fix gradio demo issue and not use chatbot component (#3)
Browse files- fix gradio demo issue and not use chatbot component (eb37cc8da5a2a15ba1535f80e1ced3bd59b95d14)
Co-authored-by: AK <[email protected]>
app.py
CHANGED
@@ -1,16 +1,8 @@
|
|
1 |
-
import
|
2 |
-
import
|
3 |
-
from openai import OpenAI
|
4 |
import time
|
5 |
import re
|
6 |
-
|
7 |
-
# Set up API key
|
8 |
-
API_KEY = os.getenv("API_KEY")
|
9 |
-
URL = os.getenv("URL")
|
10 |
-
client = OpenAI(
|
11 |
-
api_key=API_KEY,
|
12 |
-
base_url=URL
|
13 |
-
)
|
14 |
|
15 |
# Available models
|
16 |
MODELS = [
|
@@ -19,181 +11,196 @@ MODELS = [
|
|
19 |
"Meta-Llama-3.1-8B-Instruct"
|
20 |
]
|
21 |
|
22 |
-
|
23 |
-
|
24 |
-
"
|
25 |
-
"Greedy-Best-Score",
|
26 |
-
"Iterative-Refinement",
|
27 |
-
"Monte-Carlo-Tree-Search"
|
28 |
-
]
|
29 |
|
30 |
def chat_with_ai(message, chat_history, system_prompt):
|
31 |
messages = [
|
32 |
{"role": "system", "content": system_prompt},
|
33 |
]
|
34 |
-
|
35 |
-
for human, ai
|
36 |
messages.append({"role": "user", "content": human})
|
37 |
messages.append({"role": "assistant", "content": ai})
|
38 |
-
|
39 |
messages.append({"role": "user", "content": message})
|
40 |
-
|
41 |
return messages
|
42 |
|
43 |
-
def respond(message, chat_history, model, system_prompt, thinking_budget):
|
44 |
-
|
45 |
-
|
|
|
46 |
start_time = time.time()
|
47 |
-
|
48 |
-
|
|
|
|
|
49 |
model=model,
|
50 |
messages=messages,
|
51 |
-
stream=
|
52 |
-
)
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
|
|
|
|
|
|
|
|
|
59 |
answer_match = re.search(r'<answer>(.*?)</answer>', response, re.DOTALL)
|
60 |
reflection_match = re.search(r'<reflection>(.*?)</reflection>', response, re.DOTALL)
|
61 |
-
|
62 |
answer = answer_match.group(1).strip() if answer_match else ""
|
63 |
reflection = reflection_match.group(1).strip() if reflection_match else ""
|
64 |
-
|
65 |
-
# Remove answer, reflection, and final reward from the main response
|
66 |
-
response = re.sub(r'<answer>.*?</answer>', '', response, flags=re.DOTALL)
|
67 |
-
response = re.sub(r'<reflection>.*?</reflection>', '', response, flags=re.DOTALL)
|
68 |
-
response = re.sub(r'<reward>.*?</reward>\s*$', '', response, flags=re.DOTALL)
|
69 |
-
|
70 |
-
# Extract and display steps
|
71 |
steps = re.findall(r'<step>(.*?)</step>', response, re.DOTALL)
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
|
88 |
-
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
|
93 |
-
|
94 |
-
|
95 |
-
|
96 |
-
|
97 |
-
|
98 |
-
|
99 |
-
|
100 |
-
|
101 |
-
|
102 |
-
|
103 |
-
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
|
124 |
-
|
125 |
-
|
126 |
-
|
127 |
-
|
128 |
-
|
129 |
-
|
130 |
-
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
|
135 |
-
|
136 |
-
|
137 |
-
|
138 |
-
|
139 |
-
|
140 |
-
|
141 |
-
|
142 |
-
|
143 |
-
|
144 |
-
|
145 |
-
|
146 |
-
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
-
|
152 |
-
|
153 |
-
|
154 |
-
|
155 |
-
|
156 |
-
|
157 |
-
|
158 |
-
|
159 |
-
|
160 |
-
|
161 |
-
<reward> [Float between 0.0 and 1.0] </reward>
|
162 |
-
""",
|
163 |
-
height=200
|
164 |
)
|
165 |
-
|
166 |
-
|
167 |
-
|
168 |
-
|
169 |
-
|
170 |
-
|
171 |
-
|
172 |
-
|
173 |
-
|
174 |
-
|
175 |
-
|
176 |
-
|
177 |
-
|
178 |
-
|
179 |
-
|
180 |
-
|
181 |
-
|
182 |
-
|
183 |
-
|
184 |
-
|
185 |
-
|
186 |
-
|
187 |
-
|
188 |
-
|
189 |
-
|
190 |
-
|
191 |
-
|
192 |
-
|
193 |
-
|
194 |
-
|
195 |
-
|
196 |
-
|
197 |
-
|
198 |
-
|
199 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import openai
|
|
|
3 |
import time
|
4 |
import re
|
5 |
+
import os
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
6 |
|
7 |
# Available models
|
8 |
MODELS = [
|
|
|
11 |
"Meta-Llama-3.1-8B-Instruct"
|
12 |
]
|
13 |
|
14 |
+
def create_client(api_key):
|
15 |
+
openai.api_key = api_key
|
16 |
+
openai.api_base = "https://api.sambanova.ai/v1" # Fixed Base URL
|
|
|
|
|
|
|
|
|
17 |
|
18 |
def chat_with_ai(message, chat_history, system_prompt):
|
19 |
messages = [
|
20 |
{"role": "system", "content": system_prompt},
|
21 |
]
|
22 |
+
|
23 |
+
for human, ai in chat_history:
|
24 |
messages.append({"role": "user", "content": human})
|
25 |
messages.append({"role": "assistant", "content": ai})
|
26 |
+
|
27 |
messages.append({"role": "user", "content": message})
|
28 |
+
|
29 |
return messages
|
30 |
|
31 |
+
def respond(message, chat_history, model, system_prompt, thinking_budget, api_key):
|
32 |
+
print("Starting respond function...")
|
33 |
+
create_client(api_key) # Sets api_key and api_base globally
|
34 |
+
messages = chat_with_ai(message, chat_history, system_prompt.format(budget=thinking_budget))
|
35 |
start_time = time.time()
|
36 |
+
|
37 |
+
try:
|
38 |
+
print("Calling OpenAI API...")
|
39 |
+
completion = openai.ChatCompletion.create(
|
40 |
model=model,
|
41 |
messages=messages,
|
42 |
+
stream=False # Set to False for synchronous response
|
43 |
+
)
|
44 |
+
response = completion.choices[0].message['content']
|
45 |
+
thinking_time = time.time() - start_time
|
46 |
+
print("Response received from OpenAI API.")
|
47 |
+
yield response, thinking_time
|
48 |
+
except Exception as e:
|
49 |
+
error_message = f"Error: {str(e)}"
|
50 |
+
print(error_message)
|
51 |
+
yield error_message, time.time() - start_time
|
52 |
+
|
53 |
+
def parse_response(response):
|
54 |
answer_match = re.search(r'<answer>(.*?)</answer>', response, re.DOTALL)
|
55 |
reflection_match = re.search(r'<reflection>(.*?)</reflection>', response, re.DOTALL)
|
56 |
+
|
57 |
answer = answer_match.group(1).strip() if answer_match else ""
|
58 |
reflection = reflection_match.group(1).strip() if reflection_match else ""
|
59 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
60 |
steps = re.findall(r'<step>(.*?)</step>', response, re.DOTALL)
|
61 |
+
|
62 |
+
return answer, reflection, steps
|
63 |
+
|
64 |
+
def process_chat(message, history, model, system_prompt, thinking_budget, api_key):
|
65 |
+
print(f"Received message: {message}")
|
66 |
+
if not api_key:
|
67 |
+
print("API key missing")
|
68 |
+
return "Please provide your API Key before starting the chat."
|
69 |
+
|
70 |
+
try:
|
71 |
+
formatted_system_prompt = system_prompt.format(budget=thinking_budget)
|
72 |
+
except KeyError as e:
|
73 |
+
error_msg = f"System prompt missing placeholder: {str(e)}"
|
74 |
+
print(error_msg)
|
75 |
+
return error_msg
|
76 |
+
|
77 |
+
full_response = ""
|
78 |
+
thinking_time = 0
|
79 |
+
|
80 |
+
for response, elapsed_time in respond(message, history, model, formatted_system_prompt, thinking_budget, api_key):
|
81 |
+
print(f"Received response: {response}")
|
82 |
+
full_response = response
|
83 |
+
thinking_time = elapsed_time
|
84 |
+
|
85 |
+
if full_response.startswith("Error:"):
|
86 |
+
return full_response
|
87 |
+
|
88 |
+
answer, reflection, steps = parse_response(full_response)
|
89 |
+
|
90 |
+
formatted_response = f"**Answer:** {answer}\n\n**Reflection:** {reflection}\n\n**Thinking Steps:**\n"
|
91 |
+
for i, step in enumerate(steps, 1):
|
92 |
+
formatted_response += f"**Step {i}:** {step}\n"
|
93 |
+
|
94 |
+
formatted_response += f"\n**Thinking time:** {thinking_time:.2f} s"
|
95 |
+
|
96 |
+
print(f"Appended response: {formatted_response}")
|
97 |
+
history.append((message, formatted_response))
|
98 |
+
return formatted_response
|
99 |
+
|
100 |
+
# Define the default system prompt
|
101 |
+
default_system_prompt = """
|
102 |
+
You are a helpful assistant in normal conversation.
|
103 |
+
When given a problem to solve, you are an expert problem-solving assistant. Your task is to provide a detailed, step-by-step solution to a given question. Follow these instructions carefully:
|
104 |
+
|
105 |
+
1. Read the given question carefully and reset counter between <count> and </count> to {budget}
|
106 |
+
2. Generate a detailed, logical step-by-step solution.
|
107 |
+
3. Enclose each step of your solution within <step> and </step> tags.
|
108 |
+
4. You are allowed to use at most {budget} steps (starting budget), keep track of it by counting down within tags <count> </count>, STOP GENERATING MORE STEPS when hitting 0, you don't have to use all of them.
|
109 |
+
5. Do a self-reflection when you are unsure about how to proceed, based on the self-reflection and reward, decides whether you need to return to the previous steps.
|
110 |
+
6. After completing the solution steps, reorganize and synthesize the steps into the final answer within <answer> and </answer> tags.
|
111 |
+
7. Provide a critical, honest and subjective self-evaluation of your reasoning process within <reflection> and </reflection> tags.
|
112 |
+
8. Assign a quality score to your solution as a float between 0.0 (lowest quality) and 1.0 (highest quality), enclosed in <reward> and </reward> tags.
|
113 |
+
|
114 |
+
Example format:
|
115 |
+
<count> [starting budget] </count>
|
116 |
+
|
117 |
+
<step> [Content of step 1] </step>
|
118 |
+
<count> [remaining budget] </count>
|
119 |
+
|
120 |
+
<step> [Content of step 2] </step>
|
121 |
+
<reflection> [Evaluation of the steps so far] </reflection>
|
122 |
+
<reward> [Float between 0.0 and 1.0] </reward>
|
123 |
+
<count> [remaining budget] </count>
|
124 |
+
|
125 |
+
<step> [Content of step 3 or Content of some previous step] </step>
|
126 |
+
<count> [remaining budget] </count>
|
127 |
+
|
128 |
+
...
|
129 |
+
|
130 |
+
<step> [Content of final step] </step>
|
131 |
+
<count> [remaining budget] </count>
|
132 |
+
|
133 |
+
<answer> [Final Answer] </answer>
|
134 |
+
|
135 |
+
<reflection> [Evaluation of the solution] </reflection>
|
136 |
+
|
137 |
+
<reward> [Float between 0.0 and 1.0] </reward>
|
138 |
+
"""
|
139 |
+
|
140 |
+
with gr.Blocks() as demo:
|
141 |
+
gr.Markdown("# Llama3.1-Instruct-O1")
|
142 |
+
gr.Markdown("[Powered by Llama3.1 models through SN Cloud](https://sambanova.ai/fast-api?api_ref=907266)")
|
143 |
+
|
144 |
+
with gr.Row():
|
145 |
+
api_key = gr.Textbox(
|
146 |
+
label="API Key",
|
147 |
+
type="password",
|
148 |
+
placeholder="Enter your API key here"
|
|
|
|
|
|
|
|
|
149 |
)
|
150 |
+
|
151 |
+
with gr.Row():
|
152 |
+
model = gr.Dropdown(
|
153 |
+
choices=MODELS,
|
154 |
+
label="Select Model",
|
155 |
+
value=MODELS[0]
|
156 |
+
)
|
157 |
+
thinking_budget = gr.Slider(
|
158 |
+
minimum=1,
|
159 |
+
maximum=100,
|
160 |
+
value=10,
|
161 |
+
step=1,
|
162 |
+
label="Thinking Budget"
|
163 |
+
)
|
164 |
+
|
165 |
+
system_prompt = gr.Textbox(
|
166 |
+
label="System Prompt",
|
167 |
+
value=default_system_prompt,
|
168 |
+
lines=10
|
169 |
+
)
|
170 |
+
|
171 |
+
msg = gr.Textbox(
|
172 |
+
label="Type your message here...",
|
173 |
+
placeholder="Enter your message..."
|
174 |
+
)
|
175 |
+
submit = gr.Button("Submit")
|
176 |
+
clear = gr.Button("Clear Chat")
|
177 |
+
|
178 |
+
output = gr.Textbox(
|
179 |
+
label="Response",
|
180 |
+
lines=20,
|
181 |
+
interactive=False
|
182 |
+
)
|
183 |
+
|
184 |
+
# Initialize chat history
|
185 |
+
chat_history = []
|
186 |
+
|
187 |
+
def handle_submit(message, history, model, system_prompt, thinking_budget, api_key):
|
188 |
+
response = process_chat(message, history, model, system_prompt, thinking_budget, api_key)
|
189 |
+
return response
|
190 |
+
|
191 |
+
def handle_clear():
|
192 |
+
return ""
|
193 |
+
|
194 |
+
submit.click(
|
195 |
+
handle_submit,
|
196 |
+
inputs=[msg, gr.State(chat_history), model, system_prompt, thinking_budget, api_key],
|
197 |
+
outputs=output
|
198 |
+
)
|
199 |
+
|
200 |
+
clear.click(
|
201 |
+
lambda: "",
|
202 |
+
inputs=None,
|
203 |
+
outputs=output
|
204 |
+
)
|
205 |
+
|
206 |
+
demo.launch()
|