Local-first AI image memory. Point it at a folder of images, and it builds a searchable visual knowledge base on your machine.
Instead of relying on filenames or manual organization, Eyra uses vision models to understand each image, generate tags and captions, and create embeddings for semantic search. eyra search "cat on a roof" finds the image, even if the file is called IMG_4892.jpg.
- Semantic search: find images by describing what you're looking for
- Auto-tagging: vision models generate tags and captions (BLIP or Florence-2)
- Hybrid search: combines vector similarity and keyword matching
- Visual clustering: auto-group similar images using K-Means on CLIP embeddings
- OCR: extract text from images (Apple Vision or Tesseract)
- Chat UI: conversational search with natural language summaries
- Timeline: browse images by date from EXIF data
- Obsidian export: export to Markdown notes with YAML frontmatter
- Folder indexing: point it at any folder, it handles the rest
- File watching: automatically indexes new images as they appear
- Web UI: dark-mode interface with Browse, Clusters, Chat, and Timeline tabs
- 100% local: no cloud, no accounts, no data leaves your machine
- Fast: optimized for Apple Silicon (M-series Macs)
git clone https://github.com/4thul0krishn4/Eyra.git
cd Eyra
./setup.shThen index your images:
source venv/bin/activate
eyra index ~/Pictures --auto-captionSearch:
eyra search "sunset at the beach"Start the web UI:
eyra serveOpen http://localhost:8080 in your browser.
eyra index <folder> # Index all images in a folder
eyra index <folder> --auto-caption # Index with auto-captioning
eyra index <folder> --background # Index in background (server must be running)
eyra search <query> # Search (hybrid by default)
eyra search <query> --mode vector # Vector-only search
eyra search <query> --mode keyword # Tag/caption keyword search
eyra tags <keyword> # Search by tag or keyword
eyra similar <image> # Find visually similar images
eyra caption <folder> # Auto-tag/caption already-indexed images
eyra cluster # Group images by visual similarity
eyra cluster --show # View existing clusters
eyra ocr # Extract text from images
eyra export <vault-path> # Export to Obsidian Markdown
eyra timeline # Browse images by date
eyra watch <folder> # Watch folder for new images
eyra stats # Show index statistics
eyra serve # Start web UI at localhost:8080
Eyra uses BLIP to generate captions and tags for every image. Florence-2 is also available as an alternative backend.
# Index and caption at the same time
eyra index ~/Pictures --auto-caption
# Caption already-indexed images
eyra caption ~/Pictures
# Use BLIP (default) or Florence-2
eyra caption ~/Pictures --backend florence2By default, search combines:
- Vector similarity (60%): CLIP embeddings understand visual concepts
- Keyword matching (40%): matches against generated tags and captions
Use --mode vector or --mode keyword to search with only one method.
Groups your images by visual similarity using K-Means on CLIP embeddings. The number of clusters is auto-detected via silhouette score.
# Run clustering (auto-detects optimal number of groups)
eyra cluster
# View saved clusters
eyra cluster --showClusters are also visible in the Web UI under the Clusters tab.
Extract text from images using Apple Vision (macOS) or Tesseract.
# Run OCR on all unprocessed images
eyra ocr
# Re-process all images
eyra ocr --reindex
# Limit to first 100
eyra ocr --limit 100Extracted text is stored in SQLite and searchable via full-text search.
Requirements: install one of:
pip install pytesseract+brew install tesseract(cross-platform)pip install pyobjc-framework-Vision(macOS, uses built-in Apple Vision)
Export indexed images as Obsidian-compatible Markdown notes with YAML frontmatter, Dataview-compatible metadata, and embedded image references.
# Export with absolute image paths (default)
eyra export ~/MyVault
# Copy images into the vault
eyra export ~/MyVault --copy
# Organize by date instead of tags
eyra export ~/MyVault --organize date
# Only export images with a specific tag
eyra export ~/MyVault --tag screenshotEach note includes YAML frontmatter, embedded image reference, OCR text (if available), and an auto-generated index note.
Browse images by date, extracted from EXIF metadata or file modification dates.
# Group by month (default)
eyra timeline
# Group by year
eyra timeline --group year
# Group by day
eyra timeline --group dayAlso available in the Web UI Timeline tab.
The web UI (eyra serve) has four tabs:
- Browse: search, filter, and browse all images with infinite scroll
- Clusters: run and view visual clusters
- Chat: conversational search with natural language summaries
- Timeline: browse images by date
Other features: infinite scroll for large collections, sort by newest/filename/size/dimensions, filter by caption status and format, live task monitoring with progress bars, keyboard navigation in the lightbox, full-text search on captions/tags/OCR text.
Metadata is stored in SQLite alongside ChromaDB vectors:
- Fast filtering by format, dimensions, tags, caption status
- Full-text search (FTS5) on captions, tags, and OCR text
- Instant stats without loading all vectors
- Proper pagination with accurate total counts
Index large folders without blocking the UI:
eyra index ~/Pictures --background
eyra tasks- Scan: finds all images in the folder (JPG, PNG, WebP, HEIC, etc.)
- Embed: generates CLIP embeddings for each image
- Caption: BLIP generates captions and tags (optional)
- Store: embeddings in ChromaDB, metadata in SQLite
- Search: hybrid vector + keyword search
- Discover: clustering, OCR, timeline, export
Everything runs on your machine.
- Python 3.10+
- macOS (Apple Silicon recommended) or Linux
- ~2GB disk space for models (downloaded once on first run)
- OpenCLIP: CLIP embeddings for images and text
- BLIP / Florence-2: vision models for captioning
- ChromaDB: local vector database
- SQLite: metadata storage with FTS5 full-text search
- scikit-learn: K-Means clustering
- FastAPI: web backend
- Typer + Rich: CLI
- Watchdog: file system monitoring
- Pillow: image processing
MIT