How to create pinecone embeddings offline for use by online bot.

[svange]

Hello!

All of the examples I see for question/answering over docs create their embeddings and then use the index(?) made during the process of creating those embeddings immediately (i.e. db.from_documents(docs, embeddings)).

I want to populate my vector store from my home computer, and then I want my agent (which exists as a service online, available through an API) to have access to my vector store for question answering at all times. I can't follow the guides because I don't have a reference to the db object when not on my home computer, embedding docs.

How can I do this? Am I implementing bad pattern here?

Thanks!

[Hamas-ur-Rehman]

I have done something similar like this using chromadb
you might need to install

pip install chroma
pip install chromadb

from langchain.vectorstores import Chroma

//define your document loader and splitters here and add the splitted documents into `texts` variable

docsearch = Chroma.from_texts([t.page_content for t in texts], embeddings, collection_name=collection_name , persist_directory=f'db/{collection_name}')
docsearch.persist()
docsearch = None

you would load the data and store it into your embeddings this way then what you can do is make an api which would feed this vector store but you want to load it and then feed it.

vectorstore = Chroma(collection_name=f'{collection_name}',persist_directory=f'./db/{collection_name}',embedding_function=embeddings)
qa = VectorDBQA.from_chain_type(llm=OpenAI(temperature=0.2), chain_type="map_reduce", vectorstore=vectorstore,return_source_documents=False)

Now you can return qa variable to your online bot another better thing is there will be a folder created db/{collection_name} you can host that with the bot so it can access it directly and you can update this file via remote access everytime you make a change to the vectorstore

[svange]

Wow! Cool mechanism! It's a bit slow in searching, so figuring this out for a vectorstore solution will be nice, but this does the trick for prototyping! Great answer, thanks!