framedex

name: framedex description: "Build a portable knowledge base of your video (and eventually photo) archive across multiple SSDs or directly from an Apple Photos library. For each clip: GPS + reverse-geocoded place, speaker-diarized multi-lingual transcript with English translation, face detection + embeddings for later named-person queries, Claude/Gemma structured assessment (keep/review/cull rating + technical quality + lighting + time of day + dominant colors + audio quality + people count + keywords + notable timestamps), and prose scene description. Writes plain-text sidecars next to originals (or to a mirror tree for Photos libraries — never inside the .photoslibrary bundle) + persistent face DB. Photos-library mode also threads through albums, named persons, and keywords from the Photos database. Non-destructive, idempotent, resumable. Use whenever you want to: index videos, tag footage, organize a drive, build the video knowledge base, transcribe audio, describe clips, rate clips, find clips by location/lighting/person/keyword, generate folder summaries, identify duplicates or cull pile, index the Apple Photos library directly. Trigger phrases: 'index this drive', 'tag my videos', 'index my Photos library', 'index Apple Photos videos', 'what's on this SSD', 'rate these clips', 'find me clips of X', 'what should I cull', 'build the video knowledge base'."

framedex — Video Archive Knowledge Base

Cross-project, cross-drive. An entire video archive turned into a portable plain-text knowledge base + queryable face DB.

Per-clip pipeline

ffprobe — metadata
exiftool — GPS lat/lon/altitude (iPhone, DJI, drone all supported)
Nominatim — reverse-geocoded place name (rate-limited 1/sec, free, no key)
ffmpeg — 5 representative JPEG frames @ 1920px max
ffmpeg — audio extraction → WhisperX transcribe + diarization + alignment
WhisperX translate-mode — English translation for non-English clips
insightface (RetinaFace + ArcFace) — face detection + 512-dim embeddings on the same frames
Vision model (Claude Haiku/Sonnet via Max CLI / API, OR local Gemma via LM Studio) → structured YAML + prose description in one call
Write [filename].description.md sidecar + insert face rows into ~/.framedex/faces.db

Output schema

Each sidecar's YAML frontmatter:

file: IMG_4827.mov
path: /Volumes/SSD-2024/...
parent_folder: drone
duration_seconds: 12.3
resolution: 3840x2160
codec: hvc1
size_bytes: 245678912
creation_time: 2024-08-14T07:23:11Z
location:
  lat: 37.7456
  lon: -119.5936
  altitude_m: 1842.5
  place: "Yosemite Valley, Mariposa County, USA"
language_detected: es
speaker_count: 2
rating: keep                  # keep | review | cull
cull_reason: ""
technical:
  focus: sharp                # sharp | acceptable | soft
  exposure: strong            # strong | adequate | poor | clipped
  stability: smooth           # smooth | handheld | jittery
  motion_blur: clean          # clean | some | heavy
lighting: golden_hour
time_of_day: golden_hour
dominant_color_palette: "warm dusk: amber, ochre, dusty olive"
dominant_colors: [amber, ochre, olive, sky-blue]
audio_quality: clean_speech
people_count: 3               # vision model's estimate
keywords: [drone, landscape, construction, golden-hour, wide-shot, speech, workers]
notable_timestamp: ""         # MM:SS of peak moment if clip ≥ 30s
faces:                        # from insightface, separate from people_count
  - cluster_id: tmp_a3f78c    # temporary until fdx-faces labels it 'alex' / 'sam' / etc
    frame_time: 1.2
    bbox: [120, 80, 180, 240]
    detection_quality: high
face_count: 2
indexed_at: 2026-05-17T14:32:01

When indexed via fdx-photos, the sidecar also carries Photos-side fields:

original_filename: IMG_4827.MOV     # user-meaningful camera filename
photos_uuid: ABCD1234-EF56-7890-ABCD-1234567890AB
photos_persons: [Mom, Dad]          # from Photos' face recognition labels
photos_albums: [Yosemite 2024]      # Photos album membership
photos_keywords: [sunset]           # tags added in the Photos UI
photos_edited: true                 # only when Photos has edits on the clip

Body follows: ## Description (Scene/Subjects/Action/Mood/Shot type/Use cases prose), ## Transcript (with speaker labels if diarized), ## English translation (if applicable).

Three vision backends

Backend	Quality	Speed	Cost	Privacy
`cli` (default)	Claude Haiku/Sonnet via Max	~10-30s per clip	$0 (Max subscription)	Cloud (frames sent to Anthropic)
`api`	Claude Haiku/Sonnet via API	~2-3s per clip	~$0.002 (Haiku) / ~$0.008 (Sonnet) per clip	Cloud (frames sent to Anthropic)
`local`	Local model via LM Studio (Gemma 4, Qwen2-VL, etc.)	~3-90s depending on model	$0	Fully local

--vision-model haiku|sonnet picks Claude model for cli/api. --local-model NAME picks LM Studio model. The script auto-strips ANTHROPIC_API_KEY from claude -p subprocess env so CLI mode hits Max OAuth even if API key is set globally.

