Spaces:
Sleeping
Sleeping
Update langchain_helper.py
Browse files- langchain_helper.py +60 -0
langchain_helper.py
CHANGED
@@ -0,0 +1,60 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from langchain.document_loaders import YoutubeLoader
|
2 |
+
from langchain.text_splitter import RecursiveCharacterTextSplitter
|
3 |
+
from langchain.llms import OpenAI
|
4 |
+
from langchain.prompts import PromptTemplate
|
5 |
+
from langchain.chains import LLMChain
|
6 |
+
from langchain.vectorstores import FAISS # for similarity search
|
7 |
+
from langchain.embeddings.openai import OpenAIEmbeddings
|
8 |
+
|
9 |
+
|
10 |
+
from dotenv import load_dotenv
|
11 |
+
|
12 |
+
load_dotenv()
|
13 |
+
|
14 |
+
embeddings = OpenAIEmbeddings()
|
15 |
+
|
16 |
+
# video_url = "https://www.youtube.com/watch?v=OyFJWRnt_AY"
|
17 |
+
|
18 |
+
|
19 |
+
def create_vector_db_from_ytUrl(video_url: str) -> FAISS:
|
20 |
+
'''First load the youtube video. calling loader.load() will automatically transcribe the youtube video. Next we split the text with RecursiveCharacterTextSplitter. Create and return a vector store object '''
|
21 |
+
|
22 |
+
loader = YoutubeLoader.from_youtube_url(video_url)
|
23 |
+
video_transcript = loader.load()
|
24 |
+
|
25 |
+
text_splitter = RecursiveCharacterTextSplitter(
|
26 |
+
chunk_size=1000, chunk_overlap=100)
|
27 |
+
|
28 |
+
docs = text_splitter.split_documents(video_transcript)
|
29 |
+
|
30 |
+
db = FAISS.from_documents(docs, embeddings)
|
31 |
+
return db
|
32 |
+
|
33 |
+
|
34 |
+
def get_response_from_query(db, query, k=4):
|
35 |
+
''' text_davinci max tokens = 4097
|
36 |
+
Semantic search '''
|
37 |
+
|
38 |
+
docs = db.similarity_search(query, k=k)
|
39 |
+
|
40 |
+
docs_page_content = " ".join([d.page_content for d in docs])
|
41 |
+
|
42 |
+
llm = OpenAI(model="text-davinci-003")
|
43 |
+
|
44 |
+
prompt = PromptTemplate(
|
45 |
+
input_variables=["question", "docs"],
|
46 |
+
template="""You are a helpful YouTube assistant that can answer questions about videos from video transcripts.
|
47 |
+
Answer the following question: {question}
|
48 |
+
By searching the following video transcript: {docs}
|
49 |
+
|
50 |
+
If you feel like you do not have enough information to give the answer, simply say "I have not much information to answer the question!"
|
51 |
+
|
52 |
+
Your answers should be detailed.
|
53 |
+
"""
|
54 |
+
)
|
55 |
+
|
56 |
+
chain = LLMChain(llm=llm, prompt=prompt)
|
57 |
+
|
58 |
+
response = chain.run(question=query, docs=docs_page_content)
|
59 |
+
response = response.replace("\n", "")
|
60 |
+
return response, docs
|