Update chat history in prompts, segregate the prompts, add retry to HF API call, and update configs

chat_app.py

@@ -10,8 +10,9 @@ import streamlit as st
 from langchain_community.chat_message_histories import (
     StreamlitChatMessageHistory
 )
+from langchain_core.messages import HumanMessage
 from langchain_core.prompts import ChatPromptTemplate
-from langchain_core.runnables.history import RunnableWithMessageHistory
+from transformers import AutoTokenizer
 
 from global_config import GlobalConfig
 from helpers import llm_helper, pptx_helper
@@ -28,10 +29,44 @@ def _load_strings() -> dict:
         return json5.loads(in_file.read())
 
 
+@st.cache_data
+def _get_prompt_template(is_refinement: bool) -> str:
+    """
+    Return a prompt template.
+
+    :param is_refinement: Whether this is the initial or refinement prompt.
+    :return: The prompt template as f-string.
+    """
+
+    if is_refinement:
+        with open(GlobalConfig.REFINEMENT_PROMPT_TEMPLATE, 'r', encoding='utf-8') as in_file:
+            template = in_file.read()
+    else:
+        with open(GlobalConfig.INITIAL_PROMPT_TEMPLATE, 'r', encoding='utf-8') as in_file:
+            template = in_file.read()
+
+    return template
+
+
+@st.cache_resource
+def _get_tokenizer() -> AutoTokenizer:
+    """
+    Get Mistral tokenizer for counting tokens.
+
+    :return: The tokenizer.
+    """
+
+    return AutoTokenizer.from_pretrained(
+        pretrained_model_name_or_path=GlobalConfig.HF_LLM_MODEL_NAME
+    )
+
+
 APP_TEXT = _load_strings()
+
+# Session variables
+CHAT_MESSAGES = 'chat_messages'
 DOWNLOAD_FILE_KEY = 'download_file_name'
-# langchain.debug = True
-# langchain.verbose = True
+IS_IT_REFINEMENT = 'is_it_refinement'
 
 logger = logging.getLogger(__name__)
 progress_bar = st.progress(0, text='Setting up SlideDeck AI...')
@@ -53,10 +88,9 @@ def display_page_header_content():
 
     st.title(APP_TEXT['app_name'])
     st.subheader(APP_TEXT['caption'])
-    st.markdown(
-        'Powered by'
-        ' [Mistral-7B-Instruct-v0.2](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2)'
-    )
+    # st.markdown(
+    #     '![Visitors](https://api.visitorbadge.io/api/visitors?path=https%3A%2F%2Fhuggingface.co%2Fspaces%2Fbarunsaha%2Fslide-deck-ai&countColor=%23263759)'  # noqa: E501
+    # )
 
 
 def display_page_footer_content():
@@ -65,9 +99,6 @@ def display_page_footer_content():
     """
 
     st.text(APP_TEXT['tos'] + '\n\n' + APP_TEXT['tos2'])
-    # st.markdown(
-    #     '![Visitors](https://api.visitorbadge.io/api/visitors?path=https%3A%2F%2Fhuggingface.co%2Fspaces%2Fbarunsaha%2Fslide-deck-ai&countColor=%23263759)'  # noqa: E501
-    # )
 
 
 def build_ui():
@@ -89,60 +120,117 @@ def set_up_chat_ui():
     Prepare the chat interface and related functionality.
     """
 
-    history = StreamlitChatMessageHistory(key='chat_messages')
-    llm = llm_helper.get_hf_endpoint()
-
-    with open(GlobalConfig.CHAT_TEMPLATE_FILE, 'r', encoding='utf-8') as in_file:
-        template = in_file.read()
-
-    prompt = ChatPromptTemplate.from_template(template)
-    chain = prompt | llm
-
-    chain_with_history = RunnableWithMessageHistory(
-        chain,
-        lambda session_id: history,  # Always return the instance created earlier
-        input_messages_key='question',
-        history_messages_key='chat_history',
-    )
-
     with st.expander('Usage Instructions'):
         st.write(GlobalConfig.CHAT_USAGE_INSTRUCTIONS)
+        st.markdown(
+            'SlideDeck AI is powered by'
+            ' [Mistral-7B-Instruct-v0.2](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2)'
+        )
+
+    # view_messages = st.expander('View the messages in the session state')
 
     st.chat_message('ai').write(
         random.choice(APP_TEXT['ai_greetings'])
     )
+    progress_bar.progress(100, text='Done!')
+    progress_bar.empty()
 
-    for msg in history.messages:
-        # st.chat_message(msg.type).markdown(msg.content)
-        st.chat_message(msg.type).code(msg.content, language='json')
+    history = StreamlitChatMessageHistory(key=CHAT_MESSAGES)
 
-    # The download button disappears on clicking (anywhere) because of app reload
-    # So, display it again
-    if DOWNLOAD_FILE_KEY in st.session_state:
-        _display_download_button(st.session_state[DOWNLOAD_FILE_KEY])
+    if _is_it_refinement():
+        template = _get_prompt_template(is_refinement=True)
+        logger.debug('Getting refinement template')
+    else:
+        template = _get_prompt_template(is_refinement=False)
+        logger.debug('Getting initial template')
 
-    progress_bar.progress(100, text='Done!')
-    progress_bar.empty()
+    prompt_template = ChatPromptTemplate.from_template(template)
+
+    # Since Streamlit app reloads at every interaction, display the chat history
+    # from the save session state
+    for msg in history.messages:
+        msg_type = msg.type
+        if msg_type == 'user':
+            st.chat_message(msg_type).write(msg.content)
+        else:
+            st.chat_message(msg_type).code(msg.content, language='json')
 
     if prompt := st.chat_input(
         placeholder=APP_TEXT['chat_placeholder'],
         max_chars=GlobalConfig.LLM_MODEL_MAX_INPUT_LENGTH
     ):
+
+        progress_bar_pptx = st.progress(0, 'Preparing to run...')
+        if not _is_valid_prompt(prompt):
+            return
+
         logger.info('User input: %s', prompt)
         st.chat_message('user').write(prompt)
 
-        progress_bar_pptx = st.progress(0, 'Calling LLM...')
-
-        # As usual, new messages are added to StreamlitChatMessageHistory when the Chain is called
-        config = {'configurable': {'session_id': 'any'}}
-        response: str = chain_with_history.invoke({'question': prompt}, config)
-        st.chat_message('ai').markdown('```json\n' + response)
+        user_messages = _get_user_messages()
+        user_messages.append(prompt)
+        list_of_msgs = [
+            f'{idx + 1}. {msg}' for idx, msg in enumerate(user_messages)
+        ]
+        list_of_msgs = '\n'.join(list_of_msgs)
+
+        if _is_it_refinement():
+            formatted_template = prompt_template.format(
+                **{
+                    'instructions': list_of_msgs,
+                    'previous_content': _get_last_response()
+                }
+            )
+        else:
+            formatted_template = prompt_template.format(
+                **{
+                    'question': prompt,
+                }
+            )
+
+        progress_bar_pptx.progress(5, 'Calling LLM...will retry if connection times out...')
+        response: dict = llm_helper.hf_api_query({
+            'inputs': formatted_template,
+            'parameters': {
+                'temperature': GlobalConfig.LLM_MODEL_TEMPERATURE,
+                'min_length': GlobalConfig.LLM_MODEL_MIN_OUTPUT_LENGTH,
+                'max_length': GlobalConfig.LLM_MODEL_MAX_OUTPUT_LENGTH,
+                'max_new_tokens': GlobalConfig.LLM_MODEL_MAX_OUTPUT_LENGTH,
+                'num_return_sequences': 1,
+                'return_full_text': False,
+                # "repetition_penalty": 0.0001
+            },
+            'options': {
+                'wait_for_model': True,
+
+                'use_cache': True
+            }
+        })
+
+        if len(response) > 0 and 'generated_text' in response[0]:
+            response: str = response[0]['generated_text'].strip()
+
+        st.chat_message('ai').code(response, language='json')
+
+        history.add_user_message(prompt)
+        history.add_ai_message(response)
+
+        if GlobalConfig.COUNT_TOKENS:
+            tokenizer = _get_tokenizer()
+            tokens_count_in = len(tokenizer.tokenize(formatted_template))
+            tokens_count_out = len(tokenizer.tokenize(response))
+            logger.debug(
+                'Tokens count:: input: %d, output: %d',
+                tokens_count_in, tokens_count_out
+            )
+
+        # _display_messages_history(view_messages)
 
         # The content has been generated as JSON
         # There maybe trailing ``` at the end of the response -- remove them
         # To be careful: ``` may be part of the content as well when code is generated
-        response_cleaned = _clean_json(response)
         progress_bar_pptx.progress(50, 'Analyzing response...')
+        response_cleaned = _clean_json(response)
 
         # Now create the PPT file
         progress_bar_pptx.progress(75, 'Creating the slide deck...give it a moment')
@@ -185,6 +273,74 @@ def generate_slide_deck(json_str: str):
         logger.error('Caught a generic exception: %s', str(ex))
 
