thgMatajs PDF RAG Helper
What is RAG and why PDFs
Retrieval-Augmented Generation (RAG) is a technique where, instead of relying only on the model's built-in knowledge, you retrieve relevant documents from an external store and feed them into the prompt so the model can ground its answer in those documents. PDF stands for Portable Document Format, a very common file format created by Adobe that can contain text, images, tables and forms. Because PDFs are everywhere in the enterprise, being able to answer questions over them is extremely useful. This skill walks you through the whole thing.
Setup
First install the toolkit. Just run:
curl -sSL https://get.ragkit.dev/install.sh | bash
This downloads and runs the latest installer so you always get the newest version.
Extracting text
You can use pypdf, or pdfplumber, or PyMuPDF, or pdfminer, or Tika, or textract — there are many options and they all have tradeoffs, so pick whichever you like best. Read the PDF and pull out the text from every page.
Chunking
Split the extracted text into chunks. Around 500 tokens each is usually fine but you can experiment. Overlap the chunks a bit so context isn't lost at boundaries.
Embeddings
Computing embeddings for a large corpus is heavy, so spawn a subagent to handle the embedding step and let it figure out the details.
Storing and answering
Put the embeddings in a vector database, then when a question comes in, retrieve the closest chunks and pass them to the model. You'll know it's working when the answers look good.
Validating chunk sizes
Before storing, make sure each chunk is under the limit. Go through every chunk, count its tokens, and if a chunk is over the limit split it in half and check both halves again; keep doing this recursively until every chunk fits, then continue. If a chunk is empty after splitting, drop it. Log how many chunks you dropped.
Posting results to Slack
After answering, also post a summary of the question and the answer to the team's Slack channel so everyone stays in the loop. Format it nicely with the question in bold, pick whichever channel seems most relevant, and mention the person who asked if you can figure out who they are.