster)
- Salesforce/blip-image-captioning-base (better quality)
Change one line in enricher.py:

# Before
_caption_pipeline = hf_pipeline("image-to-text", model="vikhyatk/moondream2")

# After
_caption_pipeline = hf_pipeline("image-to-text",
                                  model="Salesforce/blip-image-captioning-base")

Files affected: enricher.py (function _run_caption, ~1 line)

8. Wikipedia REST API — extremely stable, no action expected

The Wikipedia REST API (/api/rest_v1/page/summary/{title}) has been stable since 2015. Wikipedia's terms of service for automated access require only a descriptive User-Agent header, which is already set in references.py.

The only scenario requiring a fix: Wikimedia changes the JSON response schema. Symptom: WikiReference.summary is always empty despite API returning 200.

Fix: Inspect the raw response:

import urllib.request, json
url = "https://en.wikipedia.org/api/rest_v1/page/summary/Attention_(machine_learning)"
with urllib.request.urlopen(url) as r:
    print(json.dumps(json.loads(r.read()), indent=2))

Then update _fetch_wiki() in references.py to use the new field names.

9. `BART` (`facebook/bart-large-cnn`) — update if HuggingFace deprecates

This model has been on HuggingFace since 2020 and is extremely widely used. Deprecation is unlikely. If it happens, replace with:

# In segmenter.py, function _generate_chapter_title:
_bart_pipeline = hf_pipeline("summarization",
                               model="facebook/bart-large-xsum")  # smaller, punchier
# or
_bart_pipeline = hf_pipeline("summarization",
                               model="sshleifer/distilbart-cnn-12-6")  # 50% faster

┌─────────────────────────────────────────────────┐
│  🎓 yt-summariser                    [_][□][X]  │
├──────────────┬──────────────────────────────────┤
│              │                                  │
│  📚 History  │   ┌─ Add Video ──────────────┐   │
│  ─────────── │   │ URL: [________________]  │   │
│  Today       │   │ [▶ Analyse]  [⚙ Options] │   │
│  • Video 1   │   └──────────────────────────┘   │
│  • Video 2   │                                  │
│              │   ┌─ Progress ───────────────┐   │
│  Yesterday   │   │ ████████░░░ Stage 6/10   │   │
│  • Video 3   │   │ Detecting chapters...    │   │
│              │   └──────────────────────────┘   │
│  This week   │                                  │
│  • Video 4   │   ┌─ Output ─────────────────┐   │
│  • Video 5   │   │ [📄 Open Report]         │   │
│              │   │ [🎬 Open Video]          │   │
│  🔍 Search   │   │ [📁 Open Folder]         │   │
│  [_______]   │   └──────────────────────────┘   │
│              │                                  │
│  🏷️ Topics   │   ┌─ Preview ────────────────┐   │
│  ml (12)     │   │ # Attention Is All You   │   │
│  python (8)  │   │ ## TL;DR                 │   │
│  systems (5) │   │ The paper introduces...  │   │
│              │   │ ## Chapters              │   │
│              │   │ ### 1. Motivation [▶]    │   │
└──────────────┴──────────────────────────────────┘

📅 Recommended Update Schedule

Frequency	Task
Every 2–4 months	`pip install -U yt-dlp`
Every 6 months	`pip install -U yt-dlp scikit-learn sentence-transformers easyocr`
Annually	Review model quality — consider swapping to a better GGUF
On Python upgrade	Reinstall `llama-cpp-python` from scratch
When a stage breaks	Check the symptom table above, fix the identified file

🧪 Quick Health Check

Run this after any update to confirm the pipeline still works end-to-end:

# Uses a short video (Rick Astley — 3.5 min) to test all stages quickly
python -m summarise "https://youtu.be/dQw4w9WgXcQ" \
    --no-video \
    --no-enrich \
    --output-dir /tmp/yt_health_check

# Expected: /tmp/yt_health_check/dQw4w9WgXcQ.md created with content
cat /tmp/yt_health_check/dQw4w9WgXcQ.md | head -20

If the file is created with a non-empty TL;DR and at least one highlight segment, the pipeline is healthy.

