Update app.py

app.py

@@ -6,6 +6,7 @@ from huggingface_hub import InferenceClient
 import os
 
 
+
 st.set_page_config(layout="wide")
 
 
@@ -16,7 +17,7 @@ my_initial_rag_text = f"""This is a RAG (Retrieval-Augmented Generation) chatbot
 - Uses Streamlit for the web interface
 - Employs SentenceTransformer for generating embeddings
 - Uses HuggingFace's InferenceClient for LLM interaction
-- Has a default text about a training module on LLMs
+- Has a default text about a GRNET training module on LLMs
 
 2. State Management:
 - Maintains several session state variables for:
@@ -128,13 +129,36 @@ By using the Software, you agree to the terms and conditions of the disclaimer."
     # Add a text area for RAG text input
     st.text_area(label="Please enter your RAG text here:", value=my_initial_rag_text, height=500, key="my_rag_text", on_change=delete_chat_messages)
 
+
+
+
+# Define parameters
+min_window_size = 5 # number of sentences per chunk
+max_window_size = 10 # number of sentences per chunk
+print(100*"-")
 # Check if the sentences are not already in the session state
 if "my_sentences" not in st.session_state:
-    my_sentences_split = st.session_state["my_rag_text"].split("\n")
-    st.session_state["my_sentences"] = []
-    for my_sentence in my_sentences_split:
-        if my_sentence.strip():
-            st.session_state["my_sentences"].append(my_sentence.strip())
+    text = st.session_state["my_rag_text"]
+
+    # Split only on .?!;: followed by space OR on \n
+    pattern = r'(?<=[.?!;:])\s+|\n'
+    sentence_split = re.split(pattern, text)
+
+    sentences = [s.strip() for s in sentence_split if s.strip()]
+
+    # Rolling window: include partial windows at end
+    chunks = []
+    for rolling_window_size in range(min_window_size, max_window_size+1):
+        for i in range(0, len(sentences)-rolling_window_size+1):
+            chunk = " ".join(sentences[i:i+rolling_window_size]).strip()
+            if chunk:   
+                chunks.append(chunk)
+                print(f"*****{chunk}*****\n")
+
+    st.session_state["my_sentences"] = chunks
+    print(len(st.session_state["my_sentences"]))
+
+
 
 # Check if the embeddings are not already in the session state
 if "my_embeddings" not in st.session_state:
@@ -166,19 +190,21 @@ with column_2:
         similarity_to_question = cosine_similarity(my_question_embedding, st.session_state.my_embeddings).flatten()
         
         # Number of sentences to keep based on similarity
-        nof_keep_sentences = 10
+        nof_keep_sentences = 3
         
         # Get the indices of the top similar sentences
-        sorted_indices = similarity_to_question.argsort()[::-1][:nof_keep_sentences]
+        sorted_indices = similarity_to_question.argsort()[::-1][:nof_keep_sentences][::-1]
         
         # Retrieve the top similar sentences
-        sorted_sentences = [st.session_state.my_sentences[i] for i in sorted_indices]
+        sorted_sentences = ["Importance: " + str(round(100*similarity_to_question[i])) + f"% {5*'>'} " + st.session_state.my_sentences[i] for i in sorted_indices]
         
         # Construct the augmented prompt with the similar sentences
-        augmented_prompt = "Here is the context:"
+        augmented_prompt = "Here is my context:"
         for sentence in sorted_sentences:
             augmented_prompt += "\n\n" + 20*"-" + f"\n\n{sentence}"
-        augmented_prompt += "\n\n" + 20*"-" + "\n\n" + "The user said:" + f"\n\n{prompt}"
+        augmented_prompt += "\n\n" + 20*"-" + "\n\n" + "use the context I gave you to reply on the following:"
+        augmented_prompt += "\n\n" + f"\n\n{prompt}"
+        
 
         # Display the user's prompt in the chat container with a specific avatar
         messages_container.chat_message("user", avatar=":material/psychology_alt:").markdown(prompt) 
@@ -204,8 +230,12 @@ with column_2:
 
     
     # Display the chat messages history
-    st.write("Messages History:")
+    st.write("Messages History All:")
     st.json(st.session_state["my_chat_messages"], expanded=False)
     # Display the augmented prompt used for generating the response
     st.write("Augmented prompt:")
     st.json({"augmented_prompt": augmented_prompt}, expanded=False)
+
+    
+
+