468db83337
This commit enables anyone to create a app and add 3 types of data sources: * pdf file * youtube video * website It exposes a function called query which first gets similar docs from vector db and then passes it to LLM to get the final answer.
36 lines
996 B
Python
36 lines
996 B
Python
import hashlib
|
|
|
|
from langchain.text_splitter import RecursiveCharacterTextSplitter
|
|
|
|
|
|
TEXT_SPLITTER_CHUNK_PARAMS = {
|
|
"chunk_size": 2000,
|
|
"chunk_overlap": 0,
|
|
"length_function": len,
|
|
}
|
|
|
|
TEXT_SPLITTER = RecursiveCharacterTextSplitter(**TEXT_SPLITTER_CHUNK_PARAMS)
|
|
|
|
|
|
class YoutubeVideoChunker:
|
|
|
|
def create_chunks(self, loader, url):
|
|
documents = []
|
|
ids = []
|
|
datas = loader.load_data(url)
|
|
metadatas = []
|
|
for data in datas:
|
|
content = data["content"]
|
|
meta_data = data["meta_data"]
|
|
chunks = TEXT_SPLITTER.split_text(content)
|
|
url = meta_data["url"]
|
|
for chunk in chunks:
|
|
chunk_id = hashlib.sha256((chunk + url).encode()).hexdigest()
|
|
ids.append(chunk_id)
|
|
documents.append(chunk)
|
|
metadatas.append(meta_data)
|
|
return {
|
|
"documents": documents,
|
|
"ids": ids,
|
|
"metadatas": metadatas,
|
|
} |