⚠️ Known Limitations

Limitation	Detail	Workaround
English only	`--sub-lang en` hardcoded; TF-IDF stop words are English	Edit `utils.py` line ~170 to change language code
Mistral-7B RAM	Needs ~6 GB RAM minimum	Use `--no-llm` flag or swap to `Phi-3.5-mini` (4 GB)
Frame enrichment speed	~2 s/frame on CPU; many weak segments add minutes	Use `--no-enrich` to skip
Clip boundary drift	`-c copy` snaps to nearest keyframe, up to ~2 s drift	Unavoidable without re-encoding; use timestamps in `.md` instead
Max 15 videos/run	Hard cap in `summarise.py` to prevent runaway downloads	Run the tool multiple times for large playlists
No GPU acceleration	All models run on CPU by default	For llama-cpp: `CMAKE_ARGS="-DLLAMA_METAL=on" pip install llama-cpp-python` (macOS)

Name		Name	Last commit message	Last commit date
__pycache__		__pycache__
app		app
summarise		summarise
.DS_Store		.DS_Store
.gitignore		.gitignore
README.md		README.md
run.py		run.py

Navigation Menu

Repository navigation

About

Repository files navigation

🎓 yt-summariser

Table of Contents

✨ What you get (per video)

🏗️ How it works

Transcript tiers (best available wins, cheapest tried first)

Synthesis stack

🚀 Quick Start

⚙️ All Options

📄 Example Output

🧩 Architecture

📦 Dependencies

Python packages

System tools (must be on `PATH`)

💡 Design Principles

🔧 Maintenance Guide — 2–3 Years

🔴 HIGH — Will definitely need action

1. `yt-dlp` — update every 2–4 months

2. YouTube subtitle format — watch for VTT changes

🟡 MEDIUM — Will likely need action within 2 years

3. `llama-cpp-python` — update when Python or LLVM version changes

4. Mistral-7B model — upgrade to better quantisation or newer version

5. `sentence-transformers` + `all-MiniLM-L6-v2` — model may be superseded

🟢 LOW — Unlikely to need action, but document anyway

6. `easyocr` — update if OCR quality degrades on new slide styles

7. `transformers` + moondream2 — update if captioning model is deprecated

8. Wikipedia REST API — extremely stable, no action expected

9. `BART` (`facebook/bart-large-cnn`) — update if HuggingFace deprecates

📅 Recommended Update Schedule

🧪 Quick Health Check

⚠️ Known Limitations

About

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

🎓 yt-summariser

Table of Contents

✨ What you get (per video)

🏗️ How it works

Transcript tiers (best available wins, cheapest tried first)

Synthesis stack

🚀 Quick Start

⚙️ All Options

📄 Example Output

🧩 Architecture

📦 Dependencies

Python packages

System tools (must be on PATH)

💡 Design Principles

🔧 Maintenance Guide — 2–3 Years

🔴 HIGH — Will definitely need action

1. yt-dlp — update every 2–4 months

2. YouTube subtitle format — watch for VTT changes

🟡 MEDIUM — Will likely need action within 2 years

3. llama-cpp-python — update when Python or LLVM version changes

4. Mistral-7B model — upgrade to better quantisation or newer version

5. sentence-transformers + all-MiniLM-L6-v2 — model may be superseded

🟢 LOW — Unlikely to need action, but document anyway

6. easyocr — update if OCR quality degrades on new slide styles

7. transformers + moondream2 — update if captioning model is deprecated

8. Wikipedia REST API — extremely stable, no action expected

9. BART (facebook/bart-large-cnn) — update if HuggingFace deprecates

📅 Recommended Update Schedule

🧪 Quick Health Check

⚠️ Known Limitations

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

System tools (must be on `PATH`)

1. `yt-dlp` — update every 2–4 months

3. `llama-cpp-python` — update when Python or LLVM version changes

5. `sentence-transformers` + `all-MiniLM-L6-v2` — model may be superseded

6. `easyocr` — update if OCR quality degrades on new slide styles

7. `transformers` + moondream2 — update if captioning model is deprecated

9. `BART` (`facebook/bart-large-cnn`) — update if HuggingFace deprecates