# -*- coding: utf-8 -*-
"""wiki_chat.ipynb
Automatically generated by Colaboratory.
Original file is located at
https://colab.research.google.com/drive/1P5rJeCXRSsDJw_1ksnHmodH6ng2Ot5NW
"""
# !pip install gradio
# !pip install -U sentence-transformers
# !pip install datasets
from azure_utils import AzureVoiceData
from polly_utils import PollyVoiceData, NEURAL_ENGINE
from langchain.prompts import PromptTemplate
from openai.error import AuthenticationError, InvalidRequestError, RateLimitError
import re
import sys
from io import StringIO
from threading import Lock
from langchain.llms import OpenAI
from langchain.chains.conversation.memory import ConversationBufferMemory
from langchain.agents import tool, load_tools, initialize_agent
from langchain.agents import Tool
from langchain import ConversationChain, LLMChain
import whisper
import warnings
import boto3
import datetime
from typing import Optional, Tuple
from contextlib import closing
# Console to variable
import io
import requests
import os
import gradio as gr
from sentence_transformers import SentenceTransformer, CrossEncoder, util
from torch import tensor as torch_tensor
from datasets import load_dataset
"""# import models"""
bi_encoder = SentenceTransformer(
'sentence-transformers/multi-qa-MiniLM-L6-cos-v1')
bi_encoder.max_seq_length = 256 # Truncate long passages to 256 tokens
# The bi-encoder will retrieve top_k documents. We use a cross-encoder, to re-rank the results list to improve the quality
cross_encoder = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')
"""# import datasets"""
dataset = load_dataset("gfhayworth/wiki_mini", split='train')
mypassages = list(dataset.to_pandas()['psg'])
dataset_embed = load_dataset("gfhayworth/wiki_mini_embed", split='train')
dataset_embed_pd = dataset_embed.to_pandas()
mycorpus_embeddings = torch_tensor(dataset_embed_pd.values)
def search(query, top_k=20, top_n=1):
question_embedding = bi_encoder.encode(query, convert_to_tensor=True)
question_embedding = question_embedding # .cuda()
hits = util.semantic_search(
question_embedding, mycorpus_embeddings, top_k=top_k)
hits = hits[0] # Get the hits for the first query
##### Re-Ranking #####
cross_inp = [[query, mypassages[hit['corpus_id']]] for hit in hits]
cross_scores = cross_encoder.predict(cross_inp)
# Sort results by the cross-encoder scores
for idx in range(len(cross_scores)):
hits[idx]['cross-score'] = cross_scores[idx]
hits = sorted(hits, key=lambda x: x['cross-score'], reverse=True)
predictions = hits[:top_n]
return predictions
# for hit in hits[0:3]:
# print("\t{:.3f}\t{}".format(hit['cross-score'], mypassages[hit['corpus_id']].replace("\n", " ")))
def get_text(qry):
predictions = search(qry)
prediction_text = []
for hit in predictions:
prediction_text.append("{}".format(mypassages[hit['corpus_id']]))
return prediction_text
@tool
def mysearch(query: str) -> str:
"""Query our own datasets.
"""
rslt = get_text(query)
return '\n'.join(rslt)
# mysearch("who is the best rapper in the world?")
# """# chat example"""
# def chat(message, history):
# history = history or []
# message = message.lower()
# responses = get_text(message)
# for response in responses:
# history.append((message, response))
# return history, history
# with gr.Blocks(css=CSS) as demo:
# history_state = gr.State()
# gr.Markdown('# WikiBot')
# title = 'Wikipedia Chatbot'
# description = 'chatbot with search on Wikipedia'
# with gr.Row():
# chatbot = gr.Chatbot()
# with gr.Row():
# message = gr.Textbox(label='Input your question here:',
# placeholder='How many countries are in Europe?',
# lines=1)
# submit = gr.Button(value='Send',
# variant='secondary').style(full_width=False)
# submit.click(chat,
# inputs=[message, history_state],
# outputs=[chatbot, history_state])
# gr.Examples(
# examples=["How many countries are in Europe?",
# "Was Roman Emperor Constantine I a Christian?",
# "Who is the best rapper in the world?"],
# inputs=message
# )
# demo.launch()
# OPENAI_API_KEY = "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"
# NOTE using locally
# AWS keys
aws_access_key_id = "AKIA3JRWKI2EOCFNK3FT"
aws_secret_access_key = "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"
aws_region_name = "us-east-2"
os.environ["AWS_ACCESS_KEY_ID"] = aws_access_key_id
os.environ["AWS_SECRET_ACCESS_KEY"] = aws_secret_access_key
os.environ["AWS_DEFAULT_REGION"] = aws_region_name
# exhumana api key
# todo: may need to pay to get one
os.environ['EXHUMAN_API_KEY'] = ''
# news, tmdb keys
os.environ["NEWS_API_KEY"] = ''
os.environ["TMDB_BEARER_TOKEN"] = ''
news_api_key = os.environ["NEWS_API_KEY"]
tmdb_bearer_token = os.environ["TMDB_BEARER_TOKEN"]
TOOLS_LIST = ['serpapi', 'wolfram-alpha', 'pal-math', 'pal-colored-objects', 'news-api', 'tmdb-api',
'open-meteo-api'] # 'google-search'
# TOOLS_DEFAULT_LIST = ['mysearch', 'serpapi', 'pal-math']
TOOLS_DEFAULT_LIST = [mysearch]
BUG_FOUND_MSG = "Congratulations, you've found a bug in this application!"
AUTH_ERR_MSG = "Please paste your OpenAI key from openai.com to use this application. It is not necessary to hit a button or key after pasting it."
MAX_TOKENS = 512
TEMPERATURE = 0
LOOPING_TALKING_HEAD = "videos/humancare.mp4"
TALKING_HEAD_WIDTH = "192"
MAX_TALKING_HEAD_TEXT_LENGTH = 155
# Pertains to Express-inator functionality
NUM_WORDS_DEFAULT = 0
MAX_WORDS = 400
FORMALITY_DEFAULT = "N/A"
TEMPERATURE_DEFAULT = 0.5
EMOTION_DEFAULT = "N/A"
LANG_LEVEL_DEFAULT = "N/A"
TRANSLATE_TO_DEFAULT = "N/A"
LITERARY_STYLE_DEFAULT = "N/A"
PROMPT_TEMPLATE = PromptTemplate(
input_variables=["original_words", "num_words", "formality",
"emotions", "lang_level", "translate_to", "literary_style"],
template="Restate {num_words}{formality}{emotions}{lang_level}{translate_to}{literary_style}the following: \n{original_words}\n",
)
POLLY_VOICE_DATA = PollyVoiceData()
AZURE_VOICE_DATA = AzureVoiceData()
VOICE_GENDER = 'Female' # "Male"
# Pertains to WHISPER functionality
WHISPER_DETECT_LANG = "Detect language"
# UNCOMMENT TO USE WHISPER
warnings.filterwarnings("ignore")
WHISPER_MODEL = whisper.load_model("tiny")
print("WHISPER_MODEL", WHISPER_MODEL)
# gradio settings
# css
CSS = ".gradio-container {background-color: lightgray}"
# placeholder for chat text input
PLACEHOLDER = "What is my plan benefit?",
# example questions
EXAMPLES = ["How many people live in Canada?",
"What is 2 to the 30th power?",
"If x+y=10 and x-y=4, what are x and y?",
"How much did it rain in SF today?",
"Get me information about the movie 'Avatar'",
"What are the top tech headlines in the US?",
"On the desk, you see two blue booklets, two purple booklets, and two yellow pairs of sunglasses - "
"if I remove all the pairs of sunglasses from the desk, how many purple items remain on it?"]
AUTHORS = """
This application, developed by Greg Hayworth, Srikanth Tangelloju, Lincoln Snyder, Michal Piekarczyk, and Xingde Jiang,
demonstrates a conversational agent implemented with OpenAI GPT-3.5 and LangChain.
When necessary, it leverages tools for complex math, searching the internet, and accessing news and weather.
Uses talking heads from Ex-Human.
For faster inference without waiting in queue, you may duplicate the space.