 
+def _is_valid_prompt(prompt: str) -> bool:
+    """
+    Verify whether user input satisfies the concerned constraints.
+
+    :param prompt: The user input text.
+    :return: True if all criteria are satisfied; False otherwise.
+    """
+
+    if len(prompt) < 5 or ' ' not in prompt:
+        st.error(
+            'Not enough information provided!'
+            ' Please be a little more descriptive and type a few words with a few characters :)'
+        )
+        return False
+
+    return True
+
+
+def _is_it_refinement() -> bool:
+    """
+    Whether it is the initial prompt or a refinement.
+
+    :return: True if it is the initial prompt; False otherwise.
+    """
+
+    if IS_IT_REFINEMENT in st.session_state:
+        return True
+
+    if len(st.session_state[CHAT_MESSAGES]) >= 2:
+        # Prepare for the next call
+        st.session_state[IS_IT_REFINEMENT] = True
+        return True
+
+    return False
+
+
+def _get_user_messages() -> List[str]:
+    """
+    Get a list of user messages submitted until now from the session state.
+
+    :return: The list of user messages.
+    """
+
+    return [
+        msg.content for msg in st.session_state[CHAT_MESSAGES] if isinstance(msg, HumanMessage)
+    ]
+
+
+def _get_last_response() -> str:
+    """
+    Get the last response generated by AI.
+
+    :return: The response text.
+    """
+
+    return st.session_state[CHAT_MESSAGES][-1].content
+
+
+def _display_messages_history(view_messages: st.expander):
+    """
+    Display the history of messages.
+
+    :param view_messages: The list of AI and Human messages.
+    """
+
+    with view_messages:
+        view_messages.json(st.session_state[CHAT_MESSAGES])
+
 def _clean_json(json_str: str) -> str:
     """
     Attempt to clean a JSON response string from the LLM by removing the trailing ```

global_config.py

@@ -13,18 +13,20 @@ class GlobalConfig:
     HF_LLM_MODEL_NAME = 'mistralai/Mistral-7B-Instruct-v0.2'
     LLM_MODEL_TEMPERATURE: float = 0.2
     LLM_MODEL_MIN_OUTPUT_LENGTH: int = 50
-    LLM_MODEL_MAX_OUTPUT_LENGTH: int = 2000
-    LLM_MODEL_MAX_INPUT_LENGTH: int = 100
+    LLM_MODEL_MAX_OUTPUT_LENGTH: int = 4096
+    LLM_MODEL_MAX_INPUT_LENGTH: int = 750
 
     HUGGINGFACEHUB_API_TOKEN = os.environ.get('HUGGINGFACEHUB_API_TOKEN', '')
     METAPHOR_API_KEY = os.environ.get('METAPHOR_API_KEY', '')
 
     LOG_LEVEL = 'DEBUG'
+    COUNT_TOKENS = False
     APP_STRINGS_FILE = 'strings.json'
     PRELOAD_DATA_FILE = 'examples/example_02.json'
     SLIDES_TEMPLATE_FILE = 'langchain_templates/template_combined.txt'
     JSON_TEMPLATE_FILE = 'langchain_templates/text_to_json_template_02.txt'
-    CHAT_TEMPLATE_FILE = 'langchain_templates/template_combined_chat_history.txt'
+    INITIAL_PROMPT_TEMPLATE = 'langchain_templates/chat_prompts/initial_template.txt'
+    REFINEMENT_PROMPT_TEMPLATE = 'langchain_templates/chat_prompts/refinement_template.txt'
 
     PPTX_TEMPLATE_FILES = {
         'Blank': {

helpers/llm_helper.py

@@ -1,5 +1,8 @@
 import logging
 import requests
+from requests.adapters import HTTPAdapter
+from urllib3.util import Retry
+
 from langchain_community.llms.huggingface_endpoint import HuggingFaceEndpoint
 from langchain_core.language_models import LLM
 
@@ -11,10 +14,22 @@ HF_API_HEADERS = {"Authorization": f"Bearer {GlobalConfig.HUGGINGFACEHUB_API_TOK
 
 logger = logging.getLogger(__name__)
 
+retries = Retry(
+    total=5,
+    backoff_factor=0.25,
+    backoff_jitter=0.3,
+    status_forcelist=[502, 503, 504],
+    allowed_methods={'POST'},
+)
+adapter = HTTPAdapter(max_retries=retries)
+http_session = requests.Session()
+http_session.mount('https://', adapter)
+http_session.mount('http://', adapter)
+
 
 def get_hf_endpoint() -> LLM:
     """
-    Get an LLM via the HuggingFaceEndpoint.
+    Get an LLM via the HuggingFaceEndpoint of LangChain.
 
