大模型語言：神經網路在推薦系統中的多元應用

# LSTM模型用於序列推薦的簡化範例
import tensorflow as tf

class SequentialRecommender(tf.keras.Model):
    def __init__(self, num_items, embedding_dim, lstm_units):
        super().__init__()
        self.item_embedding = tf.keras.layers.Embedding(
            num_items + 1, embedding_dim, mask_zero=True)
        self.lstm = tf.keras.layers.LSTM(lstm_units)
        self.dense = tf.keras.layers.Dense(num_items, activation='softmax')
        
    def call(self, inputs):
        x = self.item_embedding(inputs)
        x = self.lstm(x)
        return self.dense(x)

# 用於推薦系統的簡化VAE實作
class RecommenderVAE(tf.keras.Model):
    def __init__(self, original_dim, latent_dim):
        super().__init__()
        # 編碼器
        self.encoder = tf.keras.Sequential([
            tf.keras.layers.Dense(256, activation='relu'),
            tf.keras.layers.Dense(latent_dim * 2)  # 均值和對數方差
        ])
        
        # 解碼器
        self.decoder = tf.keras.Sequential([
            tf.keras.layers.Dense(256, activation='relu'),
            tf.keras.layers.Dense(original_dim, activation='sigmoid')
        ])
    
    def encode(self, x):
        mean_logvar = self.encoder(x)
        mean, logvar = tf.split(mean_logvar, num_or_size_splits=2, axis=1)
        return mean, logvar
        
    def reparameterize(self, mean, logvar):
        eps = tf.random.normal(shape=mean.shape)
        return eps * tf.exp(logvar * 0.5) + mean
    
    def decode(self, z):
        return self.decoder(z)
        
    def call(self, x):
        mean, logvar = self.encode(x)
        z = self.reparameterize(mean, logvar)
        reconstruction = self.decode(z)
        return reconstruction, mean, logvar

import pandas as pd
import ast

# 將字串表示的字典轉換為實際字典
md['genres'] = md['genres'].apply(ast.literal_eval)

# 轉換'genres'欄位
md['genres'] = md['genres'].apply(lambda x: [genre['name'] for genre in x])

# 計算加權評分（使用IMDb公式）
def calculate_weighted_rate(vote_average, vote_count, min_vote_count=10):
    return (vote_count / (vote_count + min_vote_count)) * vote_average + \
           (min_vote_count / (vote_count + min_vote_count)) * 5.0

# 設定最低評分次數以防止結果偏差
vote_counts = md[md['vote_count'].notnull()]['vote_count'].astype('int')
min_vote_count = vote_counts.quantile(0.95)

# 建立新欄位'weighted_rate'
md['weighted_rate'] = md.apply(
    lambda row: calculate_weighted_rate(row['vote_average'], 
                                       row['vote_count'], 
                                       min_vote_count), 
    axis=1
)

md_final['combined_info'] = md_final.apply(
    lambda row: f"Title: {row['title']}. Overview: {row['overview']} " + 
                f"Genres: {', '.join(row['genres'])}. Rating: {row['weighted_rate']}", 
    axis=1
).astype(str)

import pandas as pd
import tiktoken
import os
import openai
from openai.embeddings_utils import get_embedding

openai.api_key = os.environ["OPENAI_API_KEY"]

embedding_encoding = "cl100k_base"  # text-embedding-ada-002使用的編碼
embedding_model = "text-embedding-ada-002"
max_tokens = 8000  # text-embedding-ada-002的最大值為8191

encoding = tiktoken.get_encoding(embedding_encoding)

# 排除過長無法嵌入的評論
md_final["n_tokens"] = md_final.combined_info.apply(lambda x: len(encoding.encode(x)))
md_final = md_final[md_final.n_tokens <= max_tokens]

md_final["embedding"] = md_final.overview.apply(lambda x: get_embedding(x, engine=embedding_model))

md_final.rename(columns={'embedding': 'vector'}, inplace=True)
md_final.rename(columns={'combined_info': 'text'}, inplace=True)
md_final.to_pickle('movies.pkl')

import lancedb

uri = "data/sample-lancedb"
db = lancedb.connect(uri)
table = db.create_table("movies", md)

from langchain.embeddings import OpenAIEmbeddings
from langchain.vectorstores import LanceDB
import os

os.environ["OPENAI_API_KEY"]
embeddings = OpenAIEmbeddings()
docsearch = LanceDB(connection=table, embedding=embeddings)

query = "I'm looking for an animated action movie. What could you suggest to me?"
docs = docsearch.similarity_search(query)
docs

qa = RetrievalQA.from_chain_type(
    llm=OpenAI(), 
    chain_type="stuff",
    retriever=docsearch.as_retriever(), 
    return_source_documents=True
)

query = "I'm looking for an animated action movie. What could you suggest to me?"
result = qa({"query": query})
print(result['result'])

'I would suggest Transformers. It is an animated action movie with genres of Adventure, Science Fiction, and Action, and a rating of 6.'

result['source_documents'][0]

Document(page_content='Title: Hitman: Agent 47. Overview: An assassin teams up with a woman to help her find her father and uncover the mysteries of her ancestry. Genres: Action, Crime, Thriller. Rating: 5.365800865800866', metadata={'genres': array(['Action', 'Crime', 'Thriller'], dtype=object), 'title': 'Hitman: Agent 47', 'overview': 'An assassin teams up with a woman to help her find her father and uncover the mysteries of her ancestry.', 'weighted_rate': 5.365800865800866, 'n_tokens': 52, 'vector': array([-0.00566491, -0.01658553, -0.02255735, ..., -0.01242317, -0.01303058, -0.00709073], dtype=float32), '_distance': 0.42414575815200806})

df_filtered = md[md['genres'].apply(lambda x: 'Comedy' in x)]
qa = RetrievalQA.from_chain_type(
    llm=OpenAI(), 
    chain_type="stuff",
    retriever=docsearch.as_retriever(search_kwargs={'data': df_filtered}),
    return_source_documents=True
)

query = "I'm looking for a movie with animals and an adventurous plot."
result = qa({"query": query})

qa = RetrievalQA.from_chain_type(
    llm=OpenAI(), 
    chain_type="stuff",
    retriever=docsearch.as_retriever(search_kwargs={'filter': {'weighted_rate__gt': 7}}),
    return_source_documents=True
)

from langchain.agents.agent_toolkits import create_retriever_tool
from langchain.agents.agent_toolkits import create_conversational_retrieval_agent
from langchain.chat_models import ChatOpenAI

llm = ChatOpenAI(temperature=0)
retriever = docsearch.as_retriever(return_source_documents=True)

tool = create_retriever_tool(
    retriever,
    "movies",
    "Searches and returns recommendations about movies."
)
tools = [tool]
agent_executor = create_conversational_retrieval_agent(llm, tools, verbose=True)
result = agent_executor({"input": "suggest me some action movies"})

> Entering new AgentExecutor chain...
Invoking: `movies` with `{'genre': 'action'}`
[Document(page_content='The action continues from [REC], [...] 
Here are some action movies that you might enjoy:
1. [REC]² - The action continues from [REC], with a medical officer and a SWAT team sent into a sealed-off apartment to control the situation. It is a thriller/horror movie.
2. The Boondock Saints - Twin brothers Conner and Murphy take swift retribution into their own hands to rid Boston of criminals. It is an action/thriller/crime movie.
3. The Gamers - Four clueless players are sent on a quest to rescue a princess and must navigate dangerous forests, ancient ruins, and more. It is an action/comedy/thriller/foreign movie.
4. Atlas Shrugged Part III: Who is John Galt? - In a collapsing economy, one man has the answer while others try to control or save him. It is a drama/science fiction/mystery movie.
Please note that these recommendations are based on the genre "action" and may vary in terms of availability and personal preferences.
> Finished chain.

print(qa.combine_documents_chain.llm_chain.prompt.template)

Use the following pieces of context to answer the question at the end. If you don't know the answer, just say that you don't know, don't try to make up an answer.
{context}
Question: {question} Helpful Answer:

from langchain.prompts import PromptTemplate

template = """You are a movie recommender system that help users to find movies that match their preferences.
Use the following pieces of context to answer the question at the end. For each question, suggest three movies, with a short description of the plot and the reason why the user migth like it.
If you don't know the answer, just say that you don't know, don't try to make up an answer.
{context}
Question: {question} Your response:"""

PROMPT = PromptTemplate(
    template=template, input_variables=["context", "question"])

chain_type_kwargs = {"prompt": PROMPT}
qa = RetrievalQA.from_chain_type(
    llm=OpenAI(),
    chain_type="stuff",
    retriever=docsearch.as_retriever(),
    return_source_documents=True,
    chain_type_kwargs=chain_type_kwargs
)

query = "I'm looking for a funny action movie, any suggestion?"
result = qa({'query': query})
print(result['result'])

1. A Good Day to Die Hard: An action-packed comedy directed by John Moore, this movie follows Iconoclastic, take-no-prisoners cop John McClane as he travels to Moscow to help his wayward son Jack. With the Russian underworld in pursuit, and battling a countdown to war, the two McClanes discover that their opposing methods make them unstoppable heroes. You might enjoy this film for its blend of intense action sequences and humorous father-son dynamics.