Merge pull request #1736 from watchdogpolska/dev_ext

v1.5.16
watchdogpolska · Jun 25, 2024 · c7c2ad0 · c7c2ad0
2 parents 2b574d1 + 347188b
commit c7c2ad0
Show file tree

Hide file tree

Showing 4 changed files with 16 additions and 11 deletions.
diff --git a/feder/__init__.py b/feder/__init__.py
@@ -1,5 +1,5 @@
 # PEP 396: The __version__ attribute's value SHOULD be a string.
-__version__ = "1.5.15"
+__version__ = "1.5.16"
 
 
 # Compatibility to eg. django-rest-framework

diff --git a/feder/llm_evaluation/models.py b/feder/llm_evaluation/models.py
@@ -24,6 +24,7 @@
     NORMALIZED_RESPONSE_QUESTION_KEY,
     EMAIL_CTEGORISATiON_REFUSED,
     answer_categorization,
+    letter_categories_text,
     letter_categorization,
     letter_evaluation_intro,
     letter_response_normalization,
@@ -149,9 +150,11 @@ def categorize_letter(cls, letter):
             institution=institution_name,
             monitoring_question=monitoring_template,
         )
+        categories = letter_categories_text.format(institution=institution_name)
         test_prompt = letter_categorization.format(
             intro=intro,
             institution=institution_name,
+            letter_categories=categories,
             monitoring_response="",
         )
 
@@ -171,6 +174,7 @@ def categorize_letter(cls, letter):
         final_prompt = letter_categorization.format(
             intro=intro,
             institution=institution_name,
+            letter_categories=categories,
             monitoring_response=texts[0],
         )
         letter_llm_request = cls.objects.create(
@@ -201,6 +205,7 @@ def categorize_letter(cls, letter):
                         "intro": intro,
                         "institution": institution_name,
                         "monitoring_response": texts[0],
+                        "letter_categories": categories,
                     }
                 )
             except ValueError as e:
@@ -220,7 +225,7 @@ def categorize_letter(cls, letter):
         letter_llm_request.token_usage = llm_info_dict
         letter_llm_request.status = cls.STATUS.done
         letter_llm_request.save()
-        letter.ai_evaluation = resp
+        letter.ai_evaluation = resp.strip("`").strip()
         if "F) email nie jest odpowiedzią" in resp and "jest spamem" in resp:
             letter.is_spam = letter.SPAM.probable_spam
         letter.save()

diff --git a/feder/llm_evaluation/prompts.py b/feder/llm_evaluation/prompts.py
@@ -43,6 +43,7 @@
 ]
 
 EMAIL_IS_ANSWER = letter_categories_list[0].template[:25]
+EMAIL_IS_SPAM = letter_categories_list[-2].template[:29]
 EMAIL_CTEGORISATiON_REFUSED = (
     letter_categories_list[-1].template + " Odmowa kategoryzacji przez Azure GPT."
 )
@@ -66,7 +67,7 @@
     ```{monitoring_response}```.
     """,
     template_format="f-string",
-    partial_variables={"letter_categories": letter_categories_text},
+    partial_variables={"letter_categories": letter_categories_text.template},
 )
 
 letter_response_normalization = PromptTemplate.from_template(

diff --git a/feder/llm_evaluation/tasks.py b/feder/llm_evaluation/tasks.py
@@ -3,7 +3,7 @@
 from background_task import background
 from django.utils.translation import gettext_lazy as _
 
-from feder.llm_evaluation.prompts import EMAIL_IS_ANSWER
+from feder.llm_evaluation.prompts import EMAIL_IS_ANSWER, EMAIL_IS_SPAM
 
 from .models import LlmLetterRequest, LlmMonitoringRequest
 
@@ -44,20 +44,19 @@ def categorize_letter_in_background(letter_pk):
 
     the_same_content_evaluated = (
         Letter.objects.filter(
-            title=letter.title, body=letter.body, ai_evaluation__isnull=False
+            title=letter.title,
+            body=letter.body,
+            ai_evaluation__isnull=False,
+            ai_evaluation__icontains=EMAIL_IS_SPAM,
         )
         .exclude(pk=letter_pk)
         .first()
     )
 
-    if (
-        the_same_content_evaluated
-        and "F) email nie jest odpowiedzią" in the_same_content_evaluated.ai_evaluation
-        and "jest spamem" in the_same_content_evaluated.ai_evaluation
-    ):
+    if the_same_content_evaluated:
         message = _(
             "AI categorisation skipped for letter with the same content "
-            + "as already evaluated letter: "
+            + "as already evaluated spam letter: "
         ) + str(the_same_content_evaluated.pk)
         logger.info(f"Letter (pk={letter_pk}): {message}")
         letter.ai_evaluation = the_same_content_evaluated.ai_evaluation