Upload app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,212 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# -*- coding: utf-8 -*-
|
| 2 |
+
"""Untitled0.ipynb
|
| 3 |
+
|
| 4 |
+
Automatically generated by Colab.
|
| 5 |
+
|
| 6 |
+
Original file is located at
|
| 7 |
+
https://colab.research.google.com/drive/1jmDagWzG5dZJK1QAfpBnu3n1NXS_QIOz
|
| 8 |
+
"""
|
| 9 |
+
|
| 10 |
+
!pip install pandas datasets
|
| 11 |
+
|
| 12 |
+
import pandas as pd
|
| 13 |
+
from datasets import load_dataset
|
| 14 |
+
|
| 15 |
+
# 1. Download the raw dataset from Hugging Face
|
| 16 |
+
raw_data = load_dataset("bitext/Bitext-customer-support-llm-chatbot-training-dataset")
|
| 17 |
+
|
| 18 |
+
# 2. Convert it into a Pandas DataFrame (think of this as a code-based Excel sheet)
|
| 19 |
+
df = pd.DataFrame(raw_data['train'])
|
| 20 |
+
|
| 21 |
+
# 3. Data Cleaning: Drop rows with missing data and remove duplicate queries
|
| 22 |
+
df.dropna(subset=['instruction', 'response'], inplace=True)
|
| 23 |
+
df.drop_duplicates(subset=['instruction'], inplace=True)
|
| 24 |
+
|
| 25 |
+
# 4. Feature Selection: Keep only the columns we actually need for the AI
|
| 26 |
+
df = df[['category', 'instruction', 'response']]
|
| 27 |
+
|
| 28 |
+
# 5. Rename columns so they align with our business logic
|
| 29 |
+
df.columns = ['ticket_category', 'customer_query', 'historical_resolution']
|
| 30 |
+
|
| 31 |
+
# Display the first 5 rows to verify the cleaning worked
|
| 32 |
+
print(df.head())
|
| 33 |
+
|
| 34 |
+
!pip install sentence-transformers
|
| 35 |
+
|
| 36 |
+
from sentence_transformers import SentenceTransformer
|
| 37 |
+
import numpy as np
|
| 38 |
+
|
| 39 |
+
# 1. Load the Multilingual Embedding Model
|
| 40 |
+
# This specific model understands over 50 languages and aligns them in the same semantic space
|
| 41 |
+
print("Downloading the multilingual embedding model (this may take a minute)...")
|
| 42 |
+
model = SentenceTransformer('paraphrase-multilingual-MiniLM-L12-v2')
|
| 43 |
+
|
| 44 |
+
# 2. Extract a sample of our cleaned resolutions
|
| 45 |
+
# We use 100 rows to quickly test the pipeline
|
| 46 |
+
sample_resolutions = df['historical_resolution'].head(100).tolist()
|
| 47 |
+
|
| 48 |
+
# 3. Convert the text into vectors
|
| 49 |
+
print("Generating embeddings for the resolutions...")
|
| 50 |
+
resolution_embeddings = model.encode(sample_resolutions)
|
| 51 |
+
|
| 52 |
+
# 4. Verify the output
|
| 53 |
+
print(f"\nSuccessfully created {len(resolution_embeddings)} vector embeddings.")
|
| 54 |
+
print(f"Each vector is an array of {len(resolution_embeddings[0])} numbers.")
|
| 55 |
+
print("Here are the first 5 numbers of the first vector:")
|
| 56 |
+
print(resolution_embeddings[0][:5])
|
| 57 |
+
|
| 58 |
+
from sentence_transformers import util
|
| 59 |
+
|
| 60 |
+
# 1. Simulate an incoming customer ticket in German
|
| 61 |
+
# Meaning: "How do I create a new account?"
|
| 62 |
+
customer_query = "Wie erstelle ich ein neues Konto?"
