Turning Large PDFs into Vectors with LangChain: Split, Embed, and Retrieve

Why Retrieval Is Needed

LLMs such as GPT‑4o have a maximum context window (128k tokens), which is insufficient for whole technical documents like a 50‑page PDF.

Solution Overview

The workflow avoids sending the entire document to the LLM and instead follows four steps:

Load the document and convert it to Document objects.

Split each document into small chunks.

Embed each chunk into a vector.

Store the vectors in a vector store and retrieve the most relevant chunks at query time.

Document Loaders

LangChain provides loaders for many formats. Example for PDF:

from langchain_community.document_loaders import PyPDFLoader
loader = PyPDFLoader("path/to/document.pdf")
pages = loader.load()
for page in pages:
    print(f"Page {page.metadata['page']} length: {len(page.page_content)} characters")
    print(page.page_content[:200])

Other loaders include UnstructuredMarkdownLoader for Markdown, WebBaseLoader for web pages, and built‑in loaders such as CSVLoader, TextLoader, and DirectoryLoader.

Text Splitting

Three splitter options are demonstrated:

CharacterTextSplitter : simple character‑based splitting (e.g., 500‑char chunks with 50‑char overlap).

RecursiveCharacterTextSplitter : hierarchical splitting that prefers paragraphs, then sentences, then words.

Token‑aware splitting : uses tiktoken to count tokens for more stable chunk sizes.

from langchain_text_splitters import RecursiveCharacterTextSplitter
splitter = RecursiveCharacterTextSplitter(
    chunk_size=500,
    chunk_overlap=50,
    length_function=len,
    separators=["

", "
", "。", "！", "？", " ", ""]
)
texts = splitter.split_documents(docs)
print(f"Created {len(texts)} text chunks")

Embedding

Two embedding providers are covered:

OpenAIEmbeddings (e.g., text-embedding-3-small, 1536‑dimensional, $0.02 per 1M tokens).

OllamaEmbeddings (e.g., nomic-embed-text, 768‑dimensional, free local model).

from langchain_openai import OpenAIEmbeddings
embedding = OpenAIEmbeddings(model="text-embedding-3-small")
vector = embedding.embed_query("LangChain makes LLM app development easy")
print(f"Vector dimension: {len(vector)}")

Vector Store Selection

Common vector databases and their trade‑offs:

Chroma : lightweight, Python‑native, ideal for prototypes; not suited for large‑scale production.

FAISS : fast, single‑machine, good for up to billions of vectors; lacks native distributed support.

Milvus : distributed, cloud‑native, for massive production workloads; deployment is complex.

Qdrant : strong performance with hybrid search; relatively new.

Pinecone : fully managed SaaS, pay‑as‑you‑go; incurs latency and cost.

Example with Chroma:

# pip install langchain-chroma==1.1.0 chromadb
from langchain_chroma import Chroma
from langchain_openai import OpenAIEmbeddings
embedding = OpenAIEmbeddings(model="text-embedding-3-small")
vectorstore = Chroma.from_documents(
    documents=texts,
    embedding=embedding,
    persist_directory="./chroma_db"
)
results = vectorstore.similarity_search(query="What are the core concepts of LangChain?", k=3)
for i, doc in enumerate(results):
    print(f"Result {i+1}: {doc.page_content[:200]}...")

Example with FAISS (including score):

# pip install faiss-cpu
from langchain_community.vectorstores import FAISS
vectorstore = FAISS.from_documents(documents=texts, embedding=embedding)
results, scores = vectorstore.similarity_search_with_score("How to use LangChain?", k=3)
for doc, score in zip(results, scores):
    print(f"Score {score:.4f}, content: {doc.page_content[:100]}...")

Full End‑to‑End Pipeline

from langchain_community.document_loaders import PyPDFLoader
from langchain_text_splitters import RecursiveCharacterTextSplitter
from langchain_openai import OpenAIEmbeddings
from langchain_chroma import Chroma

loader = PyPDFLoader("tech-docs.pdf")
pages = loader.load()
splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=50)
texts = splitter.split_documents(pages)
vectorstore = Chroma.from_texts(
    texts=[t.page_content for t in texts],
    embedding=OpenAIEmbeddings(model="text-embedding-3-small"),
    persist_directory="./vector_db"
)
results = vectorstore.similarity_search("What technologies does this book cover?", k=2)
print(f"Retrieved {len(results)} relevant chunks")
for r in results:
    print(f"- {r.page_content[:100]}...")

Day 3 Recap

Key components and their primary APIs:

Document Loader – PyPDFLoader, WebBaseLoader, etc.

Text Splitter – RecursiveCharacterTextSplitter Embedding – OpenAIEmbeddings, OllamaEmbeddings Vector Store – Chroma,

FAISS