fix percentages

app/core/prompts.py

@@ -19,21 +19,21 @@ FOLLOW_SCHEMA_HUMAN = """Convert following attributes to structured schema. Keep
 
 {json_info}"""
 
-GET_PERCENTAGE_SYSTEM = "You have to assign a percentage of certainty from, don't output just 0 and 1"
+GET_PERCENTAGE_SYSTEM = "You are a fashion assistant. You have to assign percentages of cerntainty to each attribute of a product based on the image and product data. You will be given an image or a set of images of a product and set of attributes and should output the percentages of certainty into the given structure."
 
-GET_PERCENTAGE_HUMAN = """For each allowed value in each attribute, assign a percentage of certainty (in scale of 100) that the product fits that value.
-If an attribute can have multiple values, evaluate each value independently. If an attribute can have only one value, the percentages of certainty should sum up to 100.
+GET_PERCENTAGE_HUMAN = """For each allowed value in each attribute, assign a percentage of certainty (from 0 to 100) that the product fits that value.
+For attributes of type list[string], there can be multiple values, and multiple percentages of 100 are possible.
 You should use the following product data to assist you, if available:
 {product_data}
 If an attribute appears in both the image and the product data, use the value from the product data.
 """
 
-REEVALUATE_SYSTEM = "You are an expert in structured data extraction. You will be given an image or a set of images of a product and set of attributes and should reevaluate certainity of the attributes into the given structure."
+REEVALUATE_SYSTEM = "You are a fashion assistant. You have to reevaluate the attributes of a product based on the image and product data. You will be given an image or a set of images of a product and set of attributes and should output the reevaluated attributes into the given structure."
 
 REEVALUATE_HUMAN = """Reevaluate the following attributes of the main product (or {product_taxonomy}) shown in the images. Here are the attributes to reevaluate:
 {product_data}
 
-If an attribute can have multiple values, do not need to reevaluate the values, just the attribute itself. If an attribute can have only one value, reevaluate the top three values. 
+If an attribute has type of string, do not need to reevaluate the values, just the attribute itself. If an attribute has type of list[string], reevaluate the top three values. 
 """
 
 class Prompts(BaseSettings):

app/services/base.py

@@ -27,7 +27,7 @@ def cf_style_to_pydantic_percentage_shema(
         else:
             multiple = False
         class_name = "Class_" + attribute.capitalize()
-        multiple_desc = "Can have multiple values" if multiple else "Single value only"
+        multiple_desc = "multi-label classification" if multiple else "classification"
         attribute_desc = attribute_info.description
         attribute_line = f'{attribute}: {class_name} = Field("", description="{multiple_desc}, {attribute_desc}")'
 
@@ -52,6 +52,12 @@ class Product(BaseModel):
     exec(pydantic_code, globals())
     return Product
 
+def build_attributes_types_prompt(attributes):
+    list_of_types_prompt = "\n List of attributes types:\n"
+    for key, value in attributes.items():
+        list_of_types_prompt += f"- {key}: {value.data_type}\n"
+    return list_of_types_prompt
+
 
 class BaseAttributionService(ABC):
     @abstractmethod
@@ -62,6 +68,7 @@ class BaseAttributionService(ABC):
         img_urls: List[str],
         product_taxonomy: str,
         pil_images: List[Any] = None,
+        appended_prompt: str = "",
     ) -> Dict[str, Any]:
         pass
 
@@ -73,6 +80,7 @@ class BaseAttributionService(ABC):
         img_urls: List[str],
         product_taxonomy: str,
         pil_images: List[Any] = None,
+        appended_prompt: str = "",
     ) -> Dict[str, Any]:
         pass
 
@@ -91,6 +99,7 @@ class BaseAttributionService(ABC):
         product_data: Dict[str, Union[str, List[str]]],
         pil_images: List[Any] = None,
         img_paths: List[str] = None,
+        appended_prompt = str
     ) -> Dict[str, Any]:
         # validate_json_schema(schema)
 
@@ -105,6 +114,8 @@ class BaseAttributionService(ABC):
         for key, value in attributes.items():
             transformed_attributes[forward_mapping[key]] = value
 
+        attributes_types_prompt = build_attributes_types_prompt(attributes)
+
         # attributes_model = convert_attribute_to_model(transformed_attributes)
         attributes_percentage_model = cf_style_to_pydantic_percentage_shema(transformed_attributes)
         schema = attributes_percentage_model.model_json_schema()
@@ -116,6 +127,7 @@ class BaseAttributionService(ABC):
             product_data,
             # pil_images=pil_images, # temporarily removed to save cost
             img_paths=img_paths,
+            appended_prompt=attributes_types_prompt
         )
         validate_json_data(data, schema)
 
@@ -128,6 +140,7 @@ class BaseAttributionService(ABC):
             str_data,
             # pil_images=pil_images, # temporarily removed to save cost
             img_paths=img_paths,
+            appended_prompt=attributes_types_prompt
         )
 
         init_reevaluate_data = {}

app/services/service_openai.py

@@ -68,10 +68,11 @@ class OpenAIService(BaseAttributionService):
         product_data: Dict[str, Union[str, List[str]]],
         pil_images: List[Any] = None,  # do not remove, this is for weave
         img_paths: List[str] = None,
+        appended_prompt: str = "",
     ) -> Dict[str, Any]:
 
         print("Prompt: ")
-        print(prompts.GET_PERCENTAGE_HUMAN_MESSAGE.format(product_taxonomy=product_taxonomy, product_data=product_data_to_str(product_data)))
+        print(prompts.GET_PERCENTAGE_HUMAN_MESSAGE.format(product_taxonomy=product_taxonomy, product_data=product_data_to_str(product_data)) + appended_prompt)
         
         text_content = [
             {
@@ -79,7 +80,7 @@ class OpenAIService(BaseAttributionService):
                 "text": prompts.EXTRACT_INFO_HUMAN_MESSAGE.format(
                     product_taxonomy=product_taxonomy,
                     product_data=product_data_to_str(product_data),
-                ),
+                ) + appended_prompt,
             },
         ]
         if img_urls is not None:
@@ -157,10 +158,11 @@ class OpenAIService(BaseAttributionService):
         product_data: str,
         pil_images: List[Any] = None,  # do not remove, this is for weave
         img_paths: List[str] = None,
+        appended_prompt: str = "",
     ) -> Dict[str, Any]:
 
         print("Prompt: ")
-        print(prompts.REEVALUATE_HUMAN_MESSAGE.format(product_taxonomy=product_taxonomy, product_data=product_data))
+        print(prompts.REEVALUATE_HUMAN_MESSAGE.format(product_taxonomy=product_taxonomy, product_data=product_data) + appended_prompt)
         
         text_content = [
             {
@@ -168,7 +170,7 @@ class OpenAIService(BaseAttributionService):
                 "text": prompts.REEVALUATE_HUMAN_MESSAGE.format(
                     product_taxonomy=product_taxonomy,
                     product_data=product_data,
-                ),
+                ) + appended_prompt,
             },
         ]
         if img_urls is not None: