Merge pull request #59 from seanpedrick-case/dev

Added Gemma 12b and Ministral 3 14b support. Minor bug fixes

Author: seanpedrick-case

app.py

@@ -57,14 +57,12 @@
 }
 """
 
-# /* Completely hide elements with the class 'hidden_component' from view */
-# .hidden_component {
-#     display: none !important;
-#     visibility: hidden !important;
-# }
-
 # Create the gradio interface
-app = gr.Blocks(fill_width=True)
+app = gr.Blocks(fill_width=True,
+    analytics_enabled=False,
+    title="LLM topic modelling",
+    delete_cache=(43200, 43200),
+    )
 
 with app:
 

tools/aws_functions.py

@@ -35,13 +35,13 @@ def connect_to_bedrock_runtime(model_name_map:dict, model_choice:str, aws_access
             print(out_message)
             raise Exception(out_message)
     else: 
-        bedrock_runtime = list()
+        bedrock_runtime = None
 
     return bedrock_runtime
 
 def connect_to_s3_client(aws_access_key_textbox:str="", aws_secret_key_textbox:str=""):
     # If running an anthropic model, assume that running an AWS s3 model, load in s3
-    s3_client = list()
+    s3_client = None
 
     if aws_access_key_textbox and aws_secret_key_textbox:
         print("Connecting to s3 using AWS access key and secret keys from user input.")

tools/combine_sheets_into_xlsx.py

@@ -414,7 +414,7 @@ def collect_output_csvs_and_create_excel_output(in_data_files:List, chosen_cols:
     intro_text = [
         "This workbook contains outputs from the large language model topic analysis of open text data. Each sheet corresponds to a different CSV report included in the analysis.",
         f"The file analysed was {short_file_name}, the column analysed was '{chosen_cols}' and the data was grouped by column '{group}'."
-        "Please contact the LLM Topic Modelling app administrator if you need any explanation on how to use the results."
+        " Please contact the LLM Topic Modelling app administrator if you need any explanation on how to use the results."
         "Large language models are not 100% accurate and may produce biased or harmful outputs. All outputs from this analysis **need to be checked by a human** to check for harmful outputs, false information, and bias."
     ]
 

tools/config.py

@@ -154,6 +154,7 @@ def convert_string_to_boolean(value: str) -> bool:
 ACCESS_LOGS_FOLDER = get_or_create_env_var('ACCESS_LOGS_FOLDER', 'logs/')
 USAGE_LOGS_FOLDER = get_or_create_env_var('USAGE_LOGS_FOLDER', 'usage/')
 
+# Initialize full_log_subfolder based on USE_LOG_SUBFOLDERS setting
 if USE_LOG_SUBFOLDERS == "True":
     day_log_subfolder = today_rev + '/'
     host_name_subfolder = HOST_NAME + '/'
@@ -162,6 +163,8 @@ def convert_string_to_boolean(value: str) -> bool:
     FEEDBACK_LOGS_FOLDER = FEEDBACK_LOGS_FOLDER + full_log_subfolder
     ACCESS_LOGS_FOLDER = ACCESS_LOGS_FOLDER + full_log_subfolder
     USAGE_LOGS_FOLDER = USAGE_LOGS_FOLDER + full_log_subfolder
+else:
+    full_log_subfolder = ''  # Empty string when subfolders are not used
 
 S3_FEEDBACK_LOGS_FOLDER = get_or_create_env_var('S3_FEEDBACK_LOGS_FOLDER', 'feedback/' + full_log_subfolder)
 S3_ACCESS_LOGS_FOLDER = get_or_create_env_var('S3_ACCESS_LOGS_FOLDER', 'logs/' + full_log_subfolder)
@@ -261,17 +264,16 @@ def convert_string_to_boolean(value: str) -> bool:
 model_source = list()
 
 CHOSEN_LOCAL_MODEL_TYPE = get_or_create_env_var("CHOSEN_LOCAL_MODEL_TYPE", "Qwen 3 4B") # Gemma 3 1B #  "Gemma 2b" # "Gemma 3 4B"
-print("CHOSEN_LOCAL_MODEL_TYPE:", CHOSEN_LOCAL_MODEL_TYPE)
 
 if RUN_LOCAL_MODEL == "1" and CHOSEN_LOCAL_MODEL_TYPE:
     model_full_names.append(CHOSEN_LOCAL_MODEL_TYPE)
     model_short_names.append(CHOSEN_LOCAL_MODEL_TYPE)
     model_source.append("Local")
 
 if RUN_AWS_BEDROCK_MODELS == "1":
-    amazon_models = ["anthropic.claude-3-haiku-20240307-v1:0", "anthropic.claude-3-7-sonnet-20250219-v1:0", "anthropic.claude-sonnet-4-5-20250929-v1:0", "amazon.nova-micro-v1:0", "amazon.nova-lite-v1:0", "amazon.nova-pro-v1:0",  "deepseek.v3-v1:0", "openai.gpt-oss-20b-1:0", "openai.gpt-oss-120b-1:0"]
+    amazon_models = ["anthropic.claude-3-haiku-20240307-v1:0", "anthropic.claude-3-7-sonnet-20250219-v1:0", "anthropic.claude-sonnet-4-5-20250929-v1:0", "amazon.nova-micro-v1:0", "amazon.nova-lite-v1:0", "amazon.nova-pro-v1:0",  "deepseek.v3-v1:0", "openai.gpt-oss-20b-1:0", "openai.gpt-oss-120b-1:0", "google.gemma-3-12b-it", "mistral.ministral-3-14b-instruct"]
     model_full_names.extend(amazon_models)
-    model_short_names.extend(["haiku", "sonnet_3_7", "sonnet_4_5", "nova_micro", "nova_lite", "nova_pro", "deepseek_v3", "gpt_oss_20b_aws", "gpt_oss_120b_aws"])
+    model_short_names.extend(["haiku", "sonnet_3_7", "sonnet_4_5", "nova_micro", "nova_lite", "nova_pro", "deepseek_v3", "gpt_oss_20b_aws", "gpt_oss_120b_aws", "gemma_3_12b_it", "ministral_3_14b_instruct"])
     model_source.extend(["AWS"] * len(amazon_models))
 
 if RUN_GEMINI_MODELS == "1":
@@ -438,9 +440,6 @@ def update_model_choice_config(default_model_source, model_name_map):
     LOCAL_MODEL_FILE = ""
     LOCAL_MODEL_FOLDER = ""
 
-print("LOCAL_REPO_ID:", LOCAL_REPO_ID)
-
-
 USE_SPECULATIVE_DECODING = get_or_create_env_var("USE_SPECULATIVE_DECODING", "False")
 
 ASSISTANT_MODEL = get_or_create_env_var("ASSISTANT_MODEL", "")

tools/llm_funcs.py

@@ -882,7 +882,7 @@ def construct_azure_client(in_api_key: str, endpoint: str) -> Tuple[object, dict
         print("Error constructing Azure/OpenAI client:", e)
         raise
 
-def call_aws_claude(prompt: str, system_prompt: str, temperature: float, max_tokens: int, model_choice:str, bedrock_runtime:boto3.Session.client, assistant_prefill:str="") -> ResponseObject:
+def call_aws_bedrock(prompt: str, system_prompt: str, temperature: float, max_tokens: int, model_choice:str, bedrock_runtime:boto3.Session.client, assistant_prefill:str="") -> ResponseObject:
     """
     This function sends a request to AWS Claude with the following parameters:
     - prompt: The user's input prompt to be processed by the model.
@@ -902,27 +902,31 @@ def call_aws_claude(prompt: str, system_prompt: str, temperature: float, max_tok
         "temperature":temperature,
     }
 
-    if not assistant_prefill:
+    
+    # Using an assistant prefill only works for Anthropic models.
+    if assistant_prefill and "anthropic" in model_choice:
+        assistant_prefill_added = True
         messages =  [
                 {
                     "role": "user",
                     "content": [
                         {"text": prompt},
                     ],
+                },
+                {
+                    "role": "assistant",
+                    # Pre-filling with '|'
+                    "content": [{"text": assistant_prefill}]
                 }
             ]
     else:
+        assistant_prefill_added = False
         messages =  [
                 {
                     "role": "user",
                     "content": [
                         {"text": prompt},
                     ],
-                },
-                {
-                    "role": "assistant",
-                    # Pre-filling with '|'
-                    "content": [{"text": assistant_prefill}]
                 }
             ]
 
@@ -932,7 +936,7 @@ def call_aws_claude(prompt: str, system_prompt: str, temperature: float, max_tok
         }
     ]
 
-    # The converse API call itself. Note I've renamed the response variable for clarity.
+    # The converse API call.
     api_response = bedrock_runtime.converse(
         modelId=model_choice,
         messages=messages,
@@ -947,9 +951,15 @@ def call_aws_claude(prompt: str, system_prompt: str, temperature: float, max_tok
         reasoning_text = output_message['content'][0]['reasoningContent']['reasoningText']['text']
 
         # Extract the output text
-        text = assistant_prefill + output_message['content'][1]['text']
+        if assistant_prefill_added:
+            text = assistant_prefill + output_message['content'][1]['text']
+        else:
+            text = output_message['content'][1]['text']
     else:
-        text = assistant_prefill + output_message['content'][0]['text']
+        if assistant_prefill_added:
+            text = assistant_prefill + output_message['content'][0]['text']
+        else:
+            text = output_message['content'][0]['text']
 
     # The usage statistics are neatly provided in the 'usage' key.
     usage = api_response['usage']
@@ -1183,14 +1193,14 @@ def send_request(prompt: str, conversation_history: List[dict], client: ai.Clien
     elif "AWS" in model_source:
         for i in progress_bar:
             try:
-                print("Calling AWS Claude model, attempt", i + 1)
-                response = call_aws_claude(prompt, system_prompt, temperature, max_tokens, model_choice, bedrock_runtime=bedrock_runtime, assistant_prefill=assistant_prefill)
+                print("Calling AWS Bedrock model, attempt", i + 1)
+                response = call_aws_bedrock(prompt, system_prompt, temperature, max_tokens, model_choice, bedrock_runtime=bedrock_runtime, assistant_prefill=assistant_prefill)
 
                 #print("Successful call to Claude model.")
                 break
             except Exception as e:
                 # If fails, try again after X seconds in case there is a throttle limit
-                print("Call to Claude model failed:", e, " Waiting for ", str(timeout_wait), "seconds and trying again.")
+                print("Call to Bedrock model failed:", e, " Waiting for ", str(timeout_wait), "seconds and trying again.")
                 time.sleep(timeout_wait)
 
             if i == number_of_api_retry_attempts: