- Making Gemini more robust with error hnadling and parsing various kind of outputs that are possible.

- integrating gemini-2.5-flash
- integrating gpt5-mini and nano
- Estimating the prices and times per 1000 messages for each model.

Config_files/message_system_config.json

@@ -18,16 +18,17 @@
     "genres"
   ],
   "AI_Jargon": ["elevate", "enhance", "ignite", "reignite", "rekindle", "rediscover","passion", "boost", "fuel", "thrill", "revive", "spark", "performing", "fresh", "tone", "enthusiasm", "illuminate"],
+  "singeo_banned_phrases": "    Banned phrases:\n    Voice is NOT an instrument, so avoid phrases like below:\n    - Your voice is waiting\n    - Your voice awaits\n    - Your voice needs you\n    - Your voice is calling\n    - Your voice deserves more\n    - Hit the high notes / Hit those notes\n    - ...\n",
   "AI_phrases_singeo": ["your voice deserves more"],
   "header_limit": 30,
   "message_limit": 135,
-  "LLM_models": ["gpt-4o-mini", "gpt-4o", "gpt-4.1-mini", "gemini-2.5-flash-lite-preview-06-17", "claude-3-5-haiku-latest", "google/gemma-3-27b-instruct/bf-16"],
-  "openai_models": ["gpt-4o-mini", "gpt-4o", "gpt-4.1-nano", "gpt-3.5-turbo", "gpt-4.1-mini"],
-  "reasoning": ["o1", "o4-mini", "o1-mini", "o3-mini"],
+  "LLM_models": ["gpt-4o-mini", "gpt-5-mini", "gpt-5-nano", "gemini-2.5-flash","claude-3-5-haiku-latest", "google/gemma-3-27b-instruct/bf-16"],
+  "openai_models": ["gpt-4o-mini", "gpt-4o", "gpt-4.1-nano", "gpt-3.5-turbo", "gpt-4.1-mini", "gpt-5-mini", "gpt-5-nano"],
+  "reasoning": ["o1", "o4-mini", "o1-mini", "o3-mini", "gpt-5-mini", "gpt-5-nano"],
   "ollama_models": ["deepseek-r1:1.5b", "gemma3:4b", "deepseek-r1:7b", "gemma3:4b"],
   "claude_models": ["claude-3-5-haiku-latest"],
   "inference_models": ["google/gemma-3-27b-instruct/bf-16", "meta-llama/llama-3.2-11b-instruct/fp-16"],
-  "google_models": ["gemini-2.5-flash-lite-preview-06-17"]
+  "google_models": ["gemini-2.5-flash-lite", "gemini-2.5-flash", "gemini-2.0-flash"]
 }
 
 

Messaging_system/CoreConfig.py

@@ -1,7 +1,7 @@
 """
 the flow of the Program starts from create_personalized_message function
 """
-
+import os
 import time
 from Messaging_system.SnowFlakeConnection import SnowFlakeConn
 
@@ -75,16 +75,17 @@ class CoreConfig:
     # ===============================================================
     def get_brand_voice(self):
         """
-        get positive phrases for the brand to instruct the llm to talk like this
-        :return:
+        Get positive phrases for the brand to instruct the LLM to talk like this.
+        Returns:
+            str: The brand voice phrases if file exists, otherwise None.
         """
-        # assuming your file is named 'example.txt' and is in the same directory
         file_path = f'Config_files/{self.brand.lower()}_phrases.txt'
 
-        with open(file_path, 'r', encoding='utf-8') as f:
-            string = f.read()
+        if not os.path.exists(file_path):
+            return None
 
-        return string
+        with open(file_path, 'r', encoding='utf-8') as f:
+            return f.read()
 
     # --------------------------------------------------------------
     # --------------------------------------------------------------

Messaging_system/LLM.py

@@ -171,6 +171,17 @@ class LLM:
                         response_mime_type="application/json"
                     ))
 
+                tokens = {
+                    'prompt_tokens': response.usage_metadata.prompt_token_count,
+                    'completion_tokens': response.usage_metadata.candidates_token_count,
+                    'total_tokens': response.usage_metadata.total_token_count
+                }
+
+                # validating the JSON
+                self.Core.total_tokens['prompt_tokens'] += tokens['prompt_tokens']
+                self.Core.total_tokens['completion_tokens'] += tokens['completion_tokens']
+                self.Core.temp_token_counter += tokens['total_tokens']
+
                 output = self.preprocess_and_parse_json(response.text)
                 # output = json.loads(str(response.text))
 
@@ -218,7 +229,7 @@ class LLM:
                             {"role": "system", "content": instructions},
                             {"role": "user", "content": prompt}
                         ],
-                        reasoning_effort="medium",
+                        reasoning_effort="minimal",
                         n=1,
                     )
 
@@ -240,6 +251,11 @@ class LLM:
                     'total_tokens': response.usage.total_tokens
                 }
 
+                # validating the JSON
+                self.Core.total_tokens['prompt_tokens'] += tokens['prompt_tokens']
+                self.Core.total_tokens['completion_tokens'] += tokens['completion_tokens']
+                self.Core.temp_token_counter += tokens['total_tokens']
+
                 try:
                     content = response.choices[0].message.content
 
@@ -258,10 +274,6 @@ class LLM:
                                 f"'header' or 'message' is more than specified characters in response on attempt {attempt + 1}. Retrying...")
                             continue
 
-                    # validating the JSON
-                    self.Core.total_tokens['prompt_tokens'] += tokens['prompt_tokens']
-                    self.Core.total_tokens['completion_tokens'] += tokens['completion_tokens']
-                    self.Core.temp_token_counter += tokens['total_tokens']
                     return output
 
                 except json.JSONDecodeError:
@@ -398,30 +410,56 @@ class LLM:
 
     def preprocess_and_parse_json(self, response: str):
         """
-        Cleans an LLM response by removing <think> tags and extracting JSON
-        from ```json ... ``` fences (or bare text if no fence is found),
-        then returns the parsed object or None on failure.
+        Remove <think> blocks, extract JSON (from ```json fences or first {...} block),
+        and parse. Includes a repair pass to handle common LLM issues like trailing commas.
         """
-        # 1) Remove all <think>...</think> blocks
-        cleaned = re.sub(r'<think>.*?</think>', '', response, flags=re.DOTALL)
-
-        # 2) Look for a ```json ... ``` fenced block
-        fence_pattern = re.compile(r'```json(.*?)```', flags=re.DOTALL)
-        fence_match = fence_pattern.search(cleaned)
-        if fence_match:
-            json_text = fence_match.group(1).strip()
-        else:
-            # No fence; assume whole cleaned text is JSON
-            json_text = cleaned.strip()
 
-        # 3) Attempt to parse
+        def extract_json(text: str) -> str:
+            # Remove <think>...</think>
+            text = re.sub(r'<think>.*?</think>', '', text, flags=re.DOTALL).strip()
+
+            # Prefer fenced code if present
+            fence = re.search(r'```(?:json)?(.*?)```', text, flags=re.DOTALL | re.IGNORECASE)
+            if fence:
+                return fence.group(1).strip()
+
+            # Otherwise, grab the first {...} block
+            brace = re.search(r'\{.*\}', text, flags=re.DOTALL)
+            return brace.group(0).strip() if brace else text.strip()
+
+        def normalize_quotes(text: str) -> str:
+            return (text
+                    .replace('\ufeff', '')  # strip BOM if present
+                    .replace('“', '"').replace('”', '"')
+                    .replace('‘', "'").replace('’', "'"))
+
+        def strip_comments(text: str) -> str:
+            # Remove // line comments and /* block comments */
+            text = re.sub(r'//.*?$', '', text, flags=re.MULTILINE)
+            text = re.sub(r'/\*.*?\*/', '', text, flags=re.DOTALL)
+            return text
+
+        def remove_trailing_commas(text: str) -> str:
+            # Remove commas before } or ]
+            return re.sub(r',(\s*[}\]])', r'\1', text)
+
+        raw = extract_json(response)
+        raw = normalize_quotes(raw)
+
         try:
-            return json.loads(json_text)
-        except json.JSONDecodeError as e:
-            print(f"Failed to parse JSON: {e}")
-            # Optionally, log the offending text for debugging:
-            # print("Offending text:", json_text)
-            return None
+            return json.loads(raw)
+        except json.JSONDecodeError:
+            # Repair pass
+            repaired = strip_comments(raw)
+            repaired = remove_trailing_commas(repaired)
+            repaired = repaired.strip()
+
+            try:
+                return json.loads(repaired)
+            except json.JSONDecodeError as e:
+                print(f"Failed to parse JSON: {e}")
+                # print('Offending text:', repaired)
+                return None
 
     # ===============================================================
     def preprocess_and_parse_json_claude(self, response: str):

Messaging_system/Message_generator.py

@@ -175,33 +175,37 @@ class MessageGenerator:
         """
 
         jargon_list = "\n".join(f"- {word}" for word in self.Core.config_file["AI_Jargon"])
+        banned_phrases = self.Core.config_file.get(f"{self.Core.brand}_banned_phrases", None)
 
         instructions = f"""
 You are a copywriter. Your task is to write a 'header' and a 'message' as a push notification for a {self.Core.get_instrument()} student. It is critical  that the message sounds like natural, everyday speech: friendly, concise, no jargon, and it must follow the instructions.
-Write a SUPER CASUAL and NATURAL push notification, as if you are chatting over coffee. Avoid odd phrasings. The message should sound like something that a {self.Core.get_instrument()} instructor would realistically say to a student in a daily conversation. Here are some examples of things that an instructor would realistically say to a student, to give you a general sense of tone and phrasing:
+Write a SUPER CASUAL and NATURAL push notification, as if you are chatting over coffee. Avoid odd phrasings. The message should sound like something that a {self.Core.get_instrument()} instructor would realistically say to a student in a daily conversation. 
+
+"""
+        if self.Core.brand_voice is not None:
+                instructions += f"""
+Here are some examples of things that an instructor would realistically say to a student, to give you a general sense of tone and phrasing:
 
 Common instructor phrases:
 {self.Core.brand_voice}
+"""
 
 
+        rules = f"""
 ABSOLUTE RULE – OVERRIDES EVERYTHING ELSE:  
 the header and the message **MUST NOT** contain any banned word or phrases(case-insensitive; singular, plural, verb forms, or their derivatives)
 - **important Note:** header **must be** less than {self.Core.config_file["header_limit"]} characters and message **must be less** than {self.Core.config_file["message_limit"]} characters.
 
-
 Banned word:
 {jargon_list}
 
-Banned phrases:
-Voice is NOT an instrument, so avoid phrases like below:
-- Your voice is waiting
-- Your voice awaits
-- Your voice needs you
-- Your voice is calling
-- Your voice deserves more
-- Hit the high notes / Hit those notes
-- ...
-
 """
+        if banned_phrases is not None:
+            rules += banned_phrases
+
+        final_instructions = f"""
+{instructions}
+{rules}
+        """
 
-        return instructions
+        return final_instructions

Messaging_system/MultiMessage.py

@@ -172,7 +172,7 @@ Mimic the example style as much as possible and make it personalized using provi
         # Combine into a single string
         previous_text_str = "\n\n".join(previous_text)
 
-        # user_info = self.promptGen.get_user_profile(user=user)
+        user_info = self.promptGen.get_user_profile(user=user)
         input_context = self.promptGen.input_context()
         instructions = self.message_instructions(step)
         recommendation_instructions = self.promptGen.recommendations_instructions(user)
@@ -195,6 +195,8 @@ Write the *next* follow up personalized push notification following the instruct
 - The new header and message should be different from previous messages and we should not have similar words and phrases from previous messages.
 - You are **ONLY ALLOWED** to use {self.Core.get_emoji()} if we didn't use in previous messages. (ONLY ONCE, and ONLY at the end of header or message).
 
+{user_info}
+
 {recommendation_instructions}
 
 {examples}
@@ -362,33 +364,39 @@ Write the *next* follow up personalized push notification following the instruct
         """
 
         jargon_list = "\n".join(f"- {word}" for word in self.Core.config_file["AI_Jargon"])
+        banned_phrases = self.Core.config_file.get(f"{self.Core.brand}_banned_phrases", None)
 
         instructions = f"""
-You are a copywriter. Your task is to write a 'header' and a 'message' as a push notification for a {self.Core.get_instrument()} student. It is critical  that the message sounds like natural, everyday speech: friendly, concise, no jargon, and it must follow the instructions.
-Write a SUPER CASUAL and NATURAL push notification, as if you are chatting over coffee. Avoid odd phrasings. The message should sound like something that a {self.Core.get_instrument()} instructor would realistically say to a student. Here are some examples of things that an instructor would realistically say to a student, to give you a general sense of tone and phrasing:
+    You are a copywriter. Your task is to write a 'header' and a 'message' as a push notification for a {self.Core.get_instrument()} student. It is critical  that the message sounds like natural, everyday speech: friendly, concise, no jargon, and it must follow the instructions.
+    Write a SUPER CASUAL and NATURAL push notification, as if you are chatting over coffee. Avoid odd phrasings. The message should sound like something that a {self.Core.get_instrument()} instructor would realistically say to a student in a daily conversation. 
 
-Common instructor phrases:
-{self.Core.brand_voice}
+    """
+        if self.Core.brand_voice is not None:
+            instructions += f"""
+    Here are some examples of things that an instructor would realistically say to a student, to give you a general sense of tone and phrasing:
 
-ABSOLUTE RULE – OVERRIDES EVERYTHING ELSE:  
-the header and the message **MUST NOT** contain any banned word or phrases(case-insensitive; singular, plural, verb forms, or their derivatives)
-- **important Note:** header **must be** less than {self.Core.config_file["header_limit"]} characters and message **must be less** than {self.Core.config_file["message_limit"]} characters.
+    Common instructor phrases:
+    {self.Core.brand_voice}
+    """
 
-Banned word:
-{jargon_list}
+        rules = f"""
+    ABSOLUTE RULE – OVERRIDES EVERYTHING ELSE:  
+    the header and the message **MUST NOT** contain any banned word or phrases(case-insensitive; singular, plural, verb forms, or their derivatives)
+    - **important Note:** header **must be** less than {self.Core.config_file["header_limit"]} characters and message **must be less** than {self.Core.config_file["message_limit"]} characters.
 
-Banned phrases:
-Voice is NOT an instrument, so avoid phrases like below:
-- Your voice is waiting
-- Your voice awaits
-- Your voice needs you
-- Your voice is calling
-- Your voice deserves more
-- Hit the high notes / Hit those notes
-- ...
 
-"""
+    Banned word:
+    {jargon_list}
 
+    """
+        if banned_phrases is not None:
+            rules += banned_phrases
+
+        final_instructions = f"""
+{instructions}
+{rules}
+            """
+
+        return final_instructions
 
-        return instructions
 

Messaging_system/Permes.py

@@ -90,9 +90,14 @@ class Permes:
         total_prompt_tokens = personalize_message.total_tokens["prompt_tokens"]
         total_completion_tokens = personalize_message.total_tokens["completion_tokens"]
 