"""
# UNCOMMENT TO USE WHISPER
def transcribe(aud_inp, whisper_lang):
if aud_inp is None:
return ""
aud = whisper.load_audio(aud_inp)
aud = whisper.pad_or_trim(aud)
mel = whisper.log_mel_spectrogram(aud).to(WHISPER_MODEL.device)
_, probs = WHISPER_MODEL.detect_language(mel)
options = whisper.DecodingOptions(fp16=False)
if whisper_lang != WHISPER_DETECT_LANG:
whisper_lang_code = POLLY_VOICE_DATA.get_whisper_lang_code(
whisper_lang)
options = whisper.DecodingOptions(
fp16=False, language=whisper_lang_code)
result = whisper.decode(WHISPER_MODEL, mel, options)
print("result.text", result.text)
result_text = ""
if result and result.text:
result_text = result.text
return result_text
# Pertains to Express-inator functionality
def transform_text(desc, express_chain, num_words, formality,
anticipation_level, joy_level, trust_level,
fear_level, surprise_level, sadness_level, disgust_level, anger_level,
lang_level, translate_to, literary_style):
num_words_prompt = ""
if num_words and int(num_words) != 0:
num_words_prompt = "using up to " + str(num_words) + " words, "
# Change some arguments to lower case
formality = formality.lower()
anticipation_level = anticipation_level.lower()
joy_level = joy_level.lower()
trust_level = trust_level.lower()
fear_level = fear_level.lower()
surprise_level = surprise_level.lower()
sadness_level = sadness_level.lower()
disgust_level = disgust_level.lower()
anger_level = anger_level.lower()
formality_str = ""
if formality != "n/a":
formality_str = "in a " + formality + " manner, "
# put all emotions into a list
emotions = []
if anticipation_level != "n/a":
emotions.append(anticipation_level)
if joy_level != "n/a":
emotions.append(joy_level)
if trust_level != "n/a":
emotions.append(trust_level)
if fear_level != "n/a":
emotions.append(fear_level)
if surprise_level != "n/a":
emotions.append(surprise_level)
if sadness_level != "n/a":
emotions.append(sadness_level)
if disgust_level != "n/a":
emotions.append(disgust_level)
if anger_level != "n/a":
emotions.append(anger_level)
emotions_str = ""
if len(emotions) > 0:
if len(emotions) == 1:
emotions_str = "with emotion of " + emotions[0] + ", "
else:
emotions_str = "with emotions of " + \
", ".join(emotions[:-1]) + " and " + emotions[-1] + ", "
lang_level_str = ""
if lang_level != LANG_LEVEL_DEFAULT:
lang_level_str = "at a " + lang_level + \
" level, " if translate_to == TRANSLATE_TO_DEFAULT else ""
translate_to_str = ""
if translate_to != TRANSLATE_TO_DEFAULT:
translate_to_str = "translated to " + \
("" if lang_level == TRANSLATE_TO_DEFAULT else lang_level +
" level ") + translate_to + ", "
literary_style_str = ""
if literary_style != LITERARY_STYLE_DEFAULT:
if literary_style == "Prose":
literary_style_str = "as prose, "
elif literary_style == "Summary":
literary_style_str = "as a summary, "
elif literary_style == "Outline":
literary_style_str = "as an outline numbers and lower case letters, "
elif literary_style == "Bullets":
literary_style_str = "as bullet points using bullets, "
elif literary_style == "Poetry":
literary_style_str = "as a poem, "
elif literary_style == "Haiku":
literary_style_str = "as a haiku, "
elif literary_style == "Limerick":
literary_style_str = "as a limerick, "
elif literary_style == "Joke":
literary_style_str = "as a very funny joke with a setup and punchline, "
elif literary_style == "Knock-knock":
literary_style_str = "as a very funny knock-knock joke, "
formatted_prompt = PROMPT_TEMPLATE.format(
original_words=desc,
num_words=num_words_prompt,
formality=formality_str,
emotions=emotions_str,
lang_level=lang_level_str,
translate_to=translate_to_str,
literary_style=literary_style_str
)
trans_instr = num_words_prompt + formality_str + emotions_str + \
lang_level_str + translate_to_str + literary_style_str
if express_chain and len(trans_instr.strip()) > 0:
generated_text = express_chain.run(
{'original_words': desc, 'num_words': num_words_prompt, 'formality': formality_str,
'emotions': emotions_str, 'lang_level': lang_level_str, 'translate_to': translate_to_str,
'literary_style': literary_style_str}).strip()
else:
print("Not transforming text")
generated_text = desc
