Corpus images are embedded with DINOv2 (facebook/dinov2-base, 768-d CLS, L2-normalized), stored in PostgreSQL + pgvector, and indexed in Elasticsearch for kNN (dense_vector, dot product). Search returns image_id from Elasticsearch; metadata and paths come from Postgres.
Scripts (from repo root):
python3 scripts/embed_to_postgres.py: embed files underIMG_PATHinto Postgres.python3 scripts/index_to_elasticsearch.py: bulk-index embeddings intoES_INDEX_NAME(defaultimages_knn).python3 scripts/find_similar.py <query.jpg> [--top-k N]: encode the query and run kNN.
Configure .env: IMG_PATH, DATABASE_URL, ELASTICSEARCH_URL, optional ES_INDEX_NAME.
Images are stored under run_exampes/ so paths work on GitHub and locally.
Query (not in the corpus):
Top-1 match from corpus (20077.jpg):
- Command:
python3 scripts/find_similar.py run_exampes/lighthouse.jpg --top-k 5 - Top-1 result: corpus file
20077.jpg, dot-product score ≈ 0.822 (image_idb5f04aece150f3264667de7699cb44ec38d62c8daf6198a3240d7bcf970a451f). - The index contained 3000 images under
IMG_PATHwhen this was run; scores depend on corpus contents and model.
Query (not in the corpus):
Top-1 match from corpus (20250.jpg):
- Command:
python3 scripts/find_similar.py run_exampes/random_building.jpg --top-k 5 - Top-1 result: corpus file
20250.jpg, dot-product score ≈ 0.663 (image_id619ddd33ec10fa9bd6ee1dbfd6ac804008b49940cd398aecb2971ad8a7b2d642).
Top 5 (corpus filename and dot-product score):
| Rank | File | Score |
|---|---|---|
| 1 | 20250.jpg |
0.663 |
| 2 | 20436.jpg |
0.632 |
| 3 | 21948.jpg |
0.620 |
| 4 | 23660.jpg |
0.619 |
| 5 | 22801.jpg |
0.601 |
Same 3000-image index as the lighthouse run; scores are only comparable within the same model and corpus.