Merge pull request #93 from barun-saha/visual

.streamlit/config.toml

@@ -1,7 +1,7 @@
 [server]
 runOnSave = true
 headless = false
-maxUploadSize = 0
+maxUploadSize = 2
 
 [browser]
 gatherUsageStats = false

app.py

@@ -21,6 +21,7 @@ from langchain_core.messages import HumanMessage
 from langchain_core.prompts import ChatPromptTemplate
 
 import global_config as gcfg
+import helpers.file_manager as filem
 from global_config import GlobalConfig
 from helpers import llm_helper, pptx_helper, text_helper
 
@@ -141,6 +142,7 @@ APP_TEXT = _load_strings()
 CHAT_MESSAGES = 'chat_messages'
 DOWNLOAD_FILE_KEY = 'download_file_name'
 IS_IT_REFINEMENT = 'is_it_refinement'
+ADDITIONAL_INFO = 'additional_info'
 
 
 logger = logging.getLogger(__name__)
@@ -266,8 +268,17 @@ def set_up_chat_ui():
 
     if prompt := st.chat_input(
         placeholder=APP_TEXT['chat_placeholder'],
-        max_chars=GlobalConfig.LLM_MODEL_MAX_INPUT_LENGTH
+        max_chars=GlobalConfig.LLM_MODEL_MAX_INPUT_LENGTH,
+        accept_file=True,
+        file_type=['pdf', ],
     ):
+        prompt_text = prompt.text or ''
+        if prompt['files']:
+            # Apparently, Streamlit stores uploaded files in memory and clears on browser close
+            # https://docs.streamlit.io/knowledge-base/using-streamlit/where-file-uploader-store-when-deleted
+            st.session_state[ADDITIONAL_INFO] = filem.get_pdf_contents(prompt['files'][0])
+            print(f'{prompt["files"]=}')
+
         provider, llm_name = llm_helper.get_provider_model(
             llm_provider_to_use,
             use_ollama=RUN_IN_OFFLINE_MODE
@@ -279,20 +290,20 @@ def set_up_chat_ui():
         api_ver = api_version.strip()
 
         if not are_all_inputs_valid(
-                prompt, provider, llm_name, user_key,
+                prompt_text, provider, llm_name, user_key,
                 az_deployment, az_endpoint, api_ver
         ):
             return
 
         logger.info(
             'User input: %s | #characters: %d | LLM: %s',
-            prompt, len(prompt), llm_name
+            prompt_text, len(prompt_text), llm_name
         )
-        st.chat_message('user').write(prompt)
+        st.chat_message('user').write(prompt_text)
 
         if _is_it_refinement():
             user_messages = _get_user_messages()
-            user_messages.append(prompt)
+            user_messages.append(prompt_text)
             list_of_msgs = [
                 f'{idx + 1}. {msg}' for idx, msg in enumerate(user_messages)
             ]
@@ -300,10 +311,16 @@ def set_up_chat_ui():
                 **{
                     'instructions': '\n'.join(list_of_msgs),
                     'previous_content': _get_last_response(),
+                    'additional_info': st.session_state.get(ADDITIONAL_INFO, ''),
                 }
             )
         else:
-            formatted_template = prompt_template.format(**{'question': prompt})
+            formatted_template = prompt_template.format(
+                **{
+                    'question': prompt_text,
+                    'additional_info': st.session_state.get(ADDITIONAL_INFO, ''),
+                }
+            )
 
         progress_bar = st.progress(0, 'Preparing to call LLM...')
         response = ''
@@ -392,7 +409,7 @@ def set_up_chat_ui():
                 )
             return
 
-        history.add_user_message(prompt)
+        history.add_user_message(prompt_text)
         history.add_ai_message(response)
 
         # The content has been generated as JSON

global_config.py

@@ -86,6 +86,7 @@ class GlobalConfig:
     LLM_MODEL_TEMPERATURE = 0.2
     LLM_MODEL_MIN_OUTPUT_LENGTH = 100
     LLM_MODEL_MAX_INPUT_LENGTH = 400  # characters
+    MAX_PAGE_COUNT = 50
 
