-
Notifications
You must be signed in to change notification settings - Fork 10
/
signals_app.py
537 lines (449 loc) · 21.3 KB
/
signals_app.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
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
162
163
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
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
import copy
import json
import os
import uuid
from datetime import datetime
from typing import Union
import openai
import s3fs
import streamlit as st
from dotenv import load_dotenv
from genai import FunctionTemplate
from genai import MessageTemplate
from genai.eyfs import TextGenerator
from genai.message_history import InMemoryMessageHistory
from genai.message_history import TokenCounter
load_dotenv()
selected_model = "gpt-4-1106-preview"
# selected_model = "gpt-3.5-turbo-1106"
# selected_model = "gpt-4"
temperature = 0.000001
CHECK_COSTS = False
# Paths to prompts
PROMPT_PATH = "src/genai/sandbox/signals/data/"
PATH_SIGNALS_DATA = PROMPT_PATH + "signals_2024.json"
PATH_SYSTEM = PROMPT_PATH + "00_system.jsonl"
PATH_INTRO = PROMPT_PATH + "01_intro.jsonl"
PATH_ACTIONS = PROMPT_PATH + "intent_actions.json"
PATH_ILLUSTRATIONS = "src/genai/sandbox/signals/illustrations/"
# Top signal function
path_func_top_signal = PROMPT_PATH + "func_top_signal.json"
path_prompt_top_signal = PROMPT_PATH + "prompt_top_signal.jsonl"
# Top three signals function
path_func_top_three_signals = PROMPT_PATH + "func_top_three_signals.json"
path_prompt_top_three_signals = PROMPT_PATH + "prompt_top_three_signals.jsonl"
# Intent detection function
path_func_intent = PROMPT_PATH + "func_intent.json"
path_prompt_intent = PROMPT_PATH + "prompt_intent.jsonl"
# Prompt: Impact on the user
path_prompt_impact = PROMPT_PATH + "02_signal_impact.jsonl"
# Prompt: Summary of different signals
path_prompt_choice = PROMPT_PATH + "03_signal_choice.jsonl"
# Prompt: Following up on user's question
path_prompt_following_up = PROMPT_PATH + "04_follow_up.jsonl"
aws_key = os.environ["AWS_ACCESS_KEY_ID"]
aws_secret = os.environ["AWS_SECRET_ACCESS_KEY"]
s3_path = os.environ["S3_BUCKET"]
def auth_openai() -> None:
"""Authenticate with OpenAI."""
try:
openai.api_key = os.environ["OPENAI_API_KEY"]
except Exception:
openai.api_key = st.secrets["OPENAI_API_KEY"]
def read_jsonl(path: str) -> list:
"""Read a JSONL file."""
with open(path, "r") as f:
return [json.loads(line) for line in f.readlines()]
def generate_signals_texts(signals_data: dict, chosen_signals: list = None) -> str:
"""
Generate a description of the signals.
Args:
signals_data (dict): A dictionary of signals data.
chosen_signals (list, optional): A list of signals to include in the description. Defaults to None.
Returns:
str: A description of the signals.
"""
signals = [signal["short_name"] for signal in signals_data]
signals_titles = [signal["title"] for signal in signals_data]
signals_summaries = [signal["summary"] for signal in signals_data]
if chosen_signals is None:
chosen_signals = signals
# Combine titles and summaries into a single string
signals_description = ""
for short_name, title, summary in zip(signals, signals_titles, signals_summaries):
if short_name in chosen_signals:
signals_description += f"Signal '{short_name}': {title}\n{summary}\n\n"
return signals_description
def generate_action_texts(action_data: dict, active_signal: str = None) -> str:
"""
Generate a description of the actions.
Args:
action_data (dict): A dictionary of actions data.
active_signal (str, optional): The active signal. Defaults to None.
Returns:
str: A description of the actions.
"""
actions = [a["name"] for a in action_data]
action_descriptions = [a["description"] for a in action_data]
action_text = ""
for name, description in zip(actions, action_descriptions):
if (name != "following_up") or (active_signal is None):
action_text += f"Action '{name}': {description}\n\n"
else:
action_text += f"Action '{name}': User is following up with another question about the {active_signal} signal that's being discussed just now.\n\n" # noqa: B950
return action_text
# Prepare the data
signals_data = json.load(open(PATH_SIGNALS_DATA, "r"))
signals_dict = {s["short_name"]: s for s in signals_data}
signals_descriptions = generate_signals_texts(signals_data)
signals = [s["short_name"] for s in signals_data]
actions_data = json.load(open(PATH_ACTIONS, "r"))
actions_descriptions = generate_action_texts(actions_data)
actions = [a["name"] for a in actions_data]
def predict_intent(user_message: str, active_signal: str) -> str:
"""Detect the intent of the user's message.
Args:
user_message (str): The user's message.
messages (list): The history of messages.
Returns:
str: The intent of the user's message. Possible outputs are:
- "explain": The user wants to know more about a signal.
- "more_signals": The user wants to know more about a signal.
- "follow_up": The user wants to know more about a signal.
- "next_steps": The user wants to know more about a signal.
- "none": The user's message does not match any intent.
"""
func_intent = json.loads(open(path_func_intent).read())
message_history = [MessageTemplate.load(m) for m in st.session_state.messages]
message = MessageTemplate.load(path_prompt_intent)
all_messages = message_history + [message]
function = FunctionTemplate.load(func_intent)
response = TextGenerator.generate(
model=selected_model,
temperature=temperature,
messages=all_messages,
message_kwargs={
"intents": generate_action_texts(actions_data, active_signal=active_signal),
"user_input": user_message,
},
stream=False,
functions=[function.to_prompt()],
function_call={"name": "predict_intent"},
)
intent = json.loads(response["choices"][0]["message"]["function_call"]["arguments"])
if CHECK_COSTS:
cost_input = (
TokenCounter._count_tokens_from_messages(
messages=[m.to_prompt() for m in all_messages], model_name=selected_model
)
* 0.01
/ 1000
)
print(f"Intent cost: {cost_input}") # noqa: T001
return intent["prediction"]
def predict_top_signal(user_message: str, signals: list) -> str:
"""Predict the top signal from the user's message.
Args:
user_message (str): The user's message.
Returns:
str: The top signal from the user's message.
"""
# Function call
func_top_signal = json.loads(open(path_func_top_signal).read())
func_top_signal["parameters"]["properties"]["prediction"]["enum"] = signals
message = MessageTemplate.load(path_prompt_top_signal)
function = FunctionTemplate.load(func_top_signal)
response = TextGenerator.generate(
model=selected_model,
temperature=temperature,
messages=[message],
message_kwargs={"signals": signals_descriptions, "user_input": user_message},
stream=False,
functions=[function.to_prompt()],
function_call={"name": "predict_top_signal"},
)
top_signal = json.loads(response["choices"][0]["message"]["function_call"]["arguments"])
return top_signal["prediction"]
def predict_top_three_signals(user_message: str, allowed_signals: list) -> list:
"""Predict the top signal from the user's message.
Args:
user_message (str): The user's message.
Returns:
str: The top signal from the user's message.
"""
# Function call
func_top_signals = json.loads(open(path_func_top_three_signals).read())
func_top_signals["parameters"]["properties"]["prediction"]["items"]["enum"] = allowed_signals
message = MessageTemplate.load(path_prompt_top_three_signals)
function_top_three = FunctionTemplate.load(func_top_signals)
signals_descriptions_ = generate_signals_texts(signals_data, allowed_signals)
response = TextGenerator.generate(
model=selected_model,
temperature=temperature,
messages=[message],
message_kwargs={"signals": signals_descriptions_, "user_input": user_message},
stream=False,
functions=[function_top_three.to_prompt()],
function_call={"name": "predict_top_signals"},
)
top_signals = json.loads(response["choices"][0]["message"]["function_call"]["arguments"])
return top_signals["prediction"]
def signals_bot() -> None:
"""Explain me a concept like I'm 3."""
# Define custom CSS
custom_css = """
<style>
/* Adjust the selector as needed */
.stHeadingContainer {
margin-top: -100px; /* Reduce the top margin */
}
#MainMenu {visibility: hidden;}
footer {visibility: hidden;}
header {
visibility: hidden
}
[data-testid="chatAvatarIcon-assistant"] {
background-color: #A59BEE !important;
}
[data-testid="chatAvatarIcon-user"] {
background-color: #97D9E3 !important;
}
html, body, [class*="st-"] {
font-family: 'Averta', 'Avenir', "Source Sans Pro", sans-serif;
}
.stChatFloatingInputContainer {
position: relative;
}
</style>
"""
# Apply the custom CSS
st.markdown(custom_css, unsafe_allow_html=True)
st.title("Personalised futures: let our chatbot guide you through our Future Signals for 2024")
st.markdown(
"Undoubtedly, the rise of generative artificial intelligence (AI) has been one of the main trends of 2023, with ChatGPT chosen as [the word of the year](https://www.economist.com/culture/2023/12/07/our-word-of-the-year-for-2023) by *The Economist*. Reflecting on this trend, we have built an experimental generative AI chatbot of our own to help you engage more deeply with our [Signals for 2024](https://www.nesta.org.uk/feature/future-signals-2024/).\n\nThis is an experiment in creating a more interactive reading experience using 2023's big new technology. **Scroll down** to meet our chatbot, Scout, which will try to relate this year's Signals to you and your life. You can provide a little information about yourself and Scout will try to come up with ways that these Signals might be relevant to you.\n\nScout also provides a signal of the potential new ways we might interact with information in the future, with customised bots helping us explore and synthesise reams of written text, data, charts and videos to find what matters the most to us.\n\n**Guidance**\n\n*The chatbot uses OpenAI's GPT-4, a cutting-edge AI model. Nesta does not save the conversations, and OpenAI claims to delete all data in 30 days. **Nonetheless, please do not share any information that could identify you or that is sensitive or confidential.** Please remember, this is an experimental chatbot; it can make mistakes and 'hallucinate' - [another word of the year](https://www.cam.ac.uk/research/news/cambridge-dictionary-names-hallucinate-word-of-the-year-2023) - or show biases despite our efforts to instruct it to be inclusive and sensitive. After trying out the chatbot, we invite you to leave us feedback using [this form](https://forms.gle/UWcnpgKg9WG7JmPt5).*\n\n**Meet Scout...**", # noqa: B950
unsafe_allow_html=True, # noqa: B950
) # noqa: B950
# First time running the app
if "messages" not in st.session_state:
# Record of messages to display on the app
st.session_state.messages = []
# Record of messages to send to the LLM
st.session_state["memory"] = InMemoryMessageHistory()
st.session_state["messages_intent"] = []
st.session_state["messages_signal"] = []
# Keep track of which state we're in
st.session_state.state = "start"
# Fetch system and introduction messages
st.session_state.signals = []
# Add system message to the history
system_message = read_jsonl(PATH_SYSTEM)[0]
system_message = MessageTemplate.load(system_message)
system_message.format_message(**{"signals": signals_descriptions})
st.session_state["memory"].add_message(system_message.to_prompt())
# Add the intro messages
intro_messages = read_jsonl(PATH_INTRO)
for m in intro_messages:
st.session_state.messages.append(m)
st.session_state["memory"].add_message(m)
st.session_state["messages_intent"].append("intro")
st.session_state["messages_signal"].append("none")
# Keep count of the number of unique sessions
timestamp = current_time()
st.session_state["session_log"] = f"{timestamp}-{str(uuid.uuid4())}"
write_to_s3(
key=aws_key,
secret=aws_secret,
s3_path=f"{s3_path}/session-logs-signals",
filename="session_counter",
data={"session": st.session_state["session_log"], "time": timestamp},
how="a",
)
# Display chat messages on app rerun
for i, message in enumerate(st.session_state.messages):
with st.chat_message(message["role"]):
if st.session_state["messages_intent"][i] == "new_signal":
signal_to_explain = st.session_state["messages_signal"][i]
st.image(
PATH_ILLUSTRATIONS + signals_dict[signal_to_explain]["img"],
caption="Illustration by Chen Wu",
use_column_width=True,
)
st.markdown(message["content"])
# Get user message
user_message = st.chat_input("")
if user_message:
updated_css = """
<style>
.stChatFloatingInputContainer {
position: fixed;
}
</style>
"""
st.markdown(updated_css, unsafe_allow_html=True)
# Display user message
with st.chat_message("user"):
st.markdown(user_message)
st.session_state.messages.append({"role": "user", "content": user_message})
st.session_state["memory"].add_message({"role": "user", "content": user_message})
st.session_state["messages_intent"].append("user")
st.session_state["messages_signal"].append("none")
if st.session_state.state == "start":
intent = "new_signal"
st.session_state.user_info = user_message
st.session_state.state = "chatting"
else:
intent = predict_intent(user_message, active_signal=st.session_state.active_signal)
if intent == "new_signal":
# Filter out signals that have already been covered
allowed_signals = [s for s in signals if s not in st.session_state.signals]
# Determine the most relevant signal to explain
signal_to_explain = predict_top_signal(user_message, allowed_signals)
# Keep track of already discussed signals
st.session_state.signals.append(signal_to_explain)
st.session_state.active_signal = signal_to_explain
# Generate a message about the signal
instruction = MessageTemplate.load(path_prompt_impact)
message_history = st.session_state["memory"].get_messages(max_tokens=3000) + [instruction]
with st.chat_message("assistant"):
# Show the signal image
st.image(
PATH_ILLUSTRATIONS + signals_dict[signal_to_explain]["img"],
caption="Illustration by Chen Wu",
use_column_width=True,
)
# Type the response
full_response = llm_call(
selected_model,
temperature,
messages=message_history,
messages_placeholders={
"signal": signals_dict[signal_to_explain]["full_text"],
"user_input": st.session_state.user_info,
},
)
st.session_state.messages.append({"role": "assistant", "content": full_response})
st.session_state["memory"].add_message({"role": "assistant", "content": full_response})
st.session_state["messages_intent"].append(copy.deepcopy(intent))
st.session_state["messages_signal"].append(copy.deepcopy(signal_to_explain))
elif intent == "more_signals":
# Filter out signals that have already been covered
allowed_signals = [s for s in signals if s not in st.session_state.signals]
# Determine the top three signals to explain
top_signals = predict_top_three_signals(st.session_state.user_info, allowed_signals)
top_signals_text = generate_signals_texts(signals_data, top_signals)
# Generate a message about the three signals
instruction = MessageTemplate.load(path_prompt_choice)
message_history = st.session_state["memory"].get_messages(max_tokens=3000) + [instruction]
with st.chat_message("assistant"):
full_response = llm_call(
selected_model,
temperature,
messages=message_history,
messages_placeholders={"signals": top_signals_text, "user_input": st.session_state.user_info},
)
st.session_state.messages.append({"role": "assistant", "content": full_response})
st.session_state["memory"].add_message({"role": "assistant", "content": full_response})
st.session_state["messages_intent"].append(copy.deepcopy(intent))
st.session_state["messages_signal"].append("none")
elif intent == "following_up":
# Generate follow up message
instruction = MessageTemplate.load(path_prompt_following_up)
message_history = st.session_state["memory"].get_messages(max_tokens=3000) + [instruction]
with st.chat_message("assistant"):
full_response = llm_call(
selected_model,
temperature,
messages=message_history,
messages_placeholders={
"signal": signals_dict[st.session_state.active_signal]["full_text"],
"user_input": user_message,
},
)
st.session_state.messages.append({"role": "assistant", "content": full_response})
st.session_state["memory"].add_message({"role": "assistant", "content": full_response})
st.session_state["messages_intent"].append(copy.deepcopy(intent))
st.session_state["messages_signal"].append("none")
if CHECK_COSTS:
# Keep track of the number of messages
def transform_message(m: Union[MessageTemplate, dict]) -> dict:
"""Transform all messages to dictionary format (quick hack)"""
try:
return m.to_prompt()
except AttributeError:
return m
cost_input = (
TokenCounter._count_tokens_from_messages(
messages=[transform_message(m) for m in message_history], model_name=selected_model
)
* 0.01
/ 1000
)
cost_output = (
TokenCounter._count_tokens_from_messages(
messages=[{"role": "assistant", "content": full_response}], model_name=selected_model
)
* 0.03
/ 1000
)
cost_total = cost_input + cost_output
print(f"Total cost: {cost_total}") # noqa: T001
write_to_s3(
key=aws_key,
secret=aws_secret,
s3_path=f"{s3_path}/session-logs-signals",
filename="message_counter",
data={
"session": st.session_state["session_log"],
"time": current_time(),
"intent": intent,
"signal": st.session_state.active_signal,
},
how="a",
)
def llm_call(selected_model: str, temperature: float, messages: MessageTemplate, messages_placeholders: dict) -> str:
"""Call the LLM"""
message_placeholder = st.empty()
full_response = ""
for response in TextGenerator.generate(
model=selected_model,
temperature=temperature,
messages=messages,
message_kwargs=messages_placeholders,
stream=True,
):
full_response += response.choices[0].delta.get("content", "")
message_placeholder.markdown(full_response + "▌")
message_placeholder.markdown(full_response)
return full_response
def write_to_s3(key: str, secret: str, s3_path: str, filename: str, data: dict, how: str = "a") -> None:
"""Write data to a jsonl file in S3.
Parameters
----------
key
AWS access key ID.
secret
AWS secret access key.
s3_path
S3 bucket path.
filename
Name of the file to write to.
data
Data to write to the file.
how
How to write to the file. Default is "a" for append. Use "w" to overwrite.
"""
fs = s3fs.S3FileSystem(key=key, secret=secret)
with fs.open(f"{s3_path}/{filename}.jsonl", how) as f:
f.write(f"{json.dumps(data)}\n")
def current_time() -> str:
"""Return the current time as a string. Used as part of the session UUID."""
# Get current date and time
current_datetime = datetime.now()
# Convert to a long number format
datetime_string = current_datetime.strftime("%Y%m%d%H%M%S")
return datetime_string
def main() -> None:
"""Run the app."""
auth_openai()
signals_bot()
main()