Update llm_retrieval_conversation_rerank.py
Browse files- llm_retrieval_conversation_rerank.py +238 -238
llm_retrieval_conversation_rerank.py
CHANGED
@@ -1,239 +1,239 @@
|
|
1 |
-
import json
|
2 |
-
import os
|
3 |
-
from dotenv import load_dotenv
|
4 |
-
import yaml
|
5 |
-
from together import Together
|
6 |
-
from langchain.llms.together import Together as TogetherLLM
|
7 |
-
from langchain.prompts import PromptTemplate
|
8 |
-
from langchain.schema.runnable import RunnablePassthrough
|
9 |
-
from langchain.schema.output_parser import StrOutputParser
|
10 |
-
from pinecone import Pinecone
|
11 |
-
from typing import List, Dict
|
12 |
-
import cohere
|
13 |
-
load_dotenv()
|
14 |
-
|
15 |
-
|
16 |
-
API_FILE_PATH = r"
|
17 |
-
COURSES_FILE_PATH = r"
|
18 |
-
|
19 |
-
# Global list to store conversation history
|
20 |
-
conversation_history: List[Dict[str, str]] = []
|
21 |
-
|
22 |
-
def load_api_keys(api_file_path):
|
23 |
-
"""Loads API keys from a YAML file."""
|
24 |
-
with open(api_file_path, 'r') as f:
|
25 |
-
api_keys = yaml.safe_load(f)
|
26 |
-
return api_keys
|
27 |
-
|
28 |
-
def generate_query_embedding(query, together_api_key):
|
29 |
-
"""Generates embedding for the user query."""
|
30 |
-
client = Together(api_key=together_api_key)
|
31 |
-
response = client.embeddings.create(
|
32 |
-
model="WhereIsAI/UAE-Large-V1", input=query
|
33 |
-
)
|
34 |
-
return response.data[0].embedding
|
35 |
-
|
36 |
-
def initialize_pinecone(pinecone_api_key):
|
37 |
-
"""Initializes Pinecone with API key."""
|
38 |
-
return Pinecone(api_key=pinecone_api_key)
|
39 |
-
|
40 |
-
def pinecone_similarity_search(pinecone_instance, index_name, query_embedding, top_k=10):
|
41 |
-
"""Performs a similarity search in Pinecone and increase top k for reranking."""
|
42 |
-
try:
|
43 |
-
index = pinecone_instance.Index(index_name)
|
44 |
-
results = index.query(vector=query_embedding, top_k=top_k, include_metadata=True)
|
45 |
-
if not results.matches:
|
46 |
-
return None
|
47 |
-
return results
|
48 |
-
except Exception as e:
|
49 |
-
print(f"Error during similarity search: {e}")
|
50 |
-
return None
|
51 |
-
|
52 |
-
def create_prompt_template():
|
53 |
-
"""Creates a prompt template for LLM."""
|
54 |
-
template = """You are a helpful AI assistant that provides information on courses.
|
55 |
-
Based on the following context, conversation history, and new user query,
|
56 |
-
suggest relevant courses and explain why they might be useful, or respond accordingly if the user query is unrelated.
|
57 |
-
If no relevant courses are found, please indicate that.
|
58 |
-
|
59 |
-
Conversation History:
|
60 |
-
{conversation_history}
|
61 |
-
|
62 |
-
Context: {context}
|
63 |
-
User Query: {query}
|
64 |
-
|
65 |
-
Response: Let me help you find relevant courses based on your query.
|
66 |
-
"""
|
67 |
-
return PromptTemplate(template=template, input_variables=["context", "query", "conversation_history"])
|
68 |
-
|
69 |
-
def initialize_llm(together_api_key):
|
70 |
-
"""Initializes Together LLM."""
