Mnemosyne · Solr 10 · ImageSpace

Search what the pixels say.

Point ImageCat at folders of photographs. It OCRs them, reads the metadata, builds CLIP and foreground/background indexes, and opens ImageSpace — search, similar, and Lens — on port 8090.

$ bin/imagecat index ./field-notes ./scans

A demo catalog

Invented pictures. Real product.

Nothing here is from a live collection. The grid is what ImageSpace feels like after imagecat index: tiles, OCR snippets, nested folders left intact.

Aerial canyon
ocr: wash · sandstone
Canyon horseshoe
similar to canyon.jpg
Alpine ridge
ocr: granite · blue hour
Heron in marsh
ocr: heron · reeds
Building facade
File_Name: facade.jpg
Harbor lights
ocr: harbor · long exposure
Field notebook
ocr_text hits “notes”
Copper metal
lens: aperture · copper

ImageSpace

The analyst desktop, not another Solr admin.

OCR search

Paddle/RapidOCR into Solr text_ocr. A URL overlay like emmejihad.wordpress.com hits wordpress. Tika MIME/EXIF rides along.

CLIP similar

Click a tile, get neighbors in CLIP space. Incremental FAISS, same catalog, no extra crawl.

Foreground / background

rembg splits the subject from the scene, then CLIP on each. Chair vs garden, vessel vs shelf.

Lens

Mark + / −, Refine fits a tiny Keras head on the CLIP vectors you already have. Save as… Apply later.

Tray

Pin a thumb. Hover × to drop it. The working set stays on the left while you search.

OPSUI

Watch OCR 46/653, then Jaccard, CLIP, fg/bg. Status PGE EXEC, progress from the job dir.

Similar

Same canyon. Different frame.

CLIP does not need a filename match. These two aerials would sit next to each other after Similar, even if one lived under canyon1/20150416/ and the other under canyon2/.

query
neighbor
Red chair in a garden

Foreground / background

The chair, or the garden.

FG similar ranks on the subject. BG similar ranks on what is behind it. Same photograph, two indexes, two questions.

Lens

Yes, no, Refine, Save as…

A lens is not a pretrained model. You label a handful of tiles. ImageCat fits a head and reranks the catalog. Apply it next week.

yes
yes
no
no

One command

Index hides the workflow.

  1. indexCopy JPGs, nested paths kept
  2. OCRPaddle into Solr imagecat
  3. JaccardTika keys / values
  4. CLIPFAISS similar
  5. fg / bgrembg + CLIP

bin/imagecat

The operator path.

$ bin/imagecat setup
$ bin/imagecat start
$ bin/imagecat index ./field-notes ./scans
$ bin/imagecat status
$ bin/imagecat reset --yes

Adds, does not wipe. reset empties the catalog; staging stays. Lenses stay unless --lenses.

Command line