|
| 63 |
+
print(f"Incoming Customer Query (German): '{customer_query}'\n")
|
| 64 |
+
|
| 65 |
+
# 2. Embed the query using the exact same model
|
| 66 |
+
query_embedding = model.encode(customer_query)
|
| 67 |
+
|
| 68 |
+
# 3. Perform the Semantic Search
|
| 69 |
+
# This calculates the cosine similarity between the query vector and all 100 stored resolution vectors
|
| 70 |
+
# top_k=1 tells the system to only return the single best match
|
| 71 |
+
hits = util.semantic_search(query_embedding, resolution_embeddings, top_k=1)
|
| 72 |
+
|
| 73 |
+
# 4. Extract the results
|
| 74 |
+
best_match_index = hits[0][0]['corpus_id']
|
| 75 |
+
similarity_score = hits[0][0]['score']
|
| 76 |
+
|
| 77 |
+
# 5. Display the matching document from our original Pandas DataFrame
|
| 78 |
+
print(f"Similarity Score: {similarity_score:.2f} (1.0 is a perfect semantic match)\n")
|
| 79 |
+
print("--- Retrieved Internal Resolution Document (English) ---")
|
| 80 |
+
print(df['historical_resolution'].iloc[best_match_index])
|
| 81 |
+
|
| 82 |
+
!pip install huggingface_hub
|
| 83 |
+
|
| 84 |
+
from huggingface_hub import InferenceClient
|
| 85 |
+
|
| 86 |
+
# 1. Initialize the client with your token
|
| 87 |
+
# We are using a highly capable open-source instruction model
|
| 88 |
+
client = InferenceClient(
|
| 89 |
+
model="HuggingFaceH4/zephyr-7b-beta",
|
| 90 |
+
token="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"
|
| 91 |
+
)
|
| 92 |
+
|
| 93 |
+
# 2. Extract the retrieved English document from the previous step
|
| 94 |
+
retrieved_context = df['historical_resolution'].iloc[best_match_index]
|
| 95 |
+
|
| 96 |
+
# 3. Construct the Prompt
|
| 97 |
+
# This is the "Prompt Engineering" layer where we instruct the LLM on how to behave
|
| 98 |
+
prompt = f"""
|
| 99 |
+
You are a professional customer support agent for a tech company.
|
| 100 |
+
A customer has asked the following query: "{customer_query}"
|
| 101 |
+
|
| 102 |
+
Using ONLY the information provided in the internal resolution document below, write a polite and helpful email response directly to the customer.
|
| 103 |
+
The response MUST be in the same language as the customer's query. Do not invent any new steps.
|
| 104 |
+
|
| 105 |
+
Internal Resolution Document:
|
| 106 |
+
{retrieved_context}
|
| 107 |
+
|
| 108 |
+
Final Response:
|
| 109 |
+
"""
|
| 110 |
+
|
| 111 |
+
# 4. Generate the Response
|
| 112 |
+
print("Generating final response via LLM...\n")
|
| 113 |
+
try:
|
| 114 |
+
response = client.text_generation(
|
| 115 |
+
prompt,
|
| 116 |
+
max_new_tokens=250, # Limits the length of the response
|
| 117 |
+
temperature=0.3, # Keeps the response focused and factual rather than "creative"
|
| 118 |
+
)
|
| 119 |
+
print("--- Final Customer Email (German) ---")
|
| 120 |
+
print(response.strip())
|
| 121 |
+
except Exception as e:
|
| 122 |
+
print(f"Error during generation: {e}")
|
| 123 |
+
|
| 124 |
+
from huggingface_hub import InferenceClient
|
| 125 |
+
|
| 126 |
+
# 1. Initialize the client using the highly stable Llama-3 model
|
| 127 |
+
client = InferenceClient(
|
| 128 |
+
model="meta-llama/Meta-Llama-3-8B-Instruct",
|
| 129 |
+
token="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"
|
| 130 |
+
)
|
| 131 |
+
|
| 132 |
+
# 2. Extract the retrieved English document from our database
|
| 133 |
+
retrieved_context = df['historical_resolution'].iloc[best_match_index]
|
| 134 |
+
|
| 135 |
+
# 3. Construct the Chat Messages
|
| 136 |
+
# Instead of a raw string, modern APIs use a structured list of messages
|
| 137 |
+
messages = [
|
| 138 |
+
{
|
| 139 |
+
"role": "system",
|
| 140 |
+
"content": "You are a professional customer support agent. Use ONLY the provided Internal Document to answer the user. Respond in the exact same language as the user's query."
|
| 141 |
+
},
|
| 142 |
+
{
|
| 143 |
+
"role": "user",
|
| 144 |
+
"content": f"Customer Query: {customer_query}\n\nInternal Document: {retrieved_context}"
|
| 145 |
+
}
|
| 146 |
+
]
|
| 147 |
+
|
| 148 |
+
# 4. Generate the Response using the modern Chat Completion API
|
| 149 |
+
print("Generating final response via LLM...\n")
|
| 150 |
+
try:
|
| 151 |
+
response = client.chat.completions.create(
|
| 152 |
+
messages=messages,
|
| 153 |
+
max_tokens=250,
|
| 154 |
+
temperature=0.3,
|
| 155 |
+
)
|
| 156 |
+
|
| 157 |
+
print("--- Final Customer Email (German) ---")
|
| 158 |
+
# Extracting the text from the structured API response
|
| 159 |
+
print(response.choices[0].message.content.strip())
|
| 160 |
+
|
| 161 |
+
except Exception as e:
|
| 162 |
+
print(f"Error during generation: {e}")
|
| 163 |
+
|
| 164 |
+
# 1. Install Gradio (Run this in a cell if in Colab: !pip install gradio)
|
| 165 |
+
import gradio as gr
|
| 166 |
+
|
| 167 |
+
# 2. Define the core function that ties everything together
|
| 168 |
+
def resolve_ticket(customer_query):
|
| 169 |
+
# Embed the query
|
| 170 |
+
query_embedding = model.encode(customer_query)
|
| 171 |
+
|
| 172 |
+
# Retrieve the best match
|
| 173 |
+
hits = util.semantic_search(query_embedding, resolution_embeddings, top_k=1)
|
| 174 |
+
best_match_index = hits[0][0]['corpus_id']
|
| 175 |
+
retrieved_context = df['historical_resolution'].iloc[best_match_index]
|
| 176 |
+
|
| 177 |
+
# Prompt the LLM
|
| 178 |
+
messages = [
|
| 179 |
+
{"role": "system", "content": "You are a professional customer support agent. Use ONLY the provided Internal Document to answer the user. Respond in the exact same language as the user's query."},
|
| 180 |
+
{"role": "user", "content": f"Customer Query: {customer_query}\n\nInternal Document: {retrieved_context}"}
|
| 181 |
+
]
|
| 182 |
+
|
| 183 |
+
# Generate the response
|
| 184 |
+
try:
|
| 185 |
+
response = client.chat.completions.create(
|
| 186 |
+
messages=messages,
|
| 187 |
+
max_tokens=250,
|
| 188 |
+
temperature=0.3,
|
| 189 |
+
)
|
| 190 |
+
return response.choices[0].message.content.strip(), retrieved_context
|
| 191 |
+
except Exception as e:
|
| 192 |
+
return f"Error: {e}", retrieved_context
|
| 193 |
+
|
| 194 |
+
# 3. Build the Web Interface
|
| 195 |
+
with gr.Blocks() as demo:
|
| 196 |
+
gr.Markdown("# 🌍 Multilingual AI Support Triage")
|
| 197 |
+
gr.Markdown("Type a support query in any language. The system will retrieve the correct English internal document and draft a reply in your language.")
|
| 198 |
+
|
| 199 |
+
with gr.Row():
|
| 200 |
+
with gr.Column():
|
| 201 |
+
input_query = gr.Textbox(lines=4, label="Incoming Customer Ticket (Try German, French, etc.)")
|
| 202 |
+
submit_btn = gr.Button("Resolve Ticket")
|
| 203 |
+
|
| 204 |
+
with gr.Column():
|
| 205 |
+
output_reply = gr.Textbox(lines=6, label="Generated Email Response")
|
| 206 |
+
retrieved_doc = gr.Textbox(lines=4, label="Retrieved Internal Document (English)")
|
| 207 |
+
|
| 208 |
+
# Connect the button to the function
|
| 209 |
+
submit_btn.click(fn=resolve_ticket, inputs=input_query, outputs=[output_reply, retrieved_doc])
|
| 210 |
+
|
| 211 |
+
# 4. Launch the App
|
| 212 |
+
demo.launch()
|