# replace all newlines with in generated_text
generated_text = generated_text.replace("\n", "\n\n")
prompt_plus_generated = "GPT prompt: " + \
formatted_prompt + "\n\n" + generated_text
print("\n==== date/time: " + str(datetime.datetime.now() -
datetime.timedelta(hours=5)) + " ====")
print("prompt_plus_generated: " + prompt_plus_generated)
return generated_text
def load_chain(tools_list, llm):
chain = None
express_chain = None
if llm:
print("\ntools_list", tools_list)
tool_names = tools_list
# tools = load_tools(tool_names, llm=llm, news_api_key=news_api_key,
# tmdb_bearer_token=tmdb_bearer_token)
# tools = load_tools(tool_names, llm=llm)
tools = [mysearch]
memory = ConversationBufferMemory(memory_key="chat_history")
chain = initialize_agent(
tools, llm, agent="conversational-react-description", verbose=True, memory=memory)
express_chain = LLMChain(llm=llm, prompt=PROMPT_TEMPLATE, verbose=True)
return chain, express_chain
def set_openai_api_key(api_key):
"""Set the api key and return chain.
If no api_key, then None is returned.
"""
# if api_key and api_key.startswith("sk-") and len(api_key) > 50:
# # os.environ["OPENAI_API_KEY"] = api_key
# os.environ["OPENAI_API_KEY"] = OPENAI_API_KEY
# llm = OpenAI(temperature=TEMPERATURE, max_tokens=MAX_TOKENS)
# chain, express_chain = load_chain(TOOLS_DEFAULT_LIST, llm)
# os.environ["OPENAI_API_KEY"] = ""
# return chain, express_chain, llm
# return None, None, None
os.environ["OPENAI_API_KEY"] = api_key
llm = OpenAI(temperature=TEMPERATURE, max_tokens=MAX_TOKENS)
chain, express_chain = load_chain(TOOLS_DEFAULT_LIST, llm)
# NOTE hmm this looks weird!
# os.environ["OPENAI_API_KEY"] = ""
return chain, express_chain, llm
def run_chain(chain, inp, capture_hidden_text):
output = ""
hidden_text = None
if capture_hidden_text:
error_msg = None
tmp = sys.stdout
hidden_text_io = StringIO()
sys.stdout = hidden_text_io
try:
output = chain.run(input=inp)
except AuthenticationError as ae:
print("run_chain oops, in capture_hidden_text")
print(repr(ae))
error_msg = AUTH_ERR_MSG
except RateLimitError as rle:
error_msg = "\n\nRateLimitError: " + str(rle)
except ValueError as ve:
error_msg = "\n\nValueError: " + str(ve)
except InvalidRequestError as ire:
error_msg = "\n\nInvalidRequestError: " + str(ire)
except Exception as e:
error_msg = "\n\n" + BUG_FOUND_MSG + ":\n\n" + str(e)
sys.stdout = tmp
hidden_text = hidden_text_io.getvalue()
# remove escape characters from hidden_text
hidden_text = re.sub(r'\x1b[^m]*m', '', hidden_text)
# remove "Entering new AgentExecutor chain..." from hidden_text
hidden_text = re.sub(
r"Entering new AgentExecutor chain...\n", "", hidden_text)
# remove "Finished chain." from hidden_text
hidden_text = re.sub(r"Finished chain.", "", hidden_text)
# Add newline after "Thought:" "Action:" "Observation:" "Input:" and "AI:"
hidden_text = re.sub(r"Thought:", "\n\nThought:", hidden_text)
hidden_text = re.sub(r"Action:", "\n\nAction:", hidden_text)
hidden_text = re.sub(r"Observation:", "\n\nObservation:", hidden_text)
hidden_text = re.sub(r"Input:", "\n\nInput:", hidden_text)
hidden_text = re.sub(r"AI:", "\n\nAI:", hidden_text)
if error_msg:
hidden_text += error_msg
print("hidden_text: ", hidden_text)
else:
try:
output = chain.run(input=inp)
except AuthenticationError as ae:
print("run_chain oops, in else")
print(repr(ae))
output = AUTH_ERR_MSG
except RateLimitError as rle:
output = "\n\nRateLimitError: " + str(rle)
except ValueError as ve:
output = "\n\nValueError: " + str(ve)
except InvalidRequestError as ire:
output = "\n\nInvalidRequestError: " + str(ire)
except Exception as e:
output = "\n\n" + BUG_FOUND_MSG + ":\n\n" + str(e)
return output, hidden_text
class ChatWrapper:
def __init__(self):
self.lock = Lock()
def __call__(
self, api_key: str, inp: str, history: Optional[Tuple[str, str]], chain: Optional[ConversationChain],
trace_chain: bool, speak_text: bool, talking_head: bool, monologue: bool, express_chain: Optional[LLMChain],
num_words, formality, anticipation_level, joy_level, trust_level,
fear_level, surprise_level, sadness_level, disgust_level, anger_level,
lang_level, translate_to, literary_style
):
"""Execute the chat functionality."""
self.lock.acquire()
try:
print("\n==== date/time: " + str(datetime.datetime.now()) + " ====")
print("inp: " + inp)
print("trace_chain: ", trace_chain)
print("speak_text: ", speak_text)
print("talking_head: ", talking_head)
print("monologue: ", monologue)
history = history or []
# If chain is None, that is because no API key was provided.
output = AUTH_ERR_MSG
hidden_text = output
if chain:
# Set OpenAI key
import openai
openai.api_key = os.environ["OPENAI_API_KEY"] # OPENAI_API_KEY
# openai.api_key = api_key
if not monologue:
output, hidden_text = run_chain(
chain, inp, capture_hidden_text=trace_chain)
else:
output, hidden_text = inp, None
output = transform_text(output, express_chain, num_words, formality, anticipation_level, joy_level,
trust_level,
fear_level, surprise_level, sadness_level, disgust_level, anger_level,
lang_level, translate_to, literary_style)
text_to_display = output
if trace_chain:
text_to_display = hidden_text + "\n\n" + output
history.append((inp, text_to_display))
html_video, temp_file, html_audio, temp_aud_file = None, None, None, None
if speak_text:
if talking_head:
if len(output) <= MAX_TALKING_HEAD_TEXT_LENGTH:
html_video, temp_file = do_html_video_speak(
output, translate_to)
else:
temp_file = LOOPING_TALKING_HEAD
html_video = create_html_video(
temp_file, TALKING_HEAD_WIDTH)
html_audio, temp_aud_file = do_html_audio_speak(
output, translate_to)
else:
html_audio, temp_aud_file = do_html_audio_speak(
output, translate_to)
else:
if talking_head:
temp_file = LOOPING_TALKING_HEAD
html_video = create_html_video(
temp_file, TALKING_HEAD_WIDTH)
else:
# html_audio, temp_aud_file = do_html_audio_speak(output, translate_to)
# html_video = create_html_video(temp_file, "128")
pass
except Exception as e:
raise e
finally:
self.lock.release()
return history, history, html_video, temp_file, html_audio, temp_aud_file, ""
# return history, history, html_audio, temp_aud_file, ""
chat = ChatWrapper()
def do_html_audio_speak(words_to_speak, polly_language):
polly_client = boto3.Session(
aws_access_key_id=os.environ["AWS_ACCESS_KEY_ID"],
aws_secret_access_key=os.environ["AWS_SECRET_ACCESS_KEY"],
region_name=os.environ["AWS_DEFAULT_REGION"]
).client('polly')
voice_id, language_code, engine = POLLY_VOICE_DATA.get_voice(
polly_language, VOICE_GENDER)
if not voice_id:
voice_id = "Joanna"
# voice_id = "Matthew"
language_code = "en-US"
engine = NEURAL_ENGINE
response = polly_client.synthesize_speech(
Text=words_to_speak,
OutputFormat='mp3',
VoiceId=voice_id,
LanguageCode=language_code,
Engine=engine
)
html_audio = '
no audio
'
# Save the audio stream returned by Amazon Polly on Lambda's temp directory
if "AudioStream" in response:
with closing(response["AudioStream"]) as stream:
# output = os.path.join("/tmp/", "speech.mp3")
try:
with open('audios/tempfile.mp3', 'wb') as f:
f.write(stream.read())
temp_aud_file = gr.File("audios/tempfile.mp3")
temp_aud_file_url = "/file=" + temp_aud_file.value['name']
html_audio = f''
except IOError as error:
# Could not write to file, exit gracefully
print(error)
return None, None
else:
