Turning messy pharmaceutical documents into searchable data

Pharmaceutical supply chains depend on shipping labels, purchase orders, logistics forms, invoices, and regulatory records. These documents often carry the same kinds of information, but almost never in the same format — dates, tables, and key fields shift by vendor, template, and scan quality.

During my Pfizer Supply Chain Document Processing Externship, I worked on an AI-powered pipeline to ingest scanned pharmaceutical documents, extract important fields, classify document types, and make information searchable across large collections of files.

A date might appear as 07/12/26, 12 July 2026, or as a handwritten note. Important fields may sit inside tables, split across columns, print over noisy backgrounds, rotate on scanned pages, or land in different places depending on the vendor and document type.

Traditional rule-based software struggles here. A parser built for one layout can fail when a supplier changes a template, a scan is blurry, or a field shows up somewhere unexpected. Even after text is extracted, the system still has to decide what each value means, keep its relationship to nearby labels, and produce reliable structured data for supply-chain and compliance decisions.

That gap — between messy pharmaceutical documents and trustworthy structured data — is what this pipeline was built to close.

Digitally generated PDFs go through PyMuPDF and pdfplumber to pull text and layout. Scanned or image-based pages are prepared with OpenCV and PIL — resizing, cropping, contrast adjustment, and noise reduction — before OCR engines like Tesseract, PaddleOCR, and EasyOCR are evaluated across layouts, scan qualities, and document types.

Extracted fields land in structured JSON, keeping both the recognized values and their page coordinates. Classification and routing decide how each document should be processed instead of forcing one extraction strategy onto every file — making the system more resilient across shipping records, labels, purchase orders, and other clinical supply documents.

Finally, content is indexed with a RAG pipeline. LlamaIndex, FAISS, and Chroma support vector retrieval; models such as Gemini, Mistral, or Phi-2 generate answers grounded in the most relevant sections. A Gradio or Streamlit interface lets users search the collection and ask natural-language questions without paging through thousands of documents by hand.