-        total_cost = ((total_prompt_tokens / 1000000) * 0.10) + (
-                (total_completion_tokens / 1000000) * 0.4)  # Cost calculation estimation
-        print(f"Estimated Cost (USD): {total_cost:.5f}")
+        total_cost = self.calculate_cost(total_prompt_tokens, total_completion_tokens, model)
+
+        print(f"Estimated Cost (USD): {total_cost:.5f} ---> Number of messages: {(len(users_df) * number_of_messages)}")
+        st.write(f"Estimated Cost (USD): {total_cost:.5f} ---> Number of messages: {(len(users_df) * number_of_messages)}")
+
+        scale_price = (total_cost * 1000) / (len(users_df) * number_of_messages)
+        print(f"Estimated Cost (USD) for 1000 messages: {scale_price}")
+        st.write(f"Estimated Cost (USD) for 1000 messages: {scale_price}")
 
         # Storing process can also happen after some evaluation steps
         # snowflake_conn = SnowFlakeConn(session=session, brand=brand)
@@ -103,6 +108,48 @@ class Permes:
         return users_df
 
     # -----------------------------------------------------
+    def calculate_cost(self, total_prompt_tokens, total_completion_tokens, model):
+        input_price, output_price = self.get_model_price(model)
+
+        total_cost = ((total_prompt_tokens / 1000000) * input_price) + (
+                (total_completion_tokens / 1000000) * output_price)  # Cost calculation estimation
+
+        return total_cost
+
+    # ====================================================
+    def get_model_price(self, model):
+        """
+        getting the input price and output price per 1m token for the requested model
+        :param model:
+        :return:
+        """
+
+        input_prices = {
+            "gpt-4o-mini":0.15,
+            "gpt-4.1-mini":0.4,
+            "gpt-5-mini": 0.25,
+            "gpt-5-nano": 0.05,
+            "gemini-2.5-flash":0.3,
+            "gemini-2.0-flash":0.1,
+            "gemini-2.5-flash-lite":0.1
+        }
+
+        out_prices = {
+            "gpt-4o-mini":0.6,
+            "gpt-4.1-mini":1.6,
+            "gpt-5-mini": 2,
+            "gpt-5-nano": 0.4,
+            "gemini-2.5-flash":2.5,
+            "gemini-2.0-flash":0.7,
+            "gemini-2.5-flash-lite":0.4
+        }
+
+        i_price = input_prices.get(model, 0)
+        o_price= out_prices.get(model, 0)
+
+        return i_price, o_price
+
+    # ====================================================
     def identify_users(self, users_df, identifier_column):
         """
         specifying the users for identification

Messaging_system/PromptGenerator.py

@@ -88,7 +88,7 @@ Here is the information about the user:
         """
         input_context = self.input_context()
         instructions = self.message_type_instructions()
-        # user_info = self.get_user_profile(user=user)
+        user_info = self.get_user_profile(user=user)
 
         recommendation_instructions = self.recommendations_instructions(user)
 
@@ -101,6 +101,8 @@ Here is the information about the user:
 
 {instructions}
 
+{user_info}
+
 {recommendation_instructions}
 
 {example_output}

Messaging_system/PromptGenerator_2.py

@@ -117,10 +117,14 @@ Here is the information about the user:
 Your task is to write a 'header' and a 'message' as a push notification for a {self.Core.get_instrument()} student that sounds like everyday natural speech: friendly, short, no jargon by following the instructions.
 The output should sound like something that a {self.Core.get_instrument()} instructor would realistically say to a student in a daily conversation.
 
+        """
+
+        if self.Core.brand_voice is not None:
+            context += f"""
 ** Examples of actual phrases an instructor might say:**
 
 {self.Core.brand_voice}
-        """
+            """
 
         return context
 

messaging_main_test.py

@@ -161,7 +161,7 @@ if __name__ == "__main__":
     # o3-mini o1-mini o4-mini o1
 
     users_message = permes.create_personalize_messages(session=session,
-                                                model="gemini-2.5-flash-lite-preview-06-17",
+                                                model="gemini-2.5-flash",
                                                 users=users,
                                                 brand=brand,
                                                 message_style = instructions,