Ask Qwen Task

pyproject.toml

@@ -75,6 +75,7 @@ tasks = [
   "pyvips==3.1.1.8.18.1",
   "uvdat-flood-sim[large-image-writer]==1.0.4",
   "xdg-base-dirs==6.0.2",
+  "huggingface-hub>=1.14.0",
 ]
 
 [dependency-groups]

terraform/django.tf

@@ -24,6 +24,9 @@ module "django" {
     DJANGO_DATABASE_POOL_MAX_SIZE = "12"
     DJANGO_SENTRY_DSN             = "https://5302701c88f1fa6ec056e0c269071191@o267860.ingest.us.sentry.io/4510620385804288"
   }
+  additional_sensitive_django_vars = {
+    DJANGO_UVDAT_HF_TOKEN = var.DJANGO_UVDAT_HF_TOKEN
+  }
   django_cors_allowed_origins = [
     # Can't make this use "aws_route53_record.www.fqdn" because of a circular dependency
     "https://www.geodatalytics.kitware.com",

terraform/variables.tf

@@ -3,3 +3,9 @@ variable "SENTRY_AUTH_TOKEN" {
   nullable  = false
   sensitive = true
 }
+
+variable "DJANGO_UVDAT_HF_TOKEN" {
+  type      = string
+  nullable  = true
+  sensitive = true
+}

uvdat/core/tasks/analytics/__init__.py

@@ -6,6 +6,7 @@
 from .flood_network_failure import FloodNetworkFailure
 from .flood_simulation import FloodSimulation
 from .geoai_segmentation import GeoAISegmentation
+from .imagery_ask_qwen import ImageryAskQwen
 from .network_recovery import NetworkRecovery
 from .uncertainty_quantification import UncertaintyQuantification
 
@@ -15,6 +16,7 @@
 analysis_types: list[type[AnalysisType]] = [
     FloodSimulation,
     FloodNetworkFailure,
+    ImageryAskQwen,
     NetworkRecovery,
     UncertaintyQuantification,
     GeoAISegmentation,

uvdat/core/tasks/analytics/imagery_ask_qwen.py

@@ -0,0 +1,140 @@
+from __future__ import annotations
+
+import base64
+
+from celery import shared_task
+from django.conf import settings
+from django_large_image import utilities
+import large_image
+
+from uvdat.core.models import RasterData, TaskResult
+
+from .analysis_type import AnalysisInputError, AnalysisTask, AnalysisType
+
+ENDPOINT_NAMESPACE = "Kitware"
+ENDPOINT_NAME = "qwen3-5-9b-gguf-ulh"
+MODEL_CARD_URL = "https://huggingface.co/unsloth/Qwen3.5-9B-GGUF"
+PROMPT_PREFIX = (
+    "You are a geospatial analyst. Answer the following question about the provided image."
+)
+TOKEN_RANGE = {"min": 1000, "max": 10000, "step": 1000}
+THUMBNAIL_WIDTH = 2000
+
+
+class ImageryAskQwen(AnalysisType):
+    def __init__(self):
+        super().__init__()
+        self.name = "Imagery: Ask Qwen"
+        self.description = "Select an imagery layer and ask Qwen 3.5 about it."
+        self.details = (
+            "Inferencing with unsloth/Qwen3.5-9B-GGUF provided by a "
+            "Kitware-hosted Huggingface Inference Endpoint. "
+            f"See the model card at {MODEL_CARD_URL}. "
+            f'Prompts will be prefixed with "{PROMPT_PREFIX}".'
+            "Responses may cut off mid-sentence if max_tokens is reached."
+        )
+        self.db_value = "imagery_ask_qwen"
+        self.input_types = {
+            "imagery": "RasterData",
+            "text_prompt": "string",
+            "max_tokens": "number",
+        }
+        self.output_types = {
+            "response": "markdown",
+        }
+        self.attribution = "Unsloth AI, Kitware Inc."
+
+    @classmethod
+    def is_enabled(cls):
+        return settings.UVDAT_ENABLE_IMAGERY_ASK_QWEN and settings.UVDAT_HF_TOKEN
+
+    def get_input_options(self):
+        return {
+            "imagery": RasterData.objects.filter(dataset__category="imagery"),
+            "text_prompt": [],
+            "max_tokens": [TOKEN_RANGE],
+        }
+
+    def validate_inputs(self, inputs):
+        super().validate_inputs(inputs)
+        try:
+            imagery = RasterData.objects.get(id=inputs.get("imagery"))
+        except RasterData.DoesNotExist as e:
+            err_msg = "Imagery raster does not exist."
+            raise AnalysisInputError(err_msg) from e
+        if imagery.dataset.category != "imagery":
+            err_msg = 'Selected raster is not categorized as "imagery".'
+            raise AnalysisInputError(err_msg)
+        max_tokens = int(inputs.get("max_tokens"))
+        if max_tokens < TOKEN_RANGE["min"] or max_tokens > TOKEN_RANGE["max"]:
+            err_msg = f"max_tokens must be between {TOKEN_RANGE['min']} and {TOKEN_RANGE['max']}."
+            raise AnalysisInputError(err_msg)
+
+    def run_task(self, *, project, **inputs):
+        text_prompt = inputs.get("text_prompt")
+        result = TaskResult.objects.create(
+            name=text_prompt[:250],
+            task_type=self.db_value,
+            inputs=inputs,
+            project=project,
+            status="Initializing Task...",
+        )
+        imagery_ask_qwen.delay(result.id)
+        return result
+
+    def finalize(self, result):
+        pass
+
+
+@shared_task(base=AnalysisTask)
+def imagery_ask_qwen(result_id):
+    # Only available with [tasks] extra
+    from huggingface_hub import get_inference_endpoint  # noqa: PLC0415
+
+    result = TaskResult.objects.get(id=result_id)
+    imagery = RasterData.objects.get(id=result.inputs.get("imagery"))
+    text_prompt = result.inputs.get("text_prompt")
+    max_tokens = int(result.inputs.get("max_tokens"))
+
+    result.write_status("Encoding imagery...")
+    imagery_path = utilities.field_file_to_local_path(imagery.cloud_optimized_geotiff)
+    src = large_image.open(imagery_path)
+    thumbnail_bytes, _ = src.getThumbnail(THUMBNAIL_WIDTH, encoding="PNG")
+    thumbnail_b64 = base64.b64encode(thumbnail_bytes).decode("utf-8")
+    thumbnail_uri = f"data:image/jpeg;base64,{thumbnail_b64}"
+
+    result.write_status("Starting inference endpoint...")
+    endpoint = get_inference_endpoint(
+        name=ENDPOINT_NAME,
+        namespace=ENDPOINT_NAMESPACE,
+        token=settings.UVDAT_HF_TOKEN,
+    )
+    endpoint.resume()
+    endpoint.wait()
+
+    result.write_status("Sending question to Qwen...")
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "image_url", "image_url": {"url": thumbnail_uri}},
+                {"type": "text", "text": f"{PROMPT_PREFIX} {text_prompt}"},
+            ],
+        }
+    ]
+
+    result.write_status("Awaiting Qwen's response...")
+    chat = endpoint.client.chat.completions.create(
+        model="unsloth/Qwen3.5-9B-GGUF",
+        messages=messages,
+        stream=False,
+        max_tokens=max_tokens,
+    )
+    response = ""
+    for choice in chat.choices:
+        if choice.finish_reason == "length":
+            # max tokens exceeded, use reasoning content
+            response += choice.message.reasoning_content
+        else:
+            response += choice.message.content
+    result.write_outputs({"response": response})

uvdat/core/tests/test_analytics.py

@@ -19,9 +19,10 @@ def test_rest_list_analysis_types(user, authenticated_api_client, project):
     user.is_superuser = True
     user.save()
 
-    analysis_type_instances = [at() for at in analysis_types]
+    analysis_type_instances = [at() for at in analysis_types if at.is_enabled()]
     resp = authenticated_api_client.get(f"/api/v1/analytics/project/{project.id}/types/")
     data = resp.json()
+
     assert len(data) == len(analysis_type_instances)
     assert {type_info.get("name") for type_info in data} == {
         i.name for i in analysis_type_instances

uvdat/settings/base.py

@@ -158,6 +158,7 @@
 }
 
 UVDAT_WEB_URL: str = env.url("DJANGO_UVDAT_WEB_URL").geturl()
+UVDAT_HF_TOKEN: str = env.str("DJANGO_UVDAT_HF_TOKEN", default=None)
 UVDAT_ENABLE_FLOOD_SIMULATION: bool = env.bool("DJANGO_UVDAT_ENABLE_FLOOD_SIMULATION", default=True)
 UVDAT_ENABLE_FLOOD_NETWORK_FAILURE: bool = env.bool(
     "DJANGO_UVDAT_ENABLE_FLOOD_NETWORK_FAILURE", default=True
@@ -172,6 +173,7 @@
 UVDAT_ENABLE_UNCERTAINTY_QUANTIFICATION: bool = env.bool(
     "DJANGO_UVDAT_ENABLE_UNCERTAINTY_QUANTIFICATION", default=True
 )
+UVDAT_ENABLE_IMAGERY_ASK_QWEN: bool = env.bool("DJANGO_UVDAT_ENABLE_IMAGERY_ASK_QWEN", default=True)
 
 logging.getLogger("pyvips").setLevel(logging.ERROR)
 logging.getLogger("rasterio").setLevel(logging.ERROR)

web/package.json

@@ -35,6 +35,7 @@
     "vue": "3.5.34",
     "vue-chartjs": "5.3.3",
     "vue-maplibre-compare": "1.0.27",
+    "vue-markdown-render": "^2.3.1",
     "vue-router": "5.0.6",
     "vuedraggable": "4.1.0",
     "vuetify": "4.0.7"