|
71 |
-
return TogetherLLM(
|
72 |
-
model="mistralai/Mixtral-8x7B-Instruct-v0.1",
|
73 |
-
together_api_key=together_api_key,
|
74 |
-
temperature=0,
|
75 |
-
max_tokens=250
|
76 |
-
)
|
77 |
-
|
78 |
-
def create_chain(llm, prompt):
|
79 |
-
"""Creates a chain using the new RunnableSequence approach."""
|
80 |
-
chain = (
|
81 |
-
{"context": RunnablePassthrough(), "query": RunnablePassthrough(), "conversation_history": RunnablePassthrough()}
|
82 |
-
| prompt
|
83 |
-
| llm
|
84 |
-
| StrOutputParser()
|
85 |
-
)
|
86 |
-
return chain
|
87 |
-
|
88 |
-
|
89 |
-
def initialize_cohere_client(cohere_api_key):
|
90 |
-
"""Initializes the Cohere client."""
|
91 |
-
return cohere.ClientV2(api_key=cohere_api_key)
|
92 |
-
|
93 |
-
|
94 |
-
def rerank_results(cohere_client, query, documents, top_n=3):
|
95 |
-
"""Reranks documents using Cohere."""
|
96 |
-
try:
|
97 |
-
results = cohere_client.rerank(
|
98 |
-
query=query,
|
99 |
-
documents=documents,
|
100 |
-
top_n=top_n,
|
101 |
-
model="rerank-english-v3.0",
|
102 |
-
)
|
103 |
-
return results
|
104 |
-
except Exception as e:
|
105 |
-
print(f"Error reranking results: {e}")
|
106 |
-
return None
|
107 |
-
|
108 |
-
def generate_llm_response(chain, query, retrieved_data, history, cohere_client):
|
109 |
-
"""Generates an LLM response based on context and conversation history."""
|
110 |
-
try:
|
111 |
-
if not retrieved_data or not retrieved_data.matches:
|
112 |
-
return "I couldn't find any relevant courses matching your query. Please try a different search term."
|
113 |
-
|
114 |
-
# Prepare documents for reranking
|
115 |
-
documents = []
|
116 |
-
for match in retrieved_data.matches:
|
117 |
-
metadata = match.metadata
|
118 |
-
if metadata:
|
119 |
-
documents.append(
|
120 |
-
{ "text" :f"Title: {metadata.get('title', 'No title')}\nDescription: {metadata.get('text', 'No description')}\nLink: {metadata.get('course_link', 'No link')}"
|
121 |
-
}
|
122 |
-
)
|
123 |
-
|
124 |
-
if not documents:
|
125 |
-
return "I found some matches but couldn't extract course information. Please try again."
|
126 |
-
|
127 |
-
# Rerank the documents
|
128 |
-
reranked_results = rerank_results(cohere_client, query, documents)
|
129 |
-
|
130 |
-
if not reranked_results:
|
131 |
-
return "I couldn't rerank the results, please try again."
|
132 |
-
|
133 |
-
# Prepare context from reranked results
|
134 |
-
context_parts = []
|
135 |
-
for result in reranked_results.results:
|
136 |
-
context_parts.append(documents[result.index]["text"])
|
137 |
-
|
138 |
-
context = "\n\n".join(context_parts)
|
139 |
-
|
140 |
-
# Format conversation history
|
141 |
-
formatted_history = "\n".join(f"User: {item['user']}\nAssistant: {item['assistant']}" for item in history) if history else "No previous conversation."
|
142 |
-
|
143 |
-
response = chain.invoke({"context": context, "query": query, "conversation_history":formatted_history})
|
144 |
-
return response
|
145 |
-
|
146 |
-
except Exception as e:
|
147 |
-
print(f"Error generating response: {e}")
|
148 |
-
return "I encountered an error while generating the response. Please try again."
|
149 |
-
|
150 |
-
|
151 |
-
def check_context_similarity(query_embedding, previous_query_embedding, threshold=0.7):
|
152 |
-
"""Checks if the new query is related to the previous one."""
