Loaders, Indexes & Vectorstores in LangChain: Question Answering on PDF files with ChatGPT

A practical LangChain pipeline for turning PDFs, YouTube transcripts, and plain text into question-answering over embeddings is the core takeaway—and the workflow is built to show every moving part, from loaders to vector stores to retrieval-based QA. The setup starts with simple document ingestion (a text file), then scales to richer sources (YouTube transcripts and an unstructured PDF), and finally wires everything into a single retrieval + ChatGPT-style answering chain. The result is a repeatable method for asking targeted questions about a document’s contents without manually summarizing or searching through it.

The walkthrough begins with loading a custom essay from a local text file using LangChain’s text loader, then indexing it with a vector store index creator backed by Chroma DB. A short query—“Why someone in today’s world would read”—is answered using text DaVinci 003, with the response returned alongside source attribution pointing back to the essay. The emphasis is on how quickly LangChain can turn raw text into a searchable embedding space, enabling concise answers grounded in the indexed content.

Next, the pipeline demonstrates how swapping loaders changes the input type while keeping the rest of the retrieval approach intact. A YouTube loader pulls a transcript (via the YouTube transcript API) for a specific video, producing document objects with page content and metadata. Those transcript documents can then be indexed the same way and queried for information—showing how video knowledge can be converted into the same embedding-based question-answering flow.

The most important scaling step is handling PDFs. Using an unstructured PDF loader, the pipeline extracts page content from an older resume PDF (Andre Karpathy’s CV). Because long documents exceed model context limits, the transcript is split into overlapping chunks using a recursive character text splitter. The chunk size and overlap are chosen to preserve continuity across boundaries—overlap helps prevent the loss of context when a relevant idea spans two chunks. This chunking produces multiple smaller document segments suitable for embedding.

Embeddings then convert each chunk into vectors. The walkthrough compares sentence-transformers embeddings from Hugging Face (768-dimensional vectors) with OpenAI embeddings (larger dimensionality, described as roughly about twice the sentence-transformers size). After embeddings are created, they’re stored in a vector database. Chroma DB is used both in a non-persistent mode and with persistence to disk (via a persist directory and DuckDB-backed storage), so the indexed representation can be reused later.

Finally, a RetrievalQA chain combines the Chroma retriever with a ChatGPT model (temperature set to 0) to answer questions directly from the indexed PDF chunks. Example prompts include extracting the candidate’s work experience in no more than two sentences, generating a background summary in up to three sentences, and assigning a 0–10 likelihood rating for becoming a top-tier deep learning researcher two years later with an explanation. Across these examples, the pipeline demonstrates a consistent pattern: load → split (for PDFs) → embed → store (vector DB) → retrieve relevant chunks → generate grounded answers.

Overall, the workflow matters because it turns unstructured personal data (PDFs, transcripts, text) into a queryable knowledge base with minimal glue code, while explicitly addressing the two practical bottlenecks that usually break naive approaches: document heterogeneity and context-length limits.