Model Detail Template¶

Use this page as a template when adding a new per-model documentation page (for example docs/models/<model_id>.md in the future).

Goal: make every model page answer the same questions quickly, so users can compare models without re-reading source code.

Copy This Template¶

Duplicate this page into a new model doc file
Rename the title to the target model (for example RemoteCLIP (remoteclip))
Fill every section marked TODO
Link the page from Supported Models (Overview)
Add or update any model-specific caveats in Supported Models (Advanced Reference) if needed

Template Body¶

`<Model Name>` (`<model_id>`)¶

One-sentence summary: what this model is good at and what rs-embed adapter path it uses.

Quick Facts¶

Field	Value
Model ID	`<model_id>`
Family / Backbone	`TODO`
Adapter type	`precomputed` / `on-the-fly`
Typical backend	`auto` / `gee` / provider-specific
Primary input	`TODO`
Temporal mode	`year` / `range` / model-specific
Output modes	`pooled`, `grid`
Extra side inputs	none / `TODO`
Training alignment (adapter path)	Low / Medium / High + short note

When To Use This Model¶

Good fit for¶

TODO
TODO

Avoid or be careful when¶

TODO
TODO

Input Contract (What the Adapter Expects)¶

Spatial / temporal¶

SpatialSpec: TODO
TemporalSpec: TODO (for example TemporalSpec.range(...) only)
Temporal semantics in rs-embed: TODO (single composite vs multi-frame)

Sensor / channels¶

Collection (default): TODO
Band order (required): TODO
Resolution / scale_m: TODO
Fill / no-data behavior: TODO

Extra metadata / side inputs¶

Required side inputs beyond image tensor: TODO
How they are derived in rs-embed: TODO

Preprocessing Pipeline (Current rs-embed Path)¶

Document the actual adapter path, not the idealized paper pipeline.

Raw provider tensor format: TODO (for example CHW, TCHW)
Value range assumptions: TODO
Normalization steps: TODO
Resize / crop / pad policy: TODO
Any channel conversions or wavelength handling: TODO

Environment variables that change behavior¶

Env var	Default	Effect
`TODO`	`TODO`	`TODO`

Output Semantics¶

`OutputSpec.pooled()`¶

Output shape: TODO
Pooling behavior: TODO
Any caveats: TODO

`OutputSpec.grid()`¶

Output shape: TODO
Is this georeferenced raster space or token grid? TODO
Stitching / tiling behavior caveats (if relevant): TODO

Examples¶

Minimal example¶

from rs_embed import get_embedding, PointBuffer, TemporalSpec, OutputSpec

emb = get_embedding(
    "<model_id>",
    spatial=PointBuffer(lon=0.0, lat=0.0, buffer_m=2048),
    temporal=TemporalSpec.range("2022-06-01", "2022-09-01"),
    output=OutputSpec.pooled(),
    backend="gee",  # TODO adjust if not gee
)

Example with model-specific options¶

# TODO: show custom SensorSpec / env vars / output mode / input_prep as needed

Common Failure Modes / Debugging¶

TODO: missing optional dependency
TODO: wrong band order / channel count
TODO: temporal mode mismatch (year vs range)
TODO: backend mismatch (auto vs provider-specific backend)

Recommended first check:

Use inspect_provider_patch(...) / inspect_gee_patch(...) to inspect raw inputs before blaming the model.

Reproducibility Notes¶

For fair comparisons, document what must be held constant across models:

ROI definition (BBox / PointBuffer)
temporal window and compositing policy
output mode (pooled vs grid)
normalization mode / env overrides
channel order / wavelength schema

Source of Truth (Code Pointers)¶

Catalog registration: src/rs_embed/embedders/catalog.py
Adapter implementation: src/rs_embed/embedders/<file>.py
Related helpers: TODO

Doc Maintenance Checklist¶

[ ] Model ID matches MODEL_SPECS
[ ] Input channels / band order verified in adapter code
[ ] Temporal semantics verified in adapter code
[ ] Env vars verified in adapter code (not copied from stale docs)
[ ] Example snippet runs or is syntactically valid
[ ] Linked from Supported Models (Overview)