RAG changes

pages/Multimodal_Conversational_Search.py

@@ -15,6 +15,7 @@ import random
 import string
 import rag_DocumentLoader 
 import rag_DocumentSearcher
+#import colpali
 import pandas as pd
 from PIL import Image 
 import shutil
@@ -22,10 +23,9 @@ import base64
 import time
 import botocore
 #from langchain.callbacks.base import BaseCallbackHandler
-#import streamlit_nested_layout
 #from IPython.display import clear_output, display, display_markdown, Markdown
 from requests_aws4auth import AWS4Auth
-#import copali
+import colpali
 from requests.auth import HTTPBasicAuth
 
 
@@ -41,8 +41,8 @@ AI_ICON = "images/opensearch-twitter-card.png"
 REGENERATE_ICON = "images/regenerate.png"
 s3_bucket_ = "pdf-repo-uploads"
             #"pdf-repo-uploads"
-polly_client = boto3.client('polly',aws_access_key_id=st.secrets['user_access_key'],
-                aws_secret_access_key=st.secrets['user_secret_key'], region_name = 'us-east-1')
+polly_client = boto3.Session(
+            region_name='us-east-1').client('polly')
 
 # Check if the user ID is already stored in the session state
 if 'user_id' in st.session_state:
@@ -69,6 +69,13 @@ if "chats" not in st.session_state:
 
 if "questions_" not in st.session_state:
     st.session_state.questions_ = []
+    
+    
+if "show_columns" not in st.session_state:
+    st.session_state.show_columns = False
+    
+if "answer_ready" not in st.session_state:
+    st.session_state.answer_ready = False
 
 if "answers_" not in st.session_state:
     st.session_state.answers_ = []
@@ -78,6 +85,9 @@ if "input_index" not in st.session_state:
     
 if "input_is_rerank" not in st.session_state:
     st.session_state.input_is_rerank = True
+    
+if "input_is_colpali" not in st.session_state:
+    st.session_state.input_is_colpali = False
 
 if "input_copali_rerank" not in st.session_state:
     st.session_state.input_copali_rerank = False
@@ -89,8 +99,8 @@ if "input_query" not in st.session_state:
     st.session_state.input_query="which city has the highest average housing price in UK ?"#"What is the projected energy percentage from renewable sources in future?"#"Which city in United Kingdom has the highest average housing price ?"#"How many aged above 85 years died due to covid ?"# What is the projected energy from renewable sources ?"
 
 
-# if "input_rag_searchType" not in st.session_state:
-#     st.session_state.input_rag_searchType = ["Sparse Search"]
+if "input_rag_searchType" not in st.session_state:
+    st.session_state.input_rag_searchType = ["Vector Search"]
     
 
 
@@ -129,7 +139,7 @@ st.markdown("""
 
 
 credentials = boto3.Session().get_credentials()
-awsauth = HTTPBasicAuth('prasadnu',st.secrets['rag_shopping_assistant_os_api_access'])
+awsauth = AWS4Auth(credentials.access_key, credentials.secret_key, 'us-west-2', service, session_token=credentials.token)
 service = 'es'
 
 
@@ -156,7 +166,7 @@ service = 'es'
 def write_logo():
     col1, col2, col3 = st.columns([5, 1, 5])
     with col2:
-        st.image(AI_ICON, use_container_width='always') 
+        st.image(AI_ICON, use_column_width='always') 
 
 def write_top_bar():
     col1, col2 = st.columns([77,23])
@@ -164,7 +174,7 @@ def write_top_bar():
         st.write("")
         st.header("Chat with your data",divider='rainbow')
         
-        #st.image(AI_ICON, use_container_width='always')
+        #st.image(AI_ICON, use_column_width='always')
     
     with col2:
         st.write("")
@@ -188,6 +198,8 @@ if clear:
 
 
 def handle_input():
+    # st.session_state.answer_ready = True
+    # st.session_state.show_columns = False  # reset column display
     print("Question: "+st.session_state.input_query)
     print("-----------")
     print("\n\n")
@@ -208,7 +220,11 @@ def handle_input():
         'id': len(st.session_state.questions_)
     }
     st.session_state.questions_.append(question_with_id)
-    out_ = rag_DocumentSearcher.query_(awsauth, inputs, st.session_state['session_id'],st.session_state.input_rag_searchType)
+    if(st.session_state.input_is_colpali):
+        out_ = colpali.colpali_search_rerank(st.session_state.input_query)
+        #print(out_)
+    else:
+        out_ = rag_DocumentSearcher.query_(awsauth, inputs, st.session_state['session_id'],st.session_state.input_rag_searchType)
     st.session_state.answers_.append({
         'answer': out_['text'],
         'source':out_['source'],
@@ -248,7 +264,7 @@ def write_user_message(md):
     col1, col2 = st.columns([3,97])
     
     with col1:
-        st.image(USER_ICON, use_container_width='always')
+        st.image(USER_ICON, use_column_width='always')
     with col2:
         #st.warning(md['question'])
 
@@ -261,7 +277,7 @@ def render_answer(question,answer,index,res_img):
     
     col1, col2, col_3 = st.columns([4,74,22])
     with col1:
-        st.image(AI_ICON, use_container_width='always')
+        st.image(AI_ICON, use_column_width='always')
     with col2:
         ans_ = answer['answer']
         st.write(ans_)
@@ -317,38 +333,60 @@ def render_answer(question,answer,index,res_img):
     #st.write("")
     colu1,colu2,colu3 = st.columns([4,82,20])
     with colu2:
-        #with st.expander("Relevant Sources:"):
-        with st.container():
-            if(len(res_img)>0):
-                with st.expander("Relevant Sources:"):
-                    #with st.expander("Images:"):
-                    st.write("Images:")
-                    col3,col4,col5 = st.columns([33,33,33])
-                    cols = [col3,col4]
-                    idx = 0
-                    #print(res_img)
-                    for img_ in res_img:
-                        if(img_['file'].lower()!='none' and idx < 2):
-                            img = img_['file'].split(".")[0]
-                            caption = img_['caption']
+        with st.expander("Relevant Sources:"):
+            with st.container():
+                if(len(res_img)>0):
+                    with st.expander("Images:"):
+                        
+                        idx = 0
+                        print(res_img)
+                        for i in range(0,len(res_img)):
                             
-                            with cols[idx]:
+                            if(st.session_state.input_is_colpali):
+                                if(st.session_state.show_columns == True):
+                                    cols_per_row = 3
+                                    st.session_state.image_placeholder=st.empty()
+                                    with st.session_state.image_placeholder.container():
+                                        row = st.columns(cols_per_row)
+                                    for j, item in enumerate(res_img[i:i+cols_per_row]):
+                                        with row[j]:
+                                            st.image(item['file'])
+                                    
+                                else:
+                                    st.session_state.image_placeholder = st.empty()
+                                    with st.session_state.image_placeholder.container():
+                                        col3_,col4_,col5_ = st.columns([33,33,33])
+                                        with col3_:
+                                            st.image(res_img[i]['file'])
+                                    
+                                    
+
+                                
+                                
+                            else:
+                                if(res_img[i]['file'].lower()!='none' and idx < 2):
+                                    col3,col4,col5 = st.columns([33,33,33])
+                                    cols = [col3,col4]
+                                    img = res_img[i]['file'].split(".")[0]
+                                    caption = res_img[i]['caption']
                                 
-                                st.image(parent_dirname+"/figures/"+st.session_state.input_index+"/"+img+".jpg")
+                                    with cols[idx]:
+                                        
+                                        st.image(parent_dirname+"/figures/"+st.session_state.input_index+"/"+img+".jpg")
                                 #st.write(caption)
-                            idx = idx+1
+                                idx = idx+1
+                        if(st.session_state.show_columns == True):
+                            st.session_state.show_columns = False
                 #st.markdown("<div style='color:#e28743';padding:3px 7px 3px 7px;borderWidth: 0px;borderColor: red;borderStyle: solid;width: fit-content;height: fit-content;border-radius: 10px;'><b>Sources from the document:</b></div>", unsafe_allow_html = True)
-                    if(len(answer["table"] )>0):
-                        #with st.expander("Table:"):
-                        st.write("Table:")
+                if(len(answer["table"] )>0):
+                    with st.expander("Table:"):
                         df = pd.read_csv(answer["table"][0]['name'],skipinitialspace = True, on_bad_lines='skip',delimiter='`')
                         df.fillna(method='pad', inplace=True)
                         st.table(df)
-                    #with st.expander("Raw sources:"):
-                    st.write("Raw sources:")
+                with st.expander("Raw sources:"):
                     st.write(answer["source"])
-            
-    
+                
+        
         
     with col_3:
         
