Spaces:
Sleeping
Sleeping
lalanikarim
commited on
Commit
•
5f3a384
1
Parent(s):
784a3bb
rearranged code. added inline comments.
Browse files- .gitignore +1 -0
- main.py +96 -29
.gitignore
CHANGED
@@ -1,3 +1,4 @@
|
|
|
|
1 |
# Byte-compiled / optimized / DLL files
|
2 |
__pycache__/
|
3 |
*.py[cod]
|
|
|
1 |
+
models/**
|
2 |
# Byte-compiled / optimized / DLL files
|
3 |
__pycache__/
|
4 |
*.py[cod]
|
main.py
CHANGED
@@ -3,56 +3,123 @@ from langchain.llms import LlamaCpp
|
|
3 |
from langchain.prompts import PromptTemplate
|
4 |
from langchain.chains import LLMChain
|
5 |
from langchain.callbacks.manager import CallbackManager
|
6 |
-
from langchain.callbacks.
|
7 |
-
|
8 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
9 |
def main():
|
|
|
|
|
10 |
st.set_page_config(
|
11 |
page_title="Your own Chat!"
|
12 |
)
|
|
|
|
|
13 |
st.header("Your own Chat!")
|
14 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
15 |
if "messages" not in st.session_state:
|
16 |
-
st.session_state.messages = [
|
|
|
|
|
17 |
|
|
|
|
|
18 |
for message in st.session_state.messages:
|
19 |
with st.chat_message(message["role"]):
|
20 |
st.markdown(message["content"])
|
21 |
|
22 |
-
|
23 |
-
|
24 |
-
llm = LlamaCpp(
|
25 |
-
model_path="mistral-7b-instruct-v0.1.Q4_0.gguf",
|
26 |
-
temperature=0,
|
27 |
-
max_tokens=512,
|
28 |
-
top_p=1,
|
29 |
-
callback_manager=callback_manager,
|
30 |
-
verbose=True,
|
31 |
-
)
|
32 |
|
33 |
-
|
34 |
-
You are a funny AI bot who answers questions in a couple of lines.
|
35 |
-
|
36 |
-
{question}
|
37 |
-
"""
|
38 |
-
|
39 |
-
prompt = PromptTemplate(template=template, input_variables=["question"])
|
40 |
-
|
41 |
-
llm_chain = LLMChain(prompt=prompt, llm=llm)
|
42 |
-
|
43 |
-
if prompt := st.chat_input("Your message here", key="user_input"):
|
44 |
st.session_state.messages.append(
|
45 |
-
{"role": "user", "content":
|
46 |
)
|
47 |
|
|
|
48 |
with st.chat_message("user"):
|
49 |
-
st.markdown(
|
50 |
-
|
51 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
52 |
st.session_state.messages.append(
|
53 |
{"role": "assistant", "content": response}
|
54 |
)
|
55 |
|
|
|
56 |
with st.chat_message("assistant"):
|
57 |
st.markdown(response)
|
58 |
|
|
|
3 |
from langchain.prompts import PromptTemplate
|
4 |
from langchain.chains import LLMChain
|
5 |
from langchain.callbacks.manager import CallbackManager
|
6 |
+
from langchain.callbacks.base import BaseCallbackHandler
|
7 |
+
|
8 |
+
|
9 |
+
# StreamHandler to intercept streaming output from the LLM.
|
10 |
+
# This makes it appear that the Language Model is "typing"
|
11 |
+
# in realtime.
|
12 |
+
class StreamHandler(BaseCallbackHandler):
|
13 |
+
def __init__(self, container, initial_text=""):
|
14 |
+
self.container = container
|
15 |
+
self.text = initial_text
|
16 |
+
|
17 |
+
def on_llm_new_token(self, token: str, **kwargs) -> None:
|
18 |
+
self.text += token
|
19 |
+
self.container.markdown(self.text)
|
20 |
+
|
21 |
+
|
22 |
+
# The main loop of the Streamlit Application. This is not a typical main()
|
23 |
+
# function. Streamlit runs this code in its entirety everytime any inputs
|
24 |
+
# change on the webpage.
|
25 |
+
#
|
26 |
+
# P.S.: Initializing LLM and Langchain inside here seem counterproductive.
|
27 |
+
# Hopefully there is a better prescribed way to initialize and manage expensive
|
28 |
+
# resources and reference them within here. But for the sake of example, let's
|
29 |
+
# not worry about that now.
|
30 |
def main():
|
31 |
+
|
32 |
+
# Set the webpage title
|
33 |
st.set_page_config(
|
34 |
page_title="Your own Chat!"
