Merge branch 'openfoodfacts:main' into main

Author: CosmicJesterX

.env

@@ -60,3 +60,10 @@ ENABLE_ML_PREDICTIONS=False
 
 # If you want to enable listening for Redis updates, set this to True
 ENABLE_REDIS_UPDATES=False
+
+# Google specific configuration for Gemini AI
+# Gemini can use Vertex AI or Gemini API, we force it to use Vertex AI
+# as it allows us to set the location
+GOOGLE_GENAI_USE_VERTEXAI=true
+# Force location to be Paris
+GOOGLE_CLOUD_LOCATION=europe-west9

.github/workflows/container-deploy.yml

@@ -145,7 +145,9 @@ jobs:
           echo "POSTGRES_PASSWORD=${{ secrets.POSTGRES_PASSWORD }}" >> .env
           echo "ENVIRONMENT=${{ env.ENVIRONMENT }}" >> .env
           echo "GOOGLE_CLOUD_VISION_API_KEY=${{ secrets.GOOGLE_CLOUD_VISION_API_KEY }}" >> .env
-          echo "GOOGLE_GEMINI_API_KEY=${{ secrets.GOOGLE_GEMINI_API_KEY }}" >> .env
+          echo "GOOGLE_GENAI_USE_VERTEXAI=true" >> .env
+          echo "GOOGLE_CLOUD_LOCATION=europe-west9" >> .env
+          echo "GOOGLE_CREDENTIALS=${{ secrets.GOOGLE_CREDENTIALS }}" >> .env
           echo "TRITON_URI=${{ env.TRITON_URI }}" >> .env
           echo "ENABLE_ML_PREDICTIONS=True" >> .env
           # echo "ENABLE_IMPORT_OFF_DB_TASK=True" >> .env  # disabled because we have the REDIS live updates

.github/workflows/merge-conflicts-autolabel.yml

@@ -7,7 +7,7 @@ on:
 concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}
   cancel-in-progress: true
-  
+
 jobs:
   triage:
     runs-on: ubuntu-latest

config/settings.py

@@ -293,8 +293,9 @@
 # Google Gemini API
 # ------------------------------------------------------------------------------
 
-GOOGLE_GEMINI_API_KEY = os.getenv("GOOGLE_GEMINI_API_KEY")
-
+# Google service account credentials. This is a base64-encoded version of the
+# JSON file
+GOOGLE_CREDENTIALS = os.getenv("GOOGLE_CREDENTIALS")
 
 # Triton Inference Server (ML)
 # ------------------------------------------------------------------------------

docker-compose.yml

@@ -16,7 +16,9 @@ x-api-common: &api-common
     - SENTRY_DSN
     - LOG_LEVEL
     - GOOGLE_CLOUD_VISION_API_KEY
-    - GOOGLE_GEMINI_API_KEY
+    - GOOGLE_GENAI_USE_VERTEXAI
+    - GOOGLE_CLOUD_LOCATION
+    - GOOGLE_CREDENTIALS
     - TRITON_URI
     - ENABLE_ML_PREDICTIONS
     - ENABLE_IMPORT_OFF_DB_TASK

open_prices/common/google.py

@@ -1,5 +1,14 @@
-import google.generativeai as genai
+import base64
+import json
+import logging
+from functools import lru_cache
+from pathlib import Path
+
 from django.conf import settings
+from google import genai
+from google.genai import types
+
+logger = logging.getLogger(__name__)
 
 GOOGLE_CLOUD_VISION_OCR_API_URL = "https://vision.googleapis.com/v1/images:annotate"
 GOOGLE_CLOUD_VISION_OCR_FEATURES = [
@@ -10,14 +19,46 @@
     "FACE_DETECTION",
 ]
 GEMINI_MODEL_NAME = "gemini"
-GEMINI_MODEL_VERSION = "gemini-1.5-flash"
+GEMINI_MODEL_VERSION = "gemini-2.5-flash"
+
+
+def check_google_credentials() -> None:
+    """Create Google Application Default Credentials (ADC) from variable if
+    doesn't exist.
+
+    See
+    https://cloud.google.com/docs/authentication/set-up-adc-local-dev-environment#service-account
+    for more information.
+    """
+    credentials_path = Path(
+        "~/.config/gcloud/application_default_credentials.json"
+    ).expanduser()
+    if not credentials_path.is_file():
+        if settings.GOOGLE_CREDENTIALS:
+            logger.info(
+                "No google credentials found at %s. Creating credentials from GOOGLE_CREDENTIALS.",
+                credentials_path,
+            )
+            credentials_path.parent.mkdir(parents=True, exist_ok=True)
+            credentials_base64 = settings.GOOGLE_CREDENTIALS
+            credentials = json.loads(
+                base64.b64decode(credentials_base64).decode("utf-8")
+            )
+            with open(credentials_path, "w") as f:
+                json.dump(credentials, f, indent=4)
+        else:
+            logger.info(
+                "No google credentials found in environment variable GOOGLE_CREDENTIALS",
+            )
 
 
-genai.configure(api_key=settings.GOOGLE_GEMINI_API_KEY)
-gemini_model = genai.GenerativeModel(model_name=GEMINI_MODEL_VERSION)
+@lru_cache(maxsize=1)
+def get_genai_client() -> genai.Client:
+    check_google_credentials()
+    return genai.Client()
 
 
-def get_generation_config(response_schema):
-    return genai.GenerationConfig(
+def get_generation_config(response_schema) -> types.GenerateContentConfig:
+    return types.GenerateContentConfig(
         response_mime_type="application/json", response_schema=response_schema
     )

open_prices/proofs/ml.py

@@ -193,7 +193,7 @@ def extract_from_price_tag(image: Image.Image) -> Label:
     :param image: the input Pillow image
     :return: the extracted information as a dictionary
     """
-
+    client = common_google.get_genai_client()
     # Gemini model max payload size is 20MB
     # To prevent the payload from being too large, we resize the images
     max_size = 1024
@@ -209,12 +209,13 @@ def extract_from_price_tag(image: Image.Image) -> Label:
         "I expect a single JSON in your reply, no more, no less. "
         "If you cannot decode an attribute, set it to an empty string."
     )
-    response = common_google.gemini_model.generate_content(
-        [
+    response = client.models.generate_content(
+        model=common_google.GEMINI_MODEL_VERSION,
+        contents=[
             prompt,
             image,
         ],
-        generation_config=common_google.get_generation_config(Label),
+        config=common_google.get_generation_config(Label),
     )
     return json.loads(response.text)
 
@@ -249,9 +250,11 @@ def extract_from_price_tags(images: Image.Image) -> Labels:
         f"I expect a list of {len(image_list)} labels in your reply, no more, no less. "
         "If you cannot decode an attribute, set it to an empty string"
     )
-    response = common_google.gemini_model.generate_content(
-        [prompt] + image_list,
-        generation_config=common_google.get_generation_config(Labels),
+    client = common_google.get_genai_client()
+    response = client.models.generate_content(
+        model=common_google.GEMINI_MODEL_VERSION,
+        contents=[prompt] + image_list,
+        config=common_google.get_generation_config(Labels),
     )
     return json.loads(response.text)
 
@@ -267,12 +270,14 @@ def extract_from_receipt(image: Image.Image) -> Receipt:
         image.thumbnail((max_size, max_size))
 
     prompt = "Extract all relevent information, use empty strings for unknown values."
-    response = common_google.gemini_model.generate_content(
-        [
+    client = common_google.get_genai_client()
+    response = client.models.generate_content(
+        model=common_google.GEMINI_MODEL_VERSION,
+        contents=[
             prompt,
             image,
         ],
-        generation_config=common_google.get_generation_config(Receipt),
+        config=common_google.get_generation_config(Receipt),
     )
     return json.loads(response.text)