     LOG_LEVEL = 'DEBUG'
     COUNT_TOKENS = False
@@ -133,15 +134,23 @@ class GlobalConfig:
         '\n\n'
         'Finally, click on the download button at the bottom to download the slide deck.'
         ' See this [demo video](https://youtu.be/QvAKzNKtk9k) for a brief walkthrough.\n\n'
-        'Remember, the conversational interface is meant to (and will) update yor *initial*'
-        ' slide deck. If you want to create a new slide deck on a different topic,'
-        ' start a new chat session by reloading this page.\n\n'
-        'Currently, paid or *free-to-use* LLMs from five different providers are supported.'
-        ' If one is not available, choose the other from the dropdown list. A [summary of'
-        ' the supported LLMs]('
+        'Remember, the conversational interface is meant to (and will) update yor *initial*/'
+        '*previous* slide deck. If you want to create a new slide deck on a different topic,'
+        ' start a new chat session by reloading this page.'
+        '\n\nSlideDeck AI can algo generate a presentation based on a PDF file. You can upload'
+        ' a PDF file using the chat widget. Only a single file and up to max 50 pages will be'
+        ' considered. For PDF-based slide deck generation, LLMs with large context windows, such'
+        ' as Gemini, GPT, and Mistral-Nemo, are recommended. Note: images from the PDF files will'
+        ' not be used.'
+        '\n\nAlso, note that the uploaded file might disappear from the page after click.'
+        ' You do not need to upload the same file again to continue'
+        ' the interaction and refining—the contents of the PDF file will be retained in the'
+        ' same interactive session.'
+        '\n\nCurrently, paid or *free-to-use* LLMs from five different providers are supported.'
+        ' A [summary of the supported LLMs]('
         'https://github.com/barun-saha/slide-deck-ai/blob/main/README.md#summary-of-the-llms)'
-        ' is available for reference. SlideDeck AI does **NOT** store your API keys.\n\n'
-        ' SlideDeck AI does not have access to the Web, apart for searching for images relevant'
+        ' is available for reference. SlideDeck AI does **NOT** store your API keys.'
+        '\n\nSlideDeck AI does not have access to the Web, apart for searching for images relevant'
         ' to the slides. Photos are added probabilistically; transparency needs to be changed'
         ' manually, if required.\n\n'
         '[SlideDeck AI](https://github.com/barun-saha/slide-deck-ai) is an Open-Source project,'

helpers/file_manager.py

@@ -0,0 +1,40 @@
+"""
+File manager helper to work with uploaded files.
+"""
+import logging
+import os
+import sys
+
+import streamlit as st
+from pypdf import PdfReader
+
+sys.path.append('..')
+sys.path.append('../..')
+
+from global_config import GlobalConfig
+
+
+logger = logging.getLogger(__name__)
+
+
+def get_pdf_contents(
+        pdf_file: st.runtime.uploaded_file_manager.UploadedFile,
+        max_pages: int = GlobalConfig.MAX_PAGE_COUNT
+) -> str:
+    """
+    Extract the text contents from a PDF file.
+
+    :param pdf_file: The uploaded PDF file.
+    :param max_pages: The max no. of pages to extract contents from.
+    :return: The contents.
+    """
+
+    reader = PdfReader(pdf_file)
+    n_pages = min(max_pages, len(reader.pages))
+    text = ''
+
+    for page in range(n_pages):
+        page = reader.pages[page]
+        text += page.extract_text()
+
+    return text

langchain_templates/chat_prompts/initial_template_v4_two_cols_img.txt

@@ -5,6 +5,13 @@ Include main headings for each slide, detailed bullet points for each slide.
 Add relevant, detailed content to each slide. When relevant, add one or two EXAMPLES to illustrate the concept.
 For two or three important slides, generate the key message that those slides convey.
 
+The <ADDITIONAL_INFO> may provide additional information. If available, you should incorporate them while making the slides.
+Read this information carefully. Based on the contents provided, plan the structure of the presentation.
+For example, if it's a paper, you can consider having Problem, Solution, Experiments, and Results, among other sections.
+If it's a product brochure, you can have Features, Changes, Operating Conditions, and likewise relevant sections.
+Similarly, decide for other content types. Then appropriately incorporate the contents into the relevant slides, presenting in a useful way.
+If <ADDITIONAL_INFO> is empty, ignore it.
+
 Identify if a slide describes a step-by-step/sequential process, then begin the bullet points with a special marker >>.
 Limit this to max two or three slides.
 
@@ -16,11 +23,12 @@ In addition, create one slide containing 4 TO 6 icons (pictograms) illustrating
 In this slide, each line of text will begin with the name of a relevant icon enclosed between [[ and ]], e.g., [[machine-learning]] and [[fairness]].
 Insert icons only in this slide.
 
-Your output, i.e., the content of each slide should be VERBOSE, DESCRIPTIVE, and very DETAILED.
+Your output, i.e., the content of each slide should be vert detailed and descriptive but not way too verbose.
+Avoid writing like a report, but also avoid very short bullet points with just 3-4 words.
 Each bullet point should be detailed and explanatory, not just short phrases.
 
 ALWAYS add a concluding slide at the end, containing a list of the key takeaways and an optional call-to-action if relevant to the context.
-Unless explicitly instructed with the topic, create 10 TO 12 SLIDES in total. You must never create more tha 15 slides.
+Unless explicitly instructed with the topic, create 10 to 12 slides. You must never create more than 15 to 20 slides.
 
