Back to work

Context-Aware PDF Question Answering

PDFs hold useful information, but asking them direct questions is awkward. I built the retrieval path myself so answers could be generated from the document rather than from an unconstrained prompt.

View source

What stayed with me

Building the retrieval path made the limits clearer: answer quality depends as much on chunking and context selection as it does on the model.

Core behavior

  1. 01PDF ingestion and chunking
  2. 02Vector retrieval
  3. 03Context-grounded answers
PythonLangChainChromaDBGemini Flash