Update app.py

app.py

@@ -6,6 +6,7 @@ from sentence_transformers import SentenceTransformer, util
 from openai import OpenAI
 from datetime import datetime
 import pandas as pd
+import numpy as np
 import os
 import time
 import json
@@ -33,26 +34,21 @@ index = pc.Index(name=INDEX_NAME)
 system_instructions_text='''
           Your task is to extract the answer to a question from a body of text provided to you.
           The body of text will be enclosed within the delimiter tags  <text> and </text>
-
           For example,
           <text> General Preparation Tips for VARC Section:
-
           You need to develop an incessant habit of speed reading.
           Start with reading newspapers, editorials, fiction and nonfiction novels and simple passages.
           The more you read, the faster you read. Learn the basic grammar concepts like parts of speech, articles,verbs, adjectives, tenses, auxiliary verbs, modifiers, modals etc.
           Revise at least 50 new words every day
           </text>
-
           Question: What are some tips for preparing for VARC?
           Here are some tips for preparing for the VARC section:
           1. develop an incessant habit of speed reading
           2. Start reading newspapers, editorials, fiction and nonfiction novels
           3. Learn basic grammar concepts\n
           4. Revise at least 50 new words a day
-
           Question: How many new words are to be learnt in a day?
           It is advised that 50 new words are learn every day
-
           Your response should be based on the information contained in the provided text and should not included any other sources.
           If you are unable to answer the question from the text provided, please respond " Sorry. I do not have enough information to answer this"
            Do repeat the question. Do not make a pointed reference to the text provided. Directly answer the question
@@ -66,7 +62,6 @@ json_instructions='''
           ]
           You need to check which content is most appropriate to answer the question and prepare
           an answer based on the content
-
           For example,
           [
           { "id":1 , "content" : "General Preparation Tips for Verbal Section:\n
@@ -80,15 +75,12 @@ json_instructions='''
           Learn the basic  concepts like arithmetic, geometry, numbers, probability,  etc.
           Solve at least 50 new problems every day"}
           ]
-
           Question: What are some tips for preparing for Verbal exam?
           Here are some tips for preparing for the VARC section:
           1. develop an incessant habit of speed reading
           2. Start reading newspapers, editorials, fiction and nonfiction novels
           3. Learn basic grammar concepts\n
           4. Revise at least 50 new words a day
-
-
           Your response should be based on the information contained in the provided content in the json and should not included any other sources.
           If you are unable to answer the question from the content provided, please respond " Sorry. I do not have enough information to answer this"
            Do repeat the question. Do not make a pointed reference to the content provided. Directly answer the question
@@ -96,16 +88,15 @@ json_instructions='''
 
 def get_meta_score(url,question_embedding):
     qry = index.fetch(ids=[url], namespace="meta")
-    try:
-        emb=qry['vectors'][url]['values']
-        vector1 = np.array(emb).reshape(1, -1)  # Reshape to ensure compatibility with sklearn
-        vector2 = question_embedding.numpy().reshape(1, -1)
-        
-        # Calculate cosine similarity
-        cosine_scores = util.cos_sim(question_embedding, emb)
-        return  cosine_scores.item()
-    except:
-        return 0
+
+    emb=qry['vectors'][url]['values']
+    vector1 = np.array(emb).reshape(1, -1)  # Reshape to ensure compatibility with sklearn
+    vector2 = question_embedding.numpy().reshape(1, -1)
+
+    # Calculate cosine similarity
+    cosine_scores = util.cos_sim(question_embedding, emb)
+    return  cosine_scores.item()
+
 
 def query_from_pinecone(index,namespace, question_embedding, top_k=3):
     # get embedding from THE SAME embedder as the documents
@@ -140,16 +131,21 @@ def write_log(query,url, score, ans, ts):
         f.write(buffer)
 
     return
-    
+
 logfile='querylog.csv'
 try:
-    df_log=pd.read_csv(logfile, index_col=0)
+    df_log = pd.read_csv("hf://datasets/sujitb/data/querylog.csv", encoding="utf-8")
 except:
     df_log=pd.DataFrame(columns=['query','url','score','ans', 'ts'])
 
-
-
 st.title('CLLM Answering Machine')
+with st.sidebar:
+    st.markdown('*Search History*')
+    st.write('# Queries', len(df_log))
+    qrylist = df_log['query'].tail(10).tolist()
+    for q in qrylist[::-1]:
+        st.write(q)
+
 
 # Initialize chat history
 if "messages" not in st.session_state:
@@ -175,8 +171,10 @@ if QUESTION:
     # Add user message to chat history
     st.session_state.messages.append({"role": "user", "content": QUESTION})
 
+    st.write('Searching knowledgebase...')
     question_embedding = bi_encoder.encode(QUESTION, convert_to_tensor=True)
 
+    THRESHOLD=.4
     ns='webpages'
     ns='full'
     resp= query_from_pinecone(index,ns, question_embedding.tolist(), 10)
@@ -191,41 +189,42 @@ if QUESTION:
         meta_score= get_meta_score(r['id'],question_embedding)
         score=.5* r['score'] + .5*meta_score
         d['score']=score
-        st.write(d['url'], score, r['score'], meta_score)
+        #st.write(d['url'], score, r['score'], meta_score)
         resplist.append(d)
 
     if len(resplist)>0:
         sorted_indices = sorted(range(len(resplist)), key=lambda i: resplist[i]['score'], reverse=True)
-    
+
         # Get the elements with the top 2 highest values
         top_2 =  [resplist[i] for i in sorted_indices[:2]]
-                
+
         # covert to array
-        
+
         json_data = json.dumps(top_2)
-    
-        
-        goodmatch=False    
-        if resplist[sorted_indices[0]]['score']>.5:
+
+
+        goodmatch=False
+        if resplist[sorted_indices[0]]['score']>=THRESHOLD:
+            st.write('Preparing answers...')
             goodmatch=True
             mode = "two"  # two passages
-    
+
             client = OpenAI()
-    
+
             if mode=="one":
                 instr=system_instructions_text
-    
+
                 out= resplist[sorted_indices[0]]['content']
                 content="""
                 <text>
                 {}
                 </text>
                 """.format(out)
-                
+
             if mode=="two":
                 instr=json_instructions
                 content=json_data
-    
+
             response = client.chat.completions.create(
               model="gpt-3.5-turbo",
               messages=[
@@ -234,11 +233,11 @@ if QUESTION:
                 {"role": "user", "content": "Question:"+QUESTION}
               ]
             )
-    
+
             ans= response.choices[0].message.content
         else:
             ans='Weak match to your query. Please try reframing your question'
-    
+
             #st.write("Matched URL:{}  Score:{}".format(url,score))
             testing = False
             if testing:
@@ -246,27 +245,28 @@ if QUESTION:
                     st.write("2nd Matched URL:{}  Score:{}".format(resp[1]['id'],resp[1]['score']))
                 if len(resp)>=2:
                     st.write("3rd Matched URL:{}  Score:{}".format(resp[2]['id'],resp[2]['score']))
-    
-            with st.chat_message("assistant"):
-                response = st.write_stream(response_generator(ans))
-                if goodmatch:
-                    st.write('Resources:'+top_2[0]['url'])
-                    st.write(top_2[1]['url'])
-                # Add assistant response to chat history
-                st.session_state.messages.append({"role": "assistant", "content": response})
-            #st.write(ans)
-    
-            #st.write(' ----------------------')
-            #st.write(out)
-    
-            now= str(datetime.utcnow())
-            url = top_2[0]['url'] + ' ; '+top_2[1]['url']
-            df_log.loc[len(df_log)]=[QUESTION,url,score,ans,now]
-            write_log(QUESTION,url, score, ans, now)
-            #df.to_csv("hf://datasets/sujitb/data/test.csv")
-            #df_log.to_csv("hf://datasets/sujitb/data/"+logfile)
-
-    else:  ## Zero response from pinecone query 
+        ##  Send RESPONSE
+        with st.chat_message("assistant"):
+            response = st.write_stream(response_generator(ans))
+            if goodmatch:
+                st.write('Resources:')
+                st.write(top_2[0]['url'])
+                st.write(top_2[1]['url'])
+            # Add assistant response to chat history
+            st.session_state.messages.append({"role": "assistant", "content": response})
+        #st.write(ans)
+
+        #st.write(' ----------------------')
+        #st.write(out)
+
+        now= str(datetime.utcnow())
+        url = top_2[0]['url'] + ' ; '+top_2[1]['url']
+        df_log.loc[len(df_log)]=[QUESTION,url,score,ans,now]
+        write_log(QUESTION,url, score, ans, now)
+        #df.to_csv("hf://datasets/sujitb/data/test.csv")
+        #df_log.to_csv("hf://datasets/sujitb/data/"+logfile)
+
+    else:  ## Zero response from pinecone query
         #st.write("No matches for query")
         ans= "No matches for query"
         response = st.write_stream(response_generator(ans))
@@ -276,4 +276,4 @@ if QUESTION:
         now= str(datetime.utcnow())
         df_log.loc[len(df_log)]=[QUESTION,'No match',0,'-',now]
         #df_log.to_csv("hf://datasets/sujitb/data/"+logfile)
-        write_log(QUESTION,url, score, ans, now)
+        write_log(QUESTION,'No match', 0, '-', now)