|
153 |
-
if not previous_query_embedding:
|
154 |
-
return False # First query, no previous embedding to compare
|
155 |
-
|
156 |
-
from numpy import dot
|
157 |
-
from numpy.linalg import norm
|
158 |
-
|
159 |
-
cos_sim = dot(query_embedding, previous_query_embedding) / (norm(query_embedding) * norm(previous_query_embedding))
|
160 |
-
return cos_sim > threshold
|
161 |
-
|
162 |
-
def main():
|
163 |
-
global conversation_history
|
164 |
-
previous_query_embedding = None
|
165 |
-
|
166 |
-
try:
|
167 |
-
|
168 |
-
api_keys = load_api_keys(API_FILE_PATH)
|
169 |
-
together_api_key = api_keys["together_ai_api_key"]
|
170 |
-
pinecone_api_key = api_keys["pinecone_api_key"]
|
171 |
-
index_name = api_keys["pinecone_index_name"]
|
172 |
-
cohere_api_key = api_keys["cohere_api_key"]
|
173 |
-
print("Initializing services...")
|
174 |
-
|
175 |
-
# Initialize Pinecone
|
176 |
-
pinecone_instance = initialize_pinecone(pinecone_api_key)
|
177 |
-
|
178 |
-
# Initialize Together LLM
|
179 |
-
llm = initialize_llm(together_api_key)
|
180 |
-
|
181 |
-
# Initialize Cohere client
|
182 |
-
cohere_client = initialize_cohere_client(cohere_api_key)
|
183 |
-
|
184 |
-
|
185 |
-
|
186 |
-
prompt = create_prompt_template()
|
187 |
-
|
188 |
-
# Create chain
|
189 |
-
chain = create_chain(llm, prompt)
|
190 |
-
|
191 |
-
print("Ready to process queries!")
|
192 |
-
|
193 |
-
while True:
|
194 |
-
|
195 |
-
user_query = input("\nEnter your query (or 'quit' to exit): ").strip()
|
196 |
-
|
197 |
-
if user_query.lower() == 'quit':
|
198 |
-
break
|
199 |
-
|
200 |
-
if not user_query:
|
201 |
-
print("Please enter a valid query.")
|
202 |
-
continue
|
203 |
-
|
204 |
-
try:
|
205 |
-
print("Generating query embedding...")
|
206 |
-
query_embedding = generate_query_embedding(user_query, together_api_key)
|
207 |
-
|
208 |
-
# Check context similarity
|
209 |
-
if previous_query_embedding and check_context_similarity(query_embedding, previous_query_embedding):
|
210 |
-
print("Continuing the previous conversation...")
|
211 |
-
else:
|
212 |
-
print("Starting a new conversation...")
|
213 |
-
conversation_history = [] # Clear history for a new conversation
|
214 |
-
|
215 |
-
print("Searching for relevant courses...")
|
216 |
-
pinecone_results = pinecone_similarity_search(
|
217 |
-
pinecone_instance, index_name, query_embedding
|
218 |
-
)
|
219 |
-
|
220 |
-
print("Generating response...")
|
221 |
-
llm_response = generate_llm_response(chain, user_query, pinecone_results, conversation_history, cohere_client)
|
222 |
-
|
223 |
-
print("\nResponse:")
|
224 |
-
print(llm_response)
|
225 |
-
print("\n" + "="*50)
|
226 |
-
|
227 |
-
# Update conversation history
|
228 |
-
conversation_history.append({"user": user_query, "assistant": llm_response})
|
229 |
-
previous_query_embedding = query_embedding # Save for next turn
|
230 |
-
|
231 |
-
except Exception as e:
|
232 |
-
print(f"Error processing query: {e}")
|
233 |
-
print("Please try again with a different query.")
