Update retrieval.py
Browse files- retrieval.py +18 -46
retrieval.py
CHANGED
@@ -2,9 +2,6 @@
|
|
2 |
LLM chain retrieval
|
3 |
"""
|
4 |
|
5 |
-
|
6 |
-
|
7 |
-
|
8 |
import json
|
9 |
import gradio as gr
|
10 |
|
@@ -14,16 +11,6 @@ from langchain_huggingface import HuggingFaceEndpoint
|
|
14 |
from langchain_core.prompts import PromptTemplate
|
15 |
|
16 |
|
17 |
-
# Add system template for RAG application
|
18 |
-
PROMPT_TEMPLATE = """
|
19 |
-
You are an assistant for question-answering tasks. Use the following pieces of context to answer the question at the end.
|
20 |
-
If you don't know the answer, just say that you don't know, don't try to make up an answer. Keep the answer concise.
|
21 |
-
Question: {question}
|
22 |
-
Context: {context}
|
23 |
-
Helpful Answer:
|
24 |
-
"""
|
25 |
-
|
26 |
-
|
27 |
# Initialize langchain LLM chain
|
28 |
def initialize_llmchain(
|
29 |
llm_model,
|
@@ -37,22 +24,11 @@ def initialize_llmchain(
|
|
37 |
"""Initialize Langchain LLM chain"""
|
38 |
|
39 |
progress(0.1, desc="Initializing HF tokenizer...")
|
40 |
-
# HuggingFaceHub uses HF inference endpoints
|
41 |
progress(0.5, desc="Initializing HF Hub...")
|
42 |
-
# Use of trust_remote_code as model_kwargs
|
43 |
-
# Warning: langchain issue
|
44 |
-
# URL: https://github.com/langchain-ai/langchain/issues/6080
|
45 |
-
|
46 |
-
# if 'Llama' in llm_model:
|
47 |
-
# task = "conversational"
|
48 |
-
# else:
|
49 |
-
# task = "text-generation"
|
50 |
-
# print(f"Task: {task}")
|
51 |
|
52 |
llm = HuggingFaceEndpoint(
|
53 |
repo_id=llm_model,
|
54 |
task="text-generation",
|
55 |
-
#task="conversational",
|
56 |
provider="hf-inference",
|
57 |
temperature=temperature,
|
58 |
max_new_tokens=max_tokens,
|
@@ -62,18 +38,20 @@ def initialize_llmchain(
|
|
62 |
|
63 |
progress(0.75, desc="Defining buffer memory...")
|
64 |
memory = ConversationBufferMemory(
|
65 |
-
memory_key="chat_history",
|
|
|
|
|
66 |
)
|
67 |
-
|
68 |
-
retriever = vector_db.as_retriever()
|
69 |
|
70 |
progress(0.8, desc="Defining retrieval chain...")
|
71 |
with open('prompt_template.json', 'r') as file:
|
72 |
-
|
73 |
prompt_template = system_prompt["prompt"]
|
74 |
rag_prompt = PromptTemplate(
|
75 |
template=prompt_template, input_variables=["context", "question"]
|
76 |
)
|
|
|
77 |
qa_chain = ConversationalRetrievalChain.from_llm(
|
78 |
llm,
|
79 |
retriever=retriever,
|
@@ -81,17 +59,16 @@ def initialize_llmchain(
|
|
81 |
memory=memory,
|
82 |
combine_docs_chain_kwargs={"prompt": rag_prompt},
|
83 |
return_source_documents=True,
|
84 |
-
# return_generated_question=False,
|
85 |
verbose=False,
|
86 |
)
|
87 |
-
progress(0.9, desc="Done!")
