The submission explained that a large language model was trained on massive text data, generated text one token at a time by predicting probable next tokens, and produced summaries and continuations without possessing true understanding.

The module presented raw document data and scanned images. I cleaned tabular and JSON data with Pandas, standardized text (removed HTML, fixed spacing/currency/encoding), