Preparing the documents and vector database
Over the next few exercises, you'll build a full RAG workflow to have a conversation with a PDF document containing the paper, RAG VS Fine-Tuning: Pipelines, Tradeoffs, and a Case Study on Agriculture by Balaguer et al. (2024). This works by splitting the documents into chunks, storing them in a vector database, defining a prompt to connect the retrieved documents and user input, and building a retrieval chain for the LLM to access this external data.
In this exercise, you'll prepare the document for storage and ingest them into a Chroma vector database. You'll use a RecursiveCharacterTextSplitter to chunk the PDF, and ingest them into a Chroma vector database using an OpenAI embeddings function. As with the rest of the course, you don't need to provide your own OpenAI API key.
The following classes have already been imported for you: RecursiveCharacterTextSplitter, Chroma, and OpenAIEmbeddings.
This exercise is part of the course
Developing LLM Applications with LangChain
Exercise instructions
- Split the documents in
datausing a recursive character splitter with achunk_sizeof300andchunk_overlapof50; leave theseparatorsargument out, as it defaults to["\n\n", "\n", " ", ""]. - Define an OpenAI embeddings model using the
<OPENAI_API_TOKEN>placeholder provided and use it to embed and ingest the documents into a Chroma database. - Configure
vectorstoreinto a retriever object that returns the top 3 documents for use in the final RAG chain.
Hands-on interactive exercise
Have a go at this exercise by completing this sample code.
loader = PyPDFLoader('rag_vs_fine_tuning.pdf')
data = loader.load()
# Split the document using RecursiveCharacterTextSplitter
splitter = ____
docs = splitter.____(data)
# Embed the documents in a persistent Chroma vector database
embedding_function = OpenAIEmbeddings(api_key='', model='text-embedding-3-small')
vectorstore = Chroma.____(
____,
embedding=____,
persist_directory=os.getcwd()
)
# Configure the vector store as a retriever
retriever = ____(
search_type="similarity",
search_kwargs={____}
)