File size: 5,959 Bytes
e997328
 
 
 
 
 
 
 
4d2ab8e
 
 
 
e997328
 
4d2ab8e
e997328
 
 
c18c7db
e3c4cb8
4d2ab8e
 
 
 
 
 
 
 
157ebae
db2d292
157ebae
e997328
e3c4cb8
e997328
 
db2d292
 
 
 
 
 
 
9798fe7
db2d292
 
e997328
c18c7db
 
 
e997328
4d2ab8e
 
 
e997328
4d2ab8e
e997328
 
 
c18c7db
4d2ab8e
ccabaa8
4d2ab8e
 
 
e997328
4d2ab8e
ccabaa8
 
c18c7db
e997328
72998ae
4d2ab8e
 
 
 
 
 
 
 
e997328
 
 
 
 
 
4d2ab8e
 
 
 
 
 
 
 
 
 
 
 
 
 
e997328
 
 
 
4d2ab8e
 
 
e2dbbeb
4d2ab8e
 
 
 
 
 
 
 
 
 
 
 
72998ae
4d2ab8e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e3c4cb8
 
 
4d2ab8e
 
 
 
 
 
 
 
 
157ebae
c18c7db
4d2ab8e
 
 
 
 
c18c7db
 
 
4d2ab8e
 
 
 
 
 
 
 
 
c18c7db
741deae
 
 
 
 
 
4d2ab8e
e3c4cb8
4d2ab8e
 
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
from pathlib import Path

import streamlit as st
from dotenv import load_dotenv
from langchain.chains import LLMChain
from langchain.prompts import PromptTemplate

load_dotenv()
import os
from typing import List, Tuple

import numpy as np
from langchain.chat_models import ChatOpenAI
from langchain.embeddings.openai import OpenAIEmbeddings
from langchain.schema import Document

from data import load_db
from names import DATASET_ID, MODEL_ID
from storage import RedisStorage, UserInput
from utils import weighted_random_sample


class RetrievalType:
    FIRST_MATCH = "first-match"
    POOL_MATCHES = "pool-matches"


Matches = List[Tuple[Document, float]]
USE_STORAGE = os.environ.get("USE_STORAGE", "True").lower() in ("true", "t", "1")

print("USE_STORAGE", USE_STORAGE)


@st.cache_resource
def init():
    embeddings = OpenAIEmbeddings(model=MODEL_ID)
    dataset_path = f"hub://{os.environ['ACTIVELOOP_ORG_ID']}/{DATASET_ID}"

    db = load_db(
        dataset_path,
        embedding_function=embeddings,
        token=os.environ["ACTIVELOOP_TOKEN"],
        # org_id=os.environ["ACTIVELOOP_ORG_ID"],
        read_only=True,
    )

    storage = RedisStorage(
        host=os.environ["UPSTASH_URL"], password=os.environ["UPSTASH_PASSWORD"]
    )
    prompt = PromptTemplate(
        input_variables=["user_input"],
        template=Path("prompts/bot.prompt").read_text(),
    )

    llm = ChatOpenAI(temperature=0.3)

    chain = LLMChain(llm=llm, prompt=prompt)

    return db, storage, chain


# Don't show the setting sidebar
if "sidebar_state" not in st.session_state:
    st.session_state.sidebar_state = "collapsed"

st.set_page_config(initial_sidebar_state=st.session_state.sidebar_state)


db, storage, chain = init()

st.title("FairytaleDJ ๐ŸŽต๐Ÿฐ๐Ÿ”ฎ")
st.markdown(
    """
*<small>Made with [DeepLake](https://www.deeplake.ai/) ๐Ÿš€ and [LangChain](https://python.langchain.com/en/latest/index.html) ๐Ÿฆœโ›“๏ธ</small>*

๐Ÿ’ซ Unleash the magic within you with our enchanting app, turning your sentiments into a Disney soundtrack! ๐ŸŒˆ Just express your emotions, and embark on a whimsical journey as we tailor a Disney melody to match your mood. ๐Ÿ‘‘๐Ÿ’–""",
    unsafe_allow_html=True,
)
how_it_works = st.expander(label="How it works")

text_input = st.text_input(
    label="How are you feeling today?",
    placeholder="I am ready to rock and rool!",
)

run_btn = st.button("Make me sing! ๐ŸŽถ")
with how_it_works:
    st.markdown(
        """
The application follows a sequence of steps to deliver Disney songs matching the user's emotions:
- **User Input**: The application starts by collecting user's emotional state through a text input.
- **Emotion Encoding**: The user-provided emotions are then fed to a Language Model (LLM). The LLM interprets and encodes these emotions.
- **Similarity Search**: These encoded emotions are utilized to perform a similarity search within our [vector database](https://www.deeplake.ai/). This database houses Disney songs, each represented as emotional embeddings.
- **Song Selection**: From the pool of top matching songs, the application randomly selects one. The selection is weighted, giving preference to songs with higher similarity scores.
- **Song Retrieval**: The selected song's embedded player is displayed on the webpage for the user. Additionally, the LLM interpreted emotional state associated with the chosen song is displayed.
"""
    )


placeholder_emotions = st.empty()
placeholder = st.empty()


with st.sidebar:
    st.text("App settings")
    filter_threshold = st.slider(
        "Threshold used to filter out low scoring songs",
        min_value=0.0,
        max_value=1.0,
        value=0.8,
    )
    max_number_of_songs = st.slider(
        "Max number of songs we will retrieve from the db",
        min_value=5,
        max_value=50,
        value=20,
        step=1,
    )
    number_of_displayed_songs = st.slider(
        "Number of displayed songs", min_value=1, max_value=4, value=2, step=1
    )


def filter_scores(matches: Matches, th: float = 0.8) -> Matches:
    return [(doc, score) for (doc, score) in matches if score > th]


def normalize_scores_by_sum(matches: Matches) -> Matches:
    scores = [score for _, score in matches]
    tot = sum(scores)
    return [(doc, (score / tot)) for doc, score in matches]


def get_song(user_input: str, k: int = 20):
    emotions = chain.run(user_input=user_input)
    matches = db.similarity_search_with_score(emotions, distance_metric="cos", k=k)
    # [print(doc.metadata['name'], score) for doc, score in matches]
    docs, scores = zip(
        *normalize_scores_by_sum(filter_scores(matches, filter_threshold))
    )
    choosen_docs = weighted_random_sample(
        np.array(docs), np.array(scores), n=number_of_displayed_songs
    ).tolist()
    return choosen_docs, emotions


def set_song(user_input):
    if user_input == "":
        return
    # take first 120 chars
    user_input = user_input[:120]
    docs, emotions = get_song(user_input, k=max_number_of_songs)
    print(docs)
    songs = []
    with placeholder_emotions:
        st.markdown("Your emotions: `" + emotions + "`")
    with placeholder:
        iframes_html = ""
        for doc in docs:
            name = doc.metadata["name"]
            print(f"song = {name}")
            songs.append(name)
            embed_url = doc.metadata["embed_url"]
            iframes_html += (
                f'<iframe src="{embed_url}" style="border:0;height:100px"> </iframe>'
            )

        st.markdown(
            f"<div style='display:flex;flex-direction:column'>{iframes_html}</div>",
            unsafe_allow_html=True,
        )

        if USE_STORAGE:
            success_storage = storage.store(
                UserInput(text=user_input, emotions=emotions, songs=songs)
            )
            if not success_storage:
                print("[ERROR] was not able to store user_input")


if run_btn:
    set_song(text_input)