|
35 |
)
|
36 |
+
|
37 |
+
# Create a header element
|
38 |
st.header("Your own Chat!")
|
39 |
|
40 |
+
# This sets the LLM's personality.
|
41 |
+
# The initial personality privided is basic.
|
42 |
+
# Try something interesting and notice how the LLM responses are affected.
|
43 |
+
system_prompt = st.text_area(
|
44 |
+
label="System Prompt",
|
45 |
+
value="You are a helpful AI assistant who answers questions in short sentences.",
|
46 |
+
key="system_prompt")
|
47 |
+
|
48 |
+
# We store the conversation in the session state.
|
49 |
+
# This will be used to render the chat conversation.
|
50 |
+
# We initialize it with the first message we want to be greeted with.
|
51 |
if "messages" not in st.session_state:
|
52 |
+
st.session_state.messages = [
|
53 |
+
{"role": "assistant", "content": "How may I help you today?"}
|
54 |
+
]
|
55 |
|
56 |
+
# We loop through each message in the session state and render it as
|
57 |
+
# a chat message.
|
58 |
for message in st.session_state.messages:
|
59 |
with st.chat_message(message["role"]):
|
60 |
st.markdown(message["content"])
|
61 |
|
62 |
+
# We take questions/instructions from the chat input to pass to the LLM
|
63 |
+
if user_prompt := st.chat_input("Your message here", key="user_input"):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
64 |
|
65 |
+
# Add our input to the session state
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
66 |
st.session_state.messages.append(
|
67 |
+
{"role": "user", "content": user_prompt}
|
68 |
)
|
69 |
|
70 |
+
# Add our input to the chat window
|
71 |
with st.chat_message("user"):
|
72 |
+
st.markdown(user_prompt)
|
73 |
+
# A stream handler to direct streaming output on the chat screen.
|
74 |
+
# This will need to be handled somewhat differently.
|
75 |
+
# But it demonstrates what potential it carries.
|
76 |
+
stream_handler = StreamHandler(st.empty())
|
77 |
+
|
78 |
+
# Callback manager is a way to intercept streaming output from the
|
79 |
+
# LLM and take some action on it. Here we are giving it our custom
|
80 |
+
# stream handler to make it appear as if the LLM is typing the
|
81 |
+
# responses in real time.
|
82 |
+
callback_manager = CallbackManager([stream_handler])
|
83 |
+
|
84 |
+
# We initialize the quantized LLM from a local path.
|
85 |
+
# Currently most parameters are fixed but we can make them
|
86 |
+
# configurable.
|
87 |
+
llm = LlamaCpp(
|
88 |
+
model_path="models/mistral-7b-instruct-v0.1.Q4_0.gguf",
|
89 |
+
temperature=0,
|
90 |
+
max_tokens=512,
|
91 |
+
top_p=1,
|
92 |
+
callback_manager=callback_manager,
|
93 |
+
verbose=True,
|
94 |
+
)
|
95 |
+
|
96 |
+
# Template for the prompt. I am still trying to figure out what exactly
|
97 |
+
# is needed here and if we need to have parameters etc. This may
|
98 |
+
# ultimately be dictated by the model you use.
|
99 |
+
template = """
|
100 |
+
{}
|
101 |
+
|
102 |
+
{}
|
103 |
+
""".format(system_prompt, "{question}")
|
104 |
+
|
105 |
+
# We create a prompt from the template so we can use it with langchain
|
106 |
+
prompt = PromptTemplate(template=template, input_variables=["question"])
|
107 |
+
|
108 |
+
# We create an llm chain with our llm and prompt
|
109 |
+
llm_chain = LLMChain(prompt=prompt, llm=llm)
|
110 |
+
|
111 |
+
# Pass our input to the llm chain and capture the final responses.
|
112 |
+
# It is worth noting that the Stream Handler is already receiving the
|
113 |
+
# streaming response as the llm is generating. We get our response
|
114 |
+
# here once the llm has finished generating the complete response.
|
115 |
+
response = llm_chain.run(user_prompt)
|
116 |
+
|
117 |
+
# Add the response to the session state
|
118 |
st.session_state.messages.append(
|
119 |
{"role": "assistant", "content": response}
|
120 |
)
|
121 |
|
122 |
+
# Add the response to the chat window
|
123 |
with st.chat_message("assistant"):
|
124 |
st.markdown(response)
|
125 |
|