 
 ### Topic:
@@ -102,5 +110,10 @@ The output must be only a valid and syntactically correct JSON adhering to the f
 }}
 
 
+<ADDITIONAL_INFO>
+{additional_info}
+</ADDITIONAL_INFO>
+
+
 ### Output:
 ```json

langchain_templates/chat_prompts/refinement_template_v4_two_cols_img.txt

@@ -8,6 +8,13 @@ Include main headings for each slide, detailed bullet points for each slide.
 Add relevant, detailed content to each slide. When relevant, add one or two EXAMPLES to illustrate the concept.
 For two or three important slides, generate the key message that those slides convey.
 
+The <ADDITIONAL_INFO> may provide additional information. If available, you should incorporate them while making the slides.
+Read this information carefully. Based on the contents provided, plan the structure of the presentation.
+For example, if it's a paper, you can consider having Problem, Solution, Experiments, and Results, among other sections.
+If it's a product brochure, you can have Features, Changes, Operating Conditions, and likewise relevant sections.
+Similarly, decide for other content types. Then appropriately incorporate the contents into the relevant slides, presenting in a useful way.
+If <ADDITIONAL_INFO> is empty, ignore it.
+
 Identify if a slide describes a step-by-step/sequential process, then begin the bullet points with a special marker >>. Limit this to max two or three slides.
 Also, add at least one slide with a double column layout by generating appropriate content based on the description in the JSON schema provided below.
 In addition, for each slide, add image keywords based on the content of the respective slides.
@@ -18,11 +25,12 @@ In this slide, each line of text will begin with the name of a relevant icon enc
 Insert icons only in this slide.
 Do not repeat any icons or the icons slide.
 
-Your output, i.e., the content of each slide should be VERBOSE, DESCRIPTIVE, and very DETAILED.
+Your output, i.e., the content of each slide should be vert detailed and descriptive but not way too verbose.
+Avoid writing like a report, but also avoid very short bullet points with just 3-4 words.
 Each bullet point should be detailed and explanatory, not just short phrases.
 
 ALWAYS add a concluding slide at the end, containing a list of the key takeaways and an optional call-to-action if relevant to the context.
-Unless explicitly specified in the instructions below, create 10 TO 12 SLIDES in total. You must never create more tha 15 slides.
+Unless explicitly instructed with the topic, create 10 to 12 slides. You must never create more than 15 to 20 slides.
 
 
 ### List of instructions:
@@ -108,5 +116,10 @@ The output must be only a valid and syntactically correct JSON adhering to the f
 }}
 
 
+<ADDITIONAL_INFO>
+{additional_info}
+</ADDITIONAL_INFO>
+
+
 ### Output:
 ```json

requirements.txt

@@ -15,14 +15,15 @@ langchain-cohere==0.3.3
 langchain-together==0.3.0
 langchain-ollama==0.2.1
 langchain-openai==0.3.3
-streamlit~=1.38.0
+streamlit~=1.44.0
 
 python-pptx~=1.0.2
 json5~=0.9.14
 requests~=2.32.3
+pypdf~=5.4.0
 
 transformers>=4.48.0
-torch==2.4.0
+torch>=2.6.0
 
 lxml~=4.9.3
 tqdm~=4.66.5

strings.json

@@ -25,15 +25,16 @@
     "image_info": "Got some more minutes? We are also trying to deliver an AI-generated art on the presentation topic, fresh off the studio, just for you!",
     "content_generation_error": "Unfortunately, SlideDeck AI failed to generate any content for you! Please try again later.",
     "json_parsing_error": "Unfortunately, SlideDeck AI failed to parse the response from LLM! Please try again by rephrasing the query or refreshing the page.",
-    "tos": "SlideDeck AI is an experimental prototype, and it has its limitations.\nPlease carefully review any and all AI-generated content.",
+    "tos": "SlideDeck AI is an experimental prototype, and it has its limitations.\nAI-generated content may be incorrect. Please carefully review and verify the contents.",
     "tos2": "By using SlideDeck AI, you agree to fair and responsible usage.\nNo liability assumed by any party.",
     "ai_greetings": [
         "How may I help you today?",
         "Stuck with creating your presentation? Let me help you.",
         "Looks like you have a looming deadline. Can I help you get started with your slide deck?",
         "Hello! What topic do you have on your mind today?",
-        "Did you know that SlideDeck AI supports eight LLMs that are free to use and generate contents in different styles? Try them out from the dropdown list."
+        "Did you know that SlideDeck AI supports eight LLMs that generate contents in different styles?",
+        "Did you know that SlideDeck AI can create a presentation based on any uploaded PDF file?"
     ],
-    "chat_placeholder": "Write the topic or instructions here",
+    "chat_placeholder": "Write the topic or instructions here. You can also upload a PDF file.",
     "like_feedback": "If you like SlideDeck AI, please consider leaving a heart ❤\uFE0F on the [Hugging Face Space](https://huggingface.co/spaces/barunsaha/slide-deck-ai/) or a star ⭐ on [GitHub](https://github.com/barun-saha/slide-deck-ai). Your [feedback](https://forms.gle/JECFBGhjvSj7moBx9) is appreciated."
 }