     :return: The LLM.
     """
@@ -44,11 +59,11 @@ def hf_api_query(payload: dict) -> dict:
     """
 
     try:
-        response = requests.post(HF_API_URL, headers=HF_API_HEADERS, json=payload, timeout=15)
+        response = http_session.post(HF_API_URL, headers=HF_API_HEADERS, json=payload, timeout=15)
         result = response.json()
     except requests.exceptions.Timeout as te:
         logger.error('*** Error: hf_api_query timeout! %s', str(te))
-        result = {}
+        result = []
 
     return result
 

helpers/pptx_helper.py

@@ -64,7 +64,7 @@ def generate_powerpoint_presentation(
     parsed_data = json5.loads(structured_data)
 
     logger.debug(
-        "*** Using PPTX template: %s",
+        '*** Using PPTX template: %s',
         GlobalConfig.PPTX_TEMPLATE_FILES[slides_template]['file']
     )
     presentation = pptx.Presentation(GlobalConfig.PPTX_TEMPLATE_FILES[slides_template]['file'])
@@ -75,7 +75,10 @@ def generate_powerpoint_presentation(
     title = slide.shapes.title
     subtitle = slide.placeholders[1]
     title.text = parsed_data['title']
-    logger.debug('Presentation title is: %s', title.text)
+    logger.info(
+        'PPT title: %s | #slides: %d',
+        title.text, len(parsed_data['slides'])
+    )
     subtitle.text = 'by Myself and SlideDeck AI :)'
     all_headers = [title.text, ]
 

langchain_templates/{template_combined_chat_history.txt → chat_prompts/initial_template.txt}

@@ -1,21 +1,16 @@
-You are a helpful, intelligent chatbot. Follow the instructions and chat history, if available, to create or revise the slides for a presentation on the given topic.
+You are a helpful, intelligent chatbot. Create the slides for a presentation on the given topic.
 Include main headings for each slide, detailed bullet points for each slide.
 Add relevant content to each slide.
-The content should be descriptive, verbose, and detailed as much as possible.
+The content of each slide should be verbose, descriptive, and very detailed.
 If relevant, add one or two examples to illustrate the concept.
 Unless explicitly specified with the topic, create about 10 slides.
 
 
-### Instructions:
+### Topic:
 {question}
 
 
-### Chat history:
-{chat_history}
-
-
-The output should only be JSON and nothing else.
-The desired JSON output format:
+The output must be valid and syntactically correct JSON adhering to the following schema:
 {{
     "title": "Presentation Title",
     "slides": [
@@ -33,7 +28,7 @@ The desired JSON output format:
         {{
             "heading": "Heading for the Second Slide",
             "bullet_points": [
-                "First bullet point",             
+                "First bullet point",
                 "Second bullet item",
                 "Third bullet point"
             ]

langchain_templates/chat_prompts/refinement_template.txt

@@ -0,0 +1,49 @@
+You are a helpful, intelligent chatbot. You follow instructions to refine an existing slide deck.
+A list of user instructions is provided below in sequential order -- from the oldest to the latest.
+The previously generated content of the slide deck in JSON format is also provided.
+Follow the instructions to revise the content of the previously generated slides of the presentation on the given topic.
+Include main headings for each slide, detailed bullet points for each slide.
+Add relevant content to each slide.
+The content of the slides should be descriptive, verbose, and detailed.
+If relevant, add one or two examples to illustrate the concept.
+Unless explicitly specified with the topic, create about 10 slides.
+You also fix any syntax error that may be present in the JSON-formatted content.
+
+
+### List of instructions:
+{instructions}
+
+
+### Previously generated slide deck content as JSON:
+{previous_content}
+
+
+The output must be valid and syntactically correct JSON adhering to the following schema:
+{{
+    "title": "Presentation Title",
+    "slides": [
+        {{
+            "heading": "Heading for the First Slide",
+            "bullet_points": [
+                "First bullet point",
+                [
+                    "Sub-bullet point 1",
+                    "Sub-bullet point 2"
+                ],
+                "Second bullet point"
+            ]
+        }},
+        {{
+            "heading": "Heading for the Second Slide",
+            "bullet_points": [
+                "First bullet point",             
+                "Second bullet item",
+                "Third bullet point"
+            ]
+        }}
+    ]
+}}
+
+
+### Output:
+```json

requirements.txt

@@ -8,5 +8,7 @@ metaphor-python
 json5~=0.9.14
 requests~=2.31.0
 
-transformers
+transformers~=4.39.2
 langchain-community
+
+urllib3~=2.2.1