|
234 |
-
|
235 |
-
except Exception as e:
|
236 |
-
print(f"An error occurred during initialization: {str(e)}")
|
237 |
-
|
238 |
-
if __name__ == "__main__":
|
239 |
main()
|
|
|
1 |
+
import json
|
2 |
+
import os
|
3 |
+
from dotenv import load_dotenv
|
4 |
+
import yaml
|
5 |
+
from together import Together
|
6 |
+
from langchain.llms.together import Together as TogetherLLM
|
7 |
+
from langchain.prompts import PromptTemplate
|
8 |
+
from langchain.schema.runnable import RunnablePassthrough
|
9 |
+
from langchain.schema.output_parser import StrOutputParser
|
10 |
+
from pinecone import Pinecone
|
11 |
+
from typing import List, Dict
|
12 |
+
import cohere
|
13 |
+
load_dotenv()
|
14 |
+
|
15 |
+
|
16 |
+
API_FILE_PATH = r".\API.yml"
|
17 |
+
COURSES_FILE_PATH = r".\courses.json"
|
18 |
+
|
19 |
+
# Global list to store conversation history
|
20 |
+
conversation_history: List[Dict[str, str]] = []
|
21 |
+
|
22 |
+
def load_api_keys(api_file_path):
|
23 |
+
"""Loads API keys from a YAML file."""
|
24 |
+
with open(api_file_path, 'r') as f:
|
25 |
+
api_keys = yaml.safe_load(f)
|
26 |
+
return api_keys
|
27 |
+
|
28 |
+
def generate_query_embedding(query, together_api_key):
|
29 |
+
"""Generates embedding for the user query."""
|
30 |
+
client = Together(api_key=together_api_key)
|
31 |
+
response = client.embeddings.create(
|
32 |
+
model="WhereIsAI/UAE-Large-V1", input=query
|
33 |
+
)
|
34 |
+
return response.data[0].embedding
|
35 |
+
|
36 |
+
def initialize_pinecone(pinecone_api_key):
|
37 |
+
"""Initializes Pinecone with API key."""
|
38 |
+
return Pinecone(api_key=pinecone_api_key)
|
39 |
+
|
40 |
+
def pinecone_similarity_search(pinecone_instance, index_name, query_embedding, top_k=10):
|
41 |
+
"""Performs a similarity search in Pinecone and increase top k for reranking."""
|
42 |
+
try:
|
43 |
+
index = pinecone_instance.Index(index_name)
|
44 |
+
results = index.query(vector=query_embedding, top_k=top_k, include_metadata=True)
|
45 |
+
if not results.matches:
|
46 |
+
return None
|
47 |
+
return results
|
48 |
+
except Exception as e:
|
49 |
+
print(f"Error during similarity search: {e}")
|
50 |
+
return None
|
51 |
+
|
52 |
+
def create_prompt_template():
|
53 |
+
"""Creates a prompt template for LLM."""
|
54 |
+
template = """You are a helpful AI assistant that provides information on courses.
|
55 |
+
Based on the following context, conversation history, and new user query,
|
56 |
+
suggest relevant courses and explain why they might be useful, or respond accordingly if the user query is unrelated.
|
57 |
+
If no relevant courses are found, please indicate that.
|
58 |
+
|
59 |
+
Conversation History:
|
60 |
+
{conversation_history}
|
61 |
+
|
62 |
+
Context: {context}
|
63 |
+
User Query: {query}
|
64 |
+
|
65 |
+
Response: Let me help you find relevant courses based on your query.
|
66 |
+
"""
|
67 |
+
return PromptTemplate(template=template, input_variables=["context", "query", "conversation_history"])
|
68 |
+
|
69 |
+
def initialize_llm(together_api_key):
|
70 |
+
"""Initializes Together LLM."""
|
71 |
+
return TogetherLLM(
|
72 |
+
model="mistralai/Mixtral-8x7B-Instruct-v0.1",
|
73 |
+
together_api_key=together_api_key,
|
74 |
+
temperature=0,
|
75 |
+
max_tokens=250
|
76 |
+
)
|
77 |
+
|
78 |
+
def create_chain(llm, prompt):
|
79 |
+
"""Creates a chain using the new RunnableSequence approach."""