Face detection

Always on by default. ~/.framedex/faces.db is the single shared face database across all drives. Per-clip embeddings stored as 512 float32 vectors + bbox + detection score. Temporary cluster IDs (tmp_<hash>) get replaced with real names by the (not-yet-built) fdx-faces clustering tool — that tool will be a follow-up that doesn't require re-running the indexing pass, because all embeddings are captured here.

Skip with --no-faces if you don't want face data.

Companion scripts / aliases

Alias	Script	Purpose
`fdx`	`framedex.index_videos`	Main indexer for folder trees — videos and still photos (`--media images\|videos\|all`)
`fdx-photos`	`framedex.photos_indexer`	Index videos directly from an Apple Photos library (macOS, requires `[photos]` extra)
`fdx-summary`	`framedex.trip_summary`	Recursive folder summaries (`_folder-summary.md` in each ≥5-clip folder)
`fdx-master`	`framedex.master_index`	Drive-level `_INDEX.md` + `_INDEX.json`
`fdx-query`	`framedex.query`	Filter sidecars by metadata (rating, lighting, person, keyword, etc.)

Set up once

cd ~/.claude/skills/framedex

# Install Python deps (editable — changes take effect immediately). Media support
# is split into extras so a photo-only setup never pulls torch:
uv pip install -e '.[all]'        # video + photos + Apple Photos (everything)
# uv pip install -e '.[video]'    # video only (whisperx/torch)
# uv pip install -e '.[images]'   # still photos only (Pillow + pillow-heif)
# uv pip install -e '.[photos]'   # Apple Photos library (macOS, osxphotos)

# Verify system binaries + pre-download models
python3 scripts/setup.py

# HF token for pyannote diarization (one-time)
# Accept terms on https://huggingface.co/pyannote/speaker-diarization-3.1
# and https://huggingface.co/pyannote/segmentation-3.0 first
export HF_TOKEN=hf_...

# Only if using --backend api:
# export ANTHROPIC_API_KEY=sk-ant-...

# Commands are now on PATH after editable install:
#   fdx, fdx-photos, fdx-summary, fdx-master, fdx-query

Common run patterns

# Test 5 clips first — always
fdx /Volumes/SSD-2024 --max-files 5

# Full drive on default (Max CLI + Haiku)
fdx /Volumes/SSD-2024

# Higher accuracy via Max — slower, $0
fdx /Volumes/SSD-2024 --vision-model sonnet

# Local Gemma — fully offline
fdx /Volumes/SSD-2024 --backend local

# Skip movies (default cuts at 30 min)
fdx /Volumes/SSD-2024 --max-duration 30

# Re-process everything with new model
fdx /Volumes/SSD-2024 --force --vision-model sonnet

# After indexing: per-folder summaries
fdx-summary /Volumes/SSD-2024

# Drive overview
fdx-master /Volumes/SSD-2024

# Query examples
fdx-query /Volumes/SSD-2024 --rating keep --time-of-day golden_hour
fdx-query /Volumes/SSD-2024 --rating cull              # cull pile
fdx-query /Volumes/SSD-2024 --place-contains California --language es
fdx-query /Volumes/SSD-2024 --keyword drone --keyword landscape
fdx-query /Volumes/SSD-2024 --stability smooth --people-count 0
fdx-query /Volumes/SSD-2024 --rating keep --json | jq '.[] | .path'

Still photos (`fdx --media`)

fdx indexes still photos (RAW / JPEG / HEIC) through the same pipeline as video, minus the audio half, plus an EXIF camera block. One command handles mixed photo + video folders.

# Mixed drive — photos and clips both, one corpus
fdx /Volumes/SSD-photos --max-files 5
fdx /Volumes/SSD-photos

# Scope a run to one media type
fdx /Volumes/SSD-photos --media images          # stills only (no whisper stack)
fdx /Volumes/SSD-photos --media videos          # clips only

# Query just the photos
fdx-query /Volumes/SSD-photos --media images --place-contains Mara --keyword giraffe

Photo sidecars mirror the video schema with the audio/motion fields removed and a camera block added:

file: DSC_4827.RAF
media_type: image
dimensions: 7728x5152
camera: {make: FUJIFILM, model: X-T5, lens: "XF16-80mmF4", focal_length: "80.0 mm", aperture: 4.0, shutter: "1/1000", iso: 400}
location: {lat: -1.40, lon: 35.01, place: "Maasai Mara National Reserve, Narok County, Kenya"}
rating: keep                # keep | review | cull
technical: {focus: sharp, exposure: strong, composition: strong}
lighting: golden_hour
time_of_day: golden_hour
scene_type: wildlife        # wildlife | landscape | portrait | street | architecture | ...
keywords: [giraffe, waterhole, drinking, savanna]

RAW is read from the embedded full-res JPEG preview (no libraw). Requires the [images] extra (Pillow + pillow-heif); video indexing needs [video].

