agrigee_lite_client

Thin HTTP client for the agrigee_lite[api] server. Doesn't install Earth Engine, GDAL/PROJ, or DuckDB — it only knows how to build a request, upload/download bytes, and talk to /jobs. See SPECS.md for the full design and the reasoning behind it.

Install

uv pip install agrigee_lite_client

One fixed dependency set, always — geopolars, shapely, and httpx. There's no install variation (no [async], [shapely], etc. extras): the same install already covers both sync and async use.

Use uv, not plain pip, to install this package. geopolars==0.1.0a4 on PyPI ships a non-PEP-440-compliant dependency specifier (pyarrow>=4.0.*), which pip 24.1+ refuses outright ("Please use pip<24.1 if you need to use this version") — pip install agrigee_lite_client fails on any current pip. This is a bug in the published geopolars package itself, not something this package's own metadata can work around. uv's resolver tolerates it and installs cleanly; alternatively, pip<24.1 also works. The same caveat applies to agrigee_lite (the server) itself.

Quickstart — sync

from agrigee_lite_client import AgriGEEClient
from shapely.geometry import Polygon

geometry = Polygon([(-56.42, -11.20), (-56.41, -11.20), (-56.41, -11.19)])

with AgriGEEClient("http://192.168.3.204:8100") as client:
    # a single geometry — naive: internally becomes a 1-row multiple_sits
    df = client.get.sits(geometry, "2023-01-01", "2023-06-01", satellite="Sentinel2")

    # multiple geometries — uploads a GeoParquet, polls the job, downloads the result parquet
    import polars as pl
    from agrigee_lite_client._geoparquet import encode_geometries

    gdf = pl.DataFrame({
        "start_date": ["2023-01-01", "2023-01-01"],
        "end_date": ["2023-06-01", "2023-06-01"],
        "geometry": encode_geometries([geometry, geometry]),
    })
    df = client.get.multiple_sits(gdf, satellite="Sentinel2", reducers=["mean", "std"])

    # geopandas.GeoDataFrame also works (geopandas is optional — only
    # imported if you actually pass one; it's not installed alongside the client)
    import geopandas as gpd

    gdf_geopandas = gpd.GeoDataFrame(
        {"start_date": ["2023-01-01"], "end_date": ["2023-06-01"]},
        geometry=[geometry],
        crs="EPSG:4326",
    )
    df = client.get.multiple_sits(gdf_geopandas, satellite="Sentinel2")

    # image — returns an in-memory zipfile.ZipFile, nothing decoded
    zf = client.get.image(geometry, "2023-01-01", "2023-06-01", satellite="Sentinel2")
    zf.extractall("out/")

Quickstart — async

Same surface, AsyncAgriGEEClient:

import asyncio
from agrigee_lite_client import AsyncAgriGEEClient

async def main():
    async with AsyncAgriGEEClient("http://192.168.3.204:8100") as client:
        df = await client.get.sits(geometry, "2023-01-01", "2023-06-01", satellite="Sentinel2")

asyncio.run(main())

Validation before anything leaves your machine

sits()/multiple_sits() check that the requested dates actually intersect the satellite's operating period — the same check the server does — and drop (with a warning) rows that don't, raising ValueError if none are left. All of this happens before building the GeoParquet or opening a connection, using a static table embedded in the client (no network call). An unknown satellite name is also caught here (AgriGEEUnknownSatelliteError).

Since that table is a copy of what the server knows, the client checks that it's running against the same server version (via GET /version) before trusting it — a mismatch raises AgriGEEVersionMismatchError instead of validating against possibly stale data.

What this client does NOT do (on purpose)

• No CRS reprojection — input geometries must be in WGS84 (EPSG:4326 / OGC:CRS84). If you pass a geopandas.GeoDataFrame with a different CRS, the client rejects it with a clear error instead of reprojecting.
• No local caching, no sophisticated retry/backoff — the cache already lives on the server; the only "retry" here is httpx's transport timeout and a simple poll loop against GET /jobs/{id}.
• No image decoding — get.image(...) returns a raw zipfile.ZipFile, no numpy/tifffile/rasterio.
• get.sits(...) isn't a separate fast path: it's sugar over get.multiple_sits(...) with a single row, so it pays the job+poll+ download cost even for one geometry. See SPECS.md §6.

Versioning

This package's version always matches agrigee_lite's (the server), same commit — they aren't versioned independently. A test (tests/test_version_matches_server.py) enforces this in CI.

Development

cd agrigee_lite_client
uv pip install -e ".[dev]"
pytest
ruff check .

tests/test_client_against_live_api.py only runs if AGRIGEE_TEST_API_URL is set — the rest (test_client_mocked.py, test_async_client_mocked.py) use httpx.MockTransport and need no network.