Local multimodal memory with semantic search.

Embed images, audio, video, PDFs, and text into a local vector database — then find anything with a natural language query. A text search for "team dinner" surfaces the photos, even though the photos have no text metadata.

Comes with an animated setup wizard and a Raycast extension for instant visual search.

Powered by Gemini Embedding 2 (768-dim, free tier) and ChromaDB stored entirely on your machine.

You                     Gemini Embedding 2          ChromaDB (local)
 |                            |                        |
 |-- team-dinner.jpg -------->|-- 768-dim vector ------>|-- stored on disk
 |-- meeting-notes.pdf ------>|-- 768-dim vector ------>|-- stored on disk
 |-- "team dinner" (query) -->|-- query vector -------->|-- cosine search
 |<---------------------------|<-- ranked results -------|

Cross-modal search works out of the box. No tagging, no renaming, no metadata required.

git clone https://github.com/hughminhphan/vector-embedded-finder.git
cd vector-embedded-finder
pip install -e .
python setup_wizard.py

That's it. The animated setup wizard handles everything end-to-end:

Manual key setup: cp .env.example .env and add GEMINI_API_KEY=your_key

from vector_embedded_finder import search, ingest_file, ingest_directory, count

# Embed a single file — image, PDF, audio, video, or text
ingest_file("~/Photos/team-dinner.jpg")

# Embed an entire directory (recursive by default)
ingest_directory("~/Documents/", source="docs")

# Search with natural language
matches = search("team dinner at the rooftop", n_results=5)
for m in matches:
    print(f"{m['file_name']}  {m['similarity']:.0%} match  {m['file_path']}")

# Housekeeping
print(f"{count()} items indexed")

ingest_file returns {"status": "embedded" | "skipped" | "error", ...}. Files are SHA-256 deduplicated — re-ingesting the same file is a no-op.

Visual grid search with image thumbnails, right from your launcher.

cd raycast
npm install
npx ray develop

Open Raycast and search Memory Search. On first launch, go to Preferences and set:

Tip: Run python setup_wizard.py first — it prints these values pre-filled for you.

vector_embedded_finder/
  config.py      — env, API key, supported types, paths
  embedder.py    — Gemini Embedding 2 (text, image, audio, video, PDF)
  store.py       — ChromaDB layer: cosine distance, SHA-256 dedup, upsert
  search.py      — natural language search with similarity scoring
  ingest.py      — file detection → embedding → storage pipeline
  utils.py       — hashing, MIME detection, timestamp helpers

raycast/
  src/lib/runner.ts       — Python bridge: spawnSync + JSON envelope protocol
  src/search-memory.tsx   — grid search UI with thumbnails
  src/open-memory.tsx     — headless instant file opener

setup_wizard.py           — animated CLI setup: key → folders → index → Raycast card

All vectors are stored locally in data/chromadb/. The only outbound traffic is embedding API calls to Google — your files never leave your machine.

MIT