# The response didn't contain audio data, exit gracefully
print("Could not stream audio")
return None, None
return html_audio, "audios/tempfile.mp3"
def create_html_video(file_name, width):
temp_file_url = "/file=" + tmp_file.value['name']
html_video = f''
return html_video
def do_html_video_speak(words_to_speak, azure_language):
azure_voice = AZURE_VOICE_DATA.get_voice(azure_language, VOICE_GENDER)
if not azure_voice:
azure_voice = "en-US-ChristopherNeural"
headers = {"Authorization": f"Bearer {os.environ['EXHUMAN_API_KEY']}"}
body = {
'bot_name': 'humancare',
'bot_response': words_to_speak,
'azure_voice': azure_voice,
'azure_style': 'friendly',
'animation_pipeline': 'high_speed',
}
api_endpoint = "https://api.exh.ai/animations/v1/generate_lipsync"
res = requests.post(api_endpoint, json=body, headers=headers)
print("res.status_code: ", res.status_code)
html_video = '
no video
'
if isinstance(res.content, bytes):
response_stream = io.BytesIO(res.content)
print("len(res.content)): ", len(res.content))
with open('videos/tempfile.mp4', 'wb') as f:
f.write(response_stream.read())
temp_file = gr.File("videos/tempfile.mp4")
temp_file_url = "/file=" + temp_file.value['name']
html_video = f''
else:
print('video url unknown')
return html_video, "videos/tempfile.mp4"
def update_selected_tools(widget, state, llm):
if widget:
state = widget
chain, express_chain = load_chain(state, llm)
return state, llm, chain, express_chain
def update_talking_head(widget, state):
if widget:
state = widget
video_html_talking_head = create_html_video(
LOOPING_TALKING_HEAD, TALKING_HEAD_WIDTH)
return state, video_html_talking_head
else:
# return state, create_html_video(LOOPING_TALKING_HEAD, "32")
return None, ""
def update_foo(widget, state):
if widget:
state = widget
return state
with gr.Blocks(css=CSS) as block:
llm_state = gr.State()
history_state = gr.State()
chain_state = gr.State()
express_chain_state = gr.State()
tools_list_state = gr.State(TOOLS_DEFAULT_LIST)
trace_chain_state = gr.State(False)
speak_text_state = gr.State(False)
talking_head_state = gr.State(True)
# Takes the input and repeats it back to the user, optionally transforming it.
monologue_state = gr.State(False)
# Pertains to Express-inator functionality
num_words_state = gr.State(NUM_WORDS_DEFAULT)
formality_state = gr.State(FORMALITY_DEFAULT)
anticipation_level_state = gr.State(EMOTION_DEFAULT)
joy_level_state = gr.State(EMOTION_DEFAULT)
trust_level_state = gr.State(EMOTION_DEFAULT)
fear_level_state = gr.State(EMOTION_DEFAULT)
surprise_level_state = gr.State(EMOTION_DEFAULT)
sadness_level_state = gr.State(EMOTION_DEFAULT)
disgust_level_state = gr.State(EMOTION_DEFAULT)
anger_level_state = gr.State(EMOTION_DEFAULT)
lang_level_state = gr.State(LANG_LEVEL_DEFAULT)
translate_to_state = gr.State(TRANSLATE_TO_DEFAULT)
literary_style_state = gr.State(LITERARY_STYLE_DEFAULT)
# Pertains to WHISPER functionality
whisper_lang_state = gr.State(WHISPER_DETECT_LANG)
with gr.Tab("Chat"):
with gr.Row():
with gr.Column():
gr.HTML(
"""
GPT + WolframAlpha + Whisper
New feature in Translate to: Choose Language level (e.g. for conversation practice or explain like I'm five)