Using several search models

Search quality depends on the model, and the only way to know whether a larger one helps your photos is to try it. videre keeps each model’s work separate, so trying one costs time but risks nothing.

The models

Model	Download	Notes
google/siglip-base-patch16-224	~780 MB	The default
google/siglip2-base-patch16-384	~1.4 GB	Newer, higher resolution
google/siglip-so400m-patch14-384	~3.3 GB	Largest

Each links to its model card on Hugging Face, where the training data, intended use, and limitations are documented by the people who built it.

Higher resolution and more parameters generally mean better matching on fine detail, at proportionally more time per image and more disk.

Trying one

videre embed --model google/siglip2-base-patch16-384

That is a full second pass over every image, taking about as long as the first, plus its own download. The existing vectors are untouched throughout.

Then compare on queries you care about:

videre search "kids playing in snow" --scores
videre search "kids playing in snow" --scores --model google/siglip2-base-patch16-384

Use your own hard queries, not easy ones. Anything finds a sunset; the difference shows up on specific requests such as “a red bicycle against a brick wall”.

Scores are not comparable between models. Each has its own scale. Compare the ordering and which photos appear, not the numbers.

Switching

videre config set model google/siglip2-base-patch16-384
videre search "sunset"                    # now uses it by default

Switching invalidates nothing. The previous model’s vectors stay on disk and stay queryable:

videre search "sunset" --model google/siglip-base-patch16-224

Switching back is instant, because nothing was deleted.

Seeing what you have

videre stats

Embeddings:
  google/siglip-base-patch16-224              196   768-dim     400 KB
  google/siglip2-base-patch16-384             196   768-dim     400 KB

One line per prepared model, with dimensions read from the stored data rather than a hardcoded table, so an unfamiliar model reports honestly.

Asking for a model you have not prepared gives an error listing the ones you do have, rather than silently returning nothing. videre report is the exception: a missing model disables its similarity button with a note rather than failing a report that is otherwise fine.

Classification is per model too

videre classify scores a specific model’s vectors, and results are stored per model:

videre classify --model google/siglip2-base-patch16-384
videre search --category screenshot --model google/siglip2-base-patch16-384

Preparing a new model does not carry classifications across. Run classify again under it, which is quick since it reuses the vectors.

Removing one

Each model’s data is one file:

~/.videre/embeddings/<library>-<hash>/<owner>--<model>.db

ls -la ~/.videre/embeddings/*/
rm ~/.videre/embeddings/hashes-3f9a1c04e7b25d68/google--siglip2-base-patch16-384.db

Deleting it removes only that model’s vectors. Rerun videre embed --model ... to rebuild.

Also remove the weights if you are not keeping the model:

rm -rf ~/.cache/huggingface/hub/models--google--siglip2-base-patch16-384

Why they are stored per library

~/.videre/embeddings/<library>-<hash>/ is keyed by library as well as model, even though content hashes would technically allow sharing between libraries.

videre prune cannot see another library’s files, so a shared layout would let one library’s cleanup delete vectors another still needed. A thumbnail lost that way costs milliseconds; an embedding costs hours. See caches for the same tradeoff decided the other way.

The library part of the path includes a hash of its canonical path, so two databases both named photos.db in different folders never collide.

Caveats

Budget the disk. Each model is roughly 130 MB to 190 MB of vectors per 70,000 photos, plus its download. Three models is a few GB before any photos.

Do not run two embed passes at once. They both convert HEIC, and contending for that makes both dramatically slower. See long-running jobs.

--batch is capped at 96 regardless of model. Above roughly 121 the batched path silently produces wrong vectors, so higher values are reduced with a warning. See videre embed.

Libraries from before 0.10 kept vectors in the main database. That fallback was removed in 0.11, so such a library reports the model as missing. Nothing is deleted; rerun videre embed to rebuild in the current layout.