@@ -360,22 +398,25 @@ def render_answer(question,answer,index,res_img):
 
             rdn_key = ''.join([random.choice(string.ascii_letters)
                               for _ in range(10)])
+            rdn_key_1 = ''.join([random.choice(string.ascii_letters)
+                              for _ in range(10)])
             currentValue = ''.join(st.session_state.input_rag_searchType)+str(st.session_state.input_is_rerank)+str(st.session_state.input_table_with_sql)+st.session_state.input_index
             oldValue = ''.join(st.session_state.inputs_["rag_searchType"])+str(st.session_state.inputs_["is_rerank"])+str(st.session_state.inputs_["table_with_sql"])+str(st.session_state.inputs_["index"])
             #print("changing values-----------------")
             def on_button_click():
-                # print("button clicked---------------")
-                # print(currentValue)
-                # print(oldValue)
                 if(currentValue!=oldValue or 1==1): 
-                    #print("----------regenerate----------------")
                     st.session_state.input_query = st.session_state.questions_[-1]["question"]
                     st.session_state.answers_.pop()
                     st.session_state.questions_.pop()
                     
-                    handle_input()
-                    with placeholder.container():
-                        render_all()
+                    
+            def show_maxsim():
+                st.session_state.show_columns = True
+                st.session_state.maxSimImages = colpali.img_highlight(st.session_state.top_img, st.session_state.query_token_vectors, st.session_state.query_tokens)
+                handle_input()
+                with placeholder.container():
+                    render_all()
+                
 
             if("currentValue"  in st.session_state):
                 del st.session_state["currentValue"]
@@ -385,16 +426,18 @@ def render_answer(question,answer,index,res_img):
             except:
                 pass  
 
-            #print("------------------------")
-            #print(st.session_state)
-
             placeholder__ = st.empty()
-            
             placeholder__.button("🔄",key=rdn_key,on_click=on_button_click)
+            placeholder__.button("Show similarity map",key=rdn_key_1,on_click = show_maxsim)
+        
+        
      
 #Each answer will have context of the question asked in order to associate the provided feedback with the respective question
 def write_chat_message(md, q,index):
-    res_img = md['image']
+    if(st.session_state.show_columns):
+        res_img = st.session_state.maxSimImages
+    else:
+        res_img = md['image']
     #st.session_state['session_id'] = res['session_id']   to be added in memory
     chat = st.container()
     with chat:
@@ -425,7 +468,7 @@ with col_2:
     input = st.text_input( "Ask here",label_visibility = "collapsed",key="input_query")
 with col_3:
     #hidden = st.button("RUN",disabled=True,key = "hidden")
-    play = st.button("GO",on_click=handle_input,key = "play")
+    play = st.button("Go",on_click=handle_input,key = "play")
 with st.sidebar:
     st.page_link("app.py", label=":orange[Home]", icon="🏠")
     st.subheader(":blue[Sample Data]")
@@ -454,16 +497,19 @@ with st.sidebar:
     }
     </style>
     """,unsafe_allow_html=True)   
+    with st.expander("Sample questions:"):
+        st.markdown("<span style = 'color:#FF9900;'>UK Housing</span> - which city has the highest average housing price in UK ?",unsafe_allow_html=True)
+        st.markdown("<span style = 'color:#FF9900;'>Global Warming stats</span> - What is the projected energy percentage from renewable sources in future?",unsafe_allow_html=True)
+        st.markdown("<span style = 'color:#FF9900;'>Covid19 impacts</span> - How many aged above 85 years died due to covid ?",unsafe_allow_html=True)
+        
     # Initialize boto3 to use the S3 client.
-    s3_client = boto3.resource('s3',aws_access_key_id=st.secrets['user_access_key'],
-                aws_secret_access_key=st.secrets['user_secret_key'], region_name = 'us-east-1')
+    s3_client = boto3.resource('s3')
     bucket=s3_client.Bucket(s3_bucket_)
 
     objects = bucket.objects.filter(Prefix="sample_pdfs/")
     urls = []
 
-    client = boto3.client('s3',aws_access_key_id=st.secrets['user_access_key'],
-                aws_secret_access_key=st.secrets['user_secret_key'], region_name = 'us-east-1')
+    client = boto3.client('s3')
 
     for obj in objects:
         if obj.key.endswith('.pdf'): 
@@ -510,18 +556,29 @@ with st.sidebar:
     #             print('lambda done')
     #     st.success('you can start searching on your PDF')
         
-         
-    # if(pdf_doc_ is None or pdf_doc_ == ""):
-    #     if(index_select == "Global Warming stats"):
-    #         st.session_state.input_index = "globalwarmingnew"
-    #     if(index_select == "Covid19 impacts on Ireland"):
-    #         st.session_state.input_index = "covid19ie"#"choosetheknnalgorithmforyourbillionscaleusecasewithopensearchawsbigdatablog"
-    #     if(index_select == "BEIR"):
-    #         st.session_state.input_index = "2104"
-    #     if(index_select == "UK Housing"):
-    #         st.session_state.input_index = "ukhousingstats"
-   
-    
+    ############## haystach demo temporary addition ############    
+    # st.subheader(":blue[Multimodality]")
+    # colu1,colu2 = st.columns([50,50])
+    # with colu1:
+    #     in_images = st.toggle('Images', key = 'in_images', disabled = False)
+    # with colu2:
+    #     in_tables = st.toggle('Tables', key = 'in_tables', disabled = False)   
+    # if(in_tables):
+    #     st.session_state.input_table_with_sql = True
+    # else:
+    #     st.session_state.input_table_with_sql = False
+        
+     ############## haystach demo temporary addition ############       
+    #if(pdf_doc_ is None or pdf_doc_ == ""):
+    if(index_select == "Global Warming stats"):
+        st.session_state.input_index = "globalwarmingnew"
+    if(index_select == "Covid19 impacts on Ireland"):
+        st.session_state.input_index = "covid19ie"#"choosetheknnalgorithmforyourbillionscaleusecasewithopensearchawsbigdatablog"
+    if(index_select == "BEIR"):
+        st.session_state.input_index = "2104"
+    if(index_select == "UK Housing"):
+        st.session_state.input_index = "hpijan2024hometrack"
+          
     # custom_index = st.text_input("If uploaded the file already, enter the original file name", value = "")
     # if(custom_index!=""):
     #     st.session_state.input_index = re.sub('[^A-Za-z0-9]+', '', (custom_index.lower().replace(".pdf","").split("/")[-1].split(".")[0]).lower())
@@ -534,18 +591,33 @@ with st.sidebar:
     'Vector Search', 
     'Sparse Search',
     ],
-    ['Sparse Search'],
+    ['Vector Search'],
 
     key = 'input_rag_searchType',
     help = "Select the type of Search, adding more than one search type will activate hybrid search"#\n1. Conversational Search (Recommended) - This will include both the OpenSearch and LLM in the retrieval pipeline \n (note: This will put opensearch response as context to LLM to answer) \n2. OpenSearch vector search - This will put only OpenSearch's vector search in the pipeline, \n(Warning: this will lead to unformatted results )\n3. LLM Text Generation - This will include only LLM in the pipeline, \n(Warning: This will give hallucinated and out of context answers)"
     )
     
     re_rank = st.checkbox('Re-rank results', key = 'input_re_rank', disabled = False, value = True, help = "Checking this box will re-rank the results using a cross-encoder model")
-        
+     
     if(re_rank):
         st.session_state.input_is_rerank = True
     else:
         st.session_state.input_is_rerank = False
+    
+    st.subheader(":blue[Multi-vector retrieval]")
+    
+    #st.write("Dataset indexed: https://huggingface.co/datasets/vespa-engine/gpfg-QA")
+    colpali_search_rerank =  st.checkbox('Try Colpali multi-vector retrieval on the [sample dataset](https://huggingface.co/datasets/vespa-engine/gpfg-QA)', key = 'input_colpali', disabled = False, value = False, help = "Checking this box will use colpali as the embedding model and retrieval is performed using multi-vectors followed by re-ranking using MaxSim")
+        
+    if(colpali_search_rerank):
+        st.session_state.input_is_colpali = True
+        #st.session_state.input_query = ""
+    else:
+        st.session_state.input_is_colpali = False
+    
+    with st.expander("Sample questions for Colpali retriever:"):
+        st.write("1. Proportion of female new hires 2021-2023? \n\n 2. First-half 2021 return on unlisted real estate investments? \n\n 3. Trend of the fund's expected absolute volatility between January 2014 and January 2016? \n\n 4. Fund return percentage in 2017? \n\n 5. Annualized gross return of the fund from 1997 to 2008?")
+    
         
     # copali_rerank = st.checkbox("Search and Re-rank with Token level vectors",key = 'copali_rerank',help = "Enabling this option uses 'Copali' model's page level image embeddings to retrieve documents and MaxSim to re-rank the pages.\n\n Hugging Face Model: https://huggingface.co/vidore/colpali")
         
@@ -553,5 +625,8 @@ with st.sidebar:
     #     st.session_state.input_copali_rerank = True
     # else:
     #     st.session_state.input_copali_rerank = False
+    
+    
+