Custom prompts and gpt-4 model

app.py

@@ -1,5 +1,5 @@
 import streamlit as st
-from index import build_index, build_service_context, load_documents
+from index import build_index, build_service_context, change_prompts, load_documents
 
 st.title("SAIRA: Student Affairs AI Response Assistant")
 st.caption('Welcome to the SAIRA chatbot! This bot have knowledge about Innopolis University. Feel free to write your request!')
@@ -8,11 +8,11 @@ st.caption('Welcome to the SAIRA chatbot! This bot have knowledge about Innopoli
 def load_docs_and_build_index():
     service_context = build_service_context()
     docs = load_documents()
-    return build_index(docs, service_context)
+    index = build_index(docs, service_context)
+    query_engine = index.as_query_engine(streaming=True)
+    change_prompts(query_engine)
+    return query_engine
 
-index = load_docs_and_build_index()
-
-query_engine = index.as_query_engine(streaming=True)
 
 # Initialize chat history
 if "messages" not in st.session_state:

index.py

@@ -5,6 +5,7 @@ from llama_index import (
     StorageContext,
     load_index_from_storage
 )
+from llama_index.prompts import ChatPromptTemplate, ChatMessage, MessageRole
 from llama_index.vector_stores import SimpleVectorStore
 from llama_index.llms import Ollama, OpenAI
 import os
@@ -20,7 +21,7 @@ def load_documents():
 
 def build_service_context():
     # llm = Ollama(model='mistral')
-    llm = OpenAI(model="gpt-3.5-turbo")
+    llm = OpenAI(model="gpt-4-1106-preview")
     return ServiceContext.from_defaults(llm=llm, embed_model="local:BAAI/bge-large-en-v1.5")
 
 def build_index(documents, service_context):
@@ -52,3 +53,29 @@ def build_index(documents, service_context):
         # storage_context.persist(persist_dir=persist_dir)
         index.storage_context.persist(persist_dir=persist_dir)
     return index
+
+def change_prompts(query_engine):
+    message_templates = [
+        ChatMessage(content='''You are SAIRA (Student Affairs AI Response Assistant) - expert Q&A system for the students of Innopolis University.
+Always answer the query using the provided context information, and not prior knowledge.
+Never directly reference the given context in your answer. Never use file names or any other meta information in the answer.
+If you mention person or department, provide also their Telegram or E-mail.
+If you mention some Telegram chat, give the link to it''', role=MessageRole.SYSTEM),
+        ChatMessage(content='''Context information:
+{context_str}
+---------------------
+Given the context information and not prior knowledge, answer the query. 
+
+Query: {query_str}
+
+Avoid statements like 'based on the context' or 'the context information', 'in the context' or anything along those lines. 
+Never use word 'context' in the answer!
+If you can't write answer, or it is not providied in context, just write '<SPECIALIST>' as an answer, and the request will be transferred to the specialist.
+Write '<SPECIALIST>' instead of asking to contact Student Affairs.''', role=MessageRole.USER),
+    ]
+    qa_prompt_tmpl = ChatPromptTemplate(message_templates)
+
+    query_engine.update_prompts(
+        {"response_synthesizer:text_qa_template": qa_prompt_tmpl}
+    )
+