|
80 |
+
chain = (
|
81 |
+
{"context": RunnablePassthrough(), "query": RunnablePassthrough(), "conversation_history": RunnablePassthrough()}
|
82 |
+
| prompt
|
83 |
+
| llm
|
84 |
+
| StrOutputParser()
|
85 |
+
)
|
86 |
+
return chain
|
87 |
+
|
88 |
+
|
89 |
+
def initialize_cohere_client(cohere_api_key):
|
90 |
+
"""Initializes the Cohere client."""
|
91 |
+
return cohere.ClientV2(api_key=cohere_api_key)
|
92 |
+
|
93 |
+
|
94 |
+
def rerank_results(cohere_client, query, documents, top_n=3):
|
95 |
+
"""Reranks documents using Cohere."""
|
96 |
+
try:
|
97 |
+
results = cohere_client.rerank(
|
98 |
+
query=query,
|
99 |
+
documents=documents,
|
100 |
+
top_n=top_n,
|
101 |
+
model="rerank-english-v3.0",
|
102 |
+
)
|
103 |
+
return results
|
104 |
+
except Exception as e:
|
105 |
+
print(f"Error reranking results: {e}")
|
106 |
+
return None
|
107 |
+
|
108 |
+
def generate_llm_response(chain, query, retrieved_data, history, cohere_client):
|
109 |
+
"""Generates an LLM response based on context and conversation history."""
|
110 |
+
try:
|
111 |
+
if not retrieved_data or not retrieved_data.matches:
|
112 |
+
return "I couldn't find any relevant courses matching your query. Please try a different search term."
|
113 |
+
|
114 |
+
# Prepare documents for reranking
|
115 |
+
documents = []
|
116 |
+
for match in retrieved_data.matches:
|
117 |
+
metadata = match.metadata
|
118 |
+
if metadata:
|
119 |
+
documents.append(
|
120 |
+
{ "text" :f"Title: {metadata.get('title', 'No title')}\nDescription: {metadata.get('text', 'No description')}\nLink: {metadata.get('course_link', 'No link')}"
|
121 |
+
}
|
122 |
+
)
|
123 |
+
|
124 |
+
if not documents:
|
125 |
+
return "I found some matches but couldn't extract course information. Please try again."
|
126 |
+
|
127 |
+
# Rerank the documents
|
128 |
+
reranked_results = rerank_results(cohere_client, query, documents)
|
129 |
+
|
130 |
+
if not reranked_results:
|
131 |
+
return "I couldn't rerank the results, please try again."
|
132 |
+
|
133 |
+
# Prepare context from reranked results
|
134 |
+
context_parts = []
|
135 |
+
for result in reranked_results.results:
|
136 |
+
context_parts.append(documents[result.index]["text"])
|
137 |
+
|
138 |
+
context = "\n\n".join(context_parts)
|
139 |
+
|
140 |
+
# Format conversation history
|
141 |
+
formatted_history = "\n".join(f"User: {item['user']}\nAssistant: {item['assistant']}" for item in history) if history else "No previous conversation."
|
142 |
+
|
143 |
+
response = chain.invoke({"context": context, "query": query, "conversation_history":formatted_history})
|
144 |
+
return response
|
145 |
+
|
146 |
+
except Exception as e:
|
147 |
+
print(f"Error generating response: {e}")
|
148 |
+
return "I encountered an error while generating the response. Please try again."
|
149 |
+
|
150 |
+
|
151 |
+
def check_context_similarity(query_embedding, previous_query_embedding, threshold=0.7):
|
152 |
+
"""Checks if the new query is related to the previous one."""