Apple Photos library

If the library is local on disk (no iCloud, or iCloud without Optimize Storage), this is one command — no flags, no permission prompts:

fdx-photos

The full menu:

# Quick sanity check before starting — how many videos are local vs iCloud-only
.venv/bin/python scripts/diagnose_photos.py

# Smoke-test on 5 clips first
fdx-photos --max-files 5

# Full library (default: ~/Pictures/Photos Library.photoslibrary → ~/framedex-photos/)
fdx-photos

# Filter by Photos-side metadata (repeatable, OR-combined within a flag)
fdx-photos --album "Yosemite 2024" --since 2024-01-01
fdx-photos --person "Mom" --keyword sunset

# Custom mirror output (must be OUTSIDE the .photoslibrary bundle)
fdx-photos --output ~/Documents/photos-kb

# Re-process a single problem clip by UUID
fdx-photos --uuid ABCD1234-EF56-7890-ABCD-1234567890AB --force

# Only if Optimize Mac Storage is on AND you want per-clip iCloud downloads
# (needs Photos permission for the terminal in System Settings → Privacy & Security)
fdx-photos --download

# After indexing, fdx-query / fdx-summary / fdx-master work on the mirror
fdx-query ~/framedex-photos --rating keep --person "Mom"
fdx-master ~/framedex-photos

Optional folder context

Drop .video-context.md at the root of any scan target with a paragraph describing what's on that drive ("construction site, 2023-2026", "family travel, 2024", etc). The vision prompt prepends it for context-aware descriptions.

Privacy

Component	Local or cloud?
ffmpeg / exiftool / Whisper / pyannote / insightface	Local
Nominatim reverse geocoding	Sends lat/lon (not video). Skip with `--no-geocode`.
Vision (`--backend cli`/`api`)	Frames sent to Anthropic. By default not used for training.
Vision (`--backend local`)	Fully local, fully offline.
Face DB (`~/.framedex/faces.db`)	Local only, never uploaded. Back up the file manually if you care.

Multiple SSDs

Run on each drive separately. Sidecars travel with the data; the face DB is centralized at ~/.framedex/faces.db so cross-drive person queries work.

Known limitations (v1)

Frame sampling is evenly-spaced, not scene-detected (future: ffmpeg select=gt(scene,0.4))
pyannote diarization degrades on heavy ambient noise (wind, music, crowd)
WhisperX runs on CPU on Apple Silicon (CTranslate2 doesn't have M-series GPU acceleration yet; 64GB CPU is still plenty)
fdx-faces (clustering + labeling tool) not built yet — face embeddings are captured but cluster IDs are temporary hashes until that tool ships
RAW image format support not yet (videos only; photos are coming)
fdx-photos --download relies on PhotoKit, which requires the parent terminal to have "Photos" access in macOS Privacy & Security. Unsigned Python (the default .venv install) often can't trigger the OS prompt, so the terminal may never appear in that panel. The supported workaround is to turn off "Optimize Mac Storage" in Photos → Settings → iCloud, which makes all originals local and removes the need for --download entirely.

File layout

~/.claude/skills/framedex/
├── SKILL.md                       # this file
├── README.md
├── pyproject.toml                 # deps, ruff/mypy config, entry points
├── .pre-commit-config.yaml        # pre-commit hooks
├── .github/workflows/ci.yml       # CI (ruff + mypy)
├── scripts/
│   ├── setup.py                   # system binaries + model pre-download
│   └── diagnose_photos.py         # one-shot Photos-library state report
└── src/framedex/
    ├── __init__.py                # package init, version from pyproject.toml
    ├── index_videos.py            # main worker (fdx) — also exports
    │                              # ProcessOptions/ProcessContext/process_one_video
    │                              # so fdx-photos can reuse the per-clip pipeline
    ├── photos.py                  # Apple Photos library adapter (osxphotos)
    ├── photos_indexer.py          # Photos library indexer (fdx-photos)
    ├── face_db.py                 # face detection + SQLite face DB module
    ├── trip_summary.py            # recursive folder summaries (fdx-summary)
    ├── master_index.py            # drive-level KB (fdx-master)
    └── query.py                   # filter sidecars (fdx-query)