|
88 |
|
|
|
89 |
return qa_chain
|
90 |
|
91 |
|
|
|
92 |
def format_chat_history(message, chat_history):
|
93 |
-
"""Format chat history for
|
94 |
-
|
95 |
formatted_chat_history = []
|
96 |
for user_message, bot_message in chat_history:
|
97 |
formatted_chat_history.append(f"User: {user_message}")
|
@@ -99,27 +76,22 @@ def format_chat_history(message, chat_history):
|
|
99 |
return formatted_chat_history
|
100 |
|
101 |
|
|
|
102 |
def invoke_qa_chain(qa_chain, message, history):
|
103 |
"""Invoke question-answering chain"""
|
104 |
-
|
105 |
formatted_chat_history = format_chat_history(message, history)
|
106 |
-
# print("formatted_chat_history",formatted_chat_history)
|
107 |
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
)
|
112 |
|
113 |
response_sources = response["source_documents"]
|
114 |
-
|
115 |
response_answer = response["answer"]
|
116 |
-
if response_answer.find("Helpful Answer:") != -1:
|
117 |
-
response_answer = response_answer.split("Helpful Answer:")[-1]
|
118 |
-
|
119 |
-
# Append user message and response to chat history
|
120 |
-
new_history = history + [(message, response_answer)]
|
121 |
|
122 |
-
#
|
123 |
-
|
|
|
124 |
|
|
|
125 |
return qa_chain, new_history, response_sources
|
|
|
2 |
LLM chain retrieval
|
3 |
"""
|
4 |
|
|
|
|
|
|
|
5 |
import json
|
6 |
import gradio as gr
|
7 |
|
|
|
11 |
from langchain_core.prompts import PromptTemplate
|
12 |
|
13 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
# Initialize langchain LLM chain
|
15 |
def initialize_llmchain(
|
16 |
llm_model,
|
|
|
24 |
"""Initialize Langchain LLM chain"""
|
25 |
|
26 |
progress(0.1, desc="Initializing HF tokenizer...")
|
|
|
27 |
progress(0.5, desc="Initializing HF Hub...")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
28 |
|
29 |
llm = HuggingFaceEndpoint(
|
30 |
repo_id=llm_model,
|
31 |
task="text-generation",
|
|
|
32 |
provider="hf-inference",
|
33 |
temperature=temperature,
|
34 |
max_new_tokens=max_tokens,
|
|
|
38 |
|
39 |
progress(0.75, desc="Defining buffer memory...")
|
40 |
memory = ConversationBufferMemory(
|
41 |
+
memory_key="chat_history",
|
42 |
+
output_key="answer",
|
43 |
+
return_messages=True,
|
44 |
)
|
45 |
+
retriever = vector_db.as_retriever(search_type="similarity", search_kwargs={'k': top_k})
|
|
|
46 |
|
47 |
progress(0.8, desc="Defining retrieval chain...")
|
48 |
with open('prompt_template.json', 'r') as file:
|
49 |
+
system_prompt = json.load(file)
|
50 |
prompt_template = system_prompt["prompt"]
|
51 |
rag_prompt = PromptTemplate(
|
52 |
template=prompt_template, input_variables=["context", "question"]
|
53 |
)
|
54 |
+
|
55 |
qa_chain = ConversationalRetrievalChain.from_llm(
|
56 |
llm,
|
57 |
retriever=retriever,
|
|
|
59 |
memory=memory,
|
60 |
combine_docs_chain_kwargs={"prompt": rag_prompt},
|
61 |
return_source_documents=True,
|
|
|
62 |
verbose=False,
|
63 |
)
|
|
|
64 |
|
65 |
+
progress(0.9, desc="Done!")
|
66 |
return qa_chain
|
67 |
|
68 |
|
69 |
+
# Format chat history
|
70 |
def format_chat_history(message, chat_history):
|
71 |
+
"""Format chat history for LLM"""
|
|
|
72 |
formatted_chat_history = []
|
73 |
for user_message, bot_message in chat_history:
|
74 |
formatted_chat_history.append(f"User: {user_message}")
|
|
|
76 |
return formatted_chat_history
|
77 |
|
78 |
|
79 |
+
# Invoke QA chain with history
|
80 |
def invoke_qa_chain(qa_chain, message, history):
|
81 |
"""Invoke question-answering chain"""
|
|
|
82 |
formatted_chat_history = format_chat_history(message, history)
|
|
|
83 |
|
84 |
+
response = qa_chain.invoke({
|
85 |
+
"question": message,
|
86 |
+
"chat_history": formatted_chat_history,
|
87 |
+
})
|
88 |
|
89 |
response_sources = response["source_documents"]
|
|
|
90 |
response_answer = response["answer"]
|
|
|
|
|
|
|
|
|
|
|
91 |
|
92 |
+
# Clean up if "Helpful Answer:" is included
|
93 |
+
if "Helpful Answer:" in response_answer:
|
94 |
+
response_answer = response_answer.split("Helpful Answer:")[-1].strip()
|
95 |
|
96 |
+
new_history = history + [(message, response_answer)]
|
97 |
return qa_chain, new_history, response_sources
|