|
153 |
+
if not previous_query_embedding:
|
154 |
+
return False # First query, no previous embedding to compare
|
155 |
+
|
156 |
+
from numpy import dot
|
157 |
+
from numpy.linalg import norm
|
158 |
+
|
159 |
+
cos_sim = dot(query_embedding, previous_query_embedding) / (norm(query_embedding) * norm(previous_query_embedding))
|
160 |
+
return cos_sim > threshold
|
161 |
+
|
162 |
+
def main():
|
163 |
+
global conversation_history
|
164 |
+
previous_query_embedding = None
|
165 |
+
|
166 |
+
try:
|
167 |
+
|
168 |
+
api_keys = load_api_keys(API_FILE_PATH)
|
169 |
+
together_api_key = api_keys["together_ai_api_key"]
|
170 |
+
pinecone_api_key = api_keys["pinecone_api_key"]
|
171 |
+
index_name = api_keys["pinecone_index_name"]
|
172 |
+
cohere_api_key = api_keys["cohere_api_key"]
|
173 |
+
print("Initializing services...")
|
174 |
+
|
175 |
+
# Initialize Pinecone
|
176 |
+
pinecone_instance = initialize_pinecone(pinecone_api_key)
|
177 |
+
|
178 |
+
# Initialize Together LLM
|
179 |
+
llm = initialize_llm(together_api_key)
|
180 |
+
|
181 |
+
# Initialize Cohere client
|
182 |
+
cohere_client = initialize_cohere_client(cohere_api_key)
|
183 |
+
|
184 |
+
|
185 |
+
|
186 |
+
prompt = create_prompt_template()
|
187 |
+
|
188 |
+
# Create chain
|
189 |
+
chain = create_chain(llm, prompt)
|
190 |
+
|
191 |
+
print("Ready to process queries!")
|
192 |
+
|
193 |
+
while True:
|
194 |
+
|
195 |
+
user_query = input("\nEnter your query (or 'quit' to exit): ").strip()
|
196 |
+
|
197 |
+
if user_query.lower() == 'quit':
|
198 |
+
break
|
199 |
+
|
200 |
+
if not user_query:
|
201 |
+
print("Please enter a valid query.")
|
202 |
+
continue
|
203 |
+
|
204 |
+
try:
|
205 |
+
print("Generating query embedding...")
|
206 |
+
query_embedding = generate_query_embedding(user_query, together_api_key)
|
207 |
+
|
208 |
+
# Check context similarity
|
209 |
+
if previous_query_embedding and check_context_similarity(query_embedding, previous_query_embedding):
|
210 |
+
print("Continuing the previous conversation...")
|
211 |
+
else:
|
212 |
+
print("Starting a new conversation...")
|
213 |
+
conversation_history = [] # Clear history for a new conversation
|
214 |
+
|
215 |
+
print("Searching for relevant courses...")
|
216 |
+
pinecone_results = pinecone_similarity_search(
|
217 |
+
pinecone_instance, index_name, query_embedding
|
218 |
+
)
|
219 |
+
|
220 |
+
print("Generating response...")
|
221 |
+
llm_response = generate_llm_response(chain, user_query, pinecone_results, conversation_history, cohere_client)
|
222 |
+
|
223 |
+
print("\nResponse:")
|
224 |
+
print(llm_response)
|
225 |
+
print("\n" + "="*50)
|
226 |
+
|
227 |
+
# Update conversation history
|
228 |
+
conversation_history.append({"user": user_query, "assistant": llm_response})
|
229 |
+
previous_query_embedding = query_embedding # Save for next turn
|
230 |
+
|
231 |
+
except Exception as e:
|
232 |
+
print(f"Error processing query: {e}")
|
233 |
+
print("Please try again with a different query.")
|
234 |
+
|
235 |
+
except Exception as e:
|
236 |
+
print(f"An error occurred during initialization: {str(e)}")
|
237 |
+
|
238 |
+
if __name__ == "__main__":
|
239 |
main()
|