Add initial version of PDF-based PPT generation

.streamlit/config.toml

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

app.py

@@ -19,6 +19,7 @@ from dotenv import load_dotenv
 from langchain_community.chat_message_histories import StreamlitChatMessageHistory
 from langchain_core.messages import HumanMessage
 from langchain_core.prompts import ChatPromptTemplate
+from pypdf import PdfReader
 
 import global_config as gcfg
 from global_config import GlobalConfig
@@ -266,8 +267,17 @@ 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', ],
     ):
+        print(f'{prompt=}')
+        prompt_text = prompt.text or ''
+        if prompt['files']:
+            additional_text = get_pdf_contents(prompt['files'][0])
+        else:
+            additional_text = ''
+
         provider, llm_name = llm_helper.get_provider_model(
             llm_provider_to_use,
             use_ollama=RUN_IN_OFFLINE_MODE
@@ -279,20 +289,20 @@ 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 +310,16 @@ def set_up_chat_ui():
                 **{
                     'instructions': '\n'.join(list_of_msgs),
                     'previous_content': _get_last_response(),
+                    'additional_info': additional_text,
                 }
             )
         else:
-            formatted_template = prompt_template.format(**{'question': prompt})
+            formatted_template = prompt_template.format(
+                **{
+                    'question': prompt_text,
+                    'additional_info': additional_text,
+                }
+            )
 
         progress_bar = st.progress(0, 'Preparing to call LLM...')
         response = ''
@@ -392,7 +408,7 @@ 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
@@ -487,6 +503,30 @@ def generate_slide_deck(json_str: str) -> Union[pathlib.Path, None]:
     return path
 
 
+def get_pdf_contents(
+        pdf_file: st.runtime.uploaded_file_manager.UploadedFile,
+        max_pages: int = 20
+) -> 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.
+    """
+
+    print(f'{type(pdf_file)=}')
+    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
+
+
 def _is_it_refinement() -> bool:
     """
     Whether it is the initial prompt or a refinement.

langchain_templates/chat_prompts/initial_template_v4_two_cols_img.txt

@@ -5,6 +5,10 @@ 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.
+Rather than simply listing them line by line, try to understand these concepts and data provided and present them appropriately in the slides.
+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.
 
@@ -16,7 +20,7 @@ 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 (you're creating a presentation, not a report).
 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.
@@ -102,5 +106,10 @@ 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 +8,10 @@ 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.
+Rather than simply listing them line by line, try to understand these concepts and data provided and present them appropriately in the slides.
+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.
@@ -18,7 +22,7 @@ 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 (you're creating a presentation, not a report).
 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.
@@ -108,5 +112,10 @@ 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

@@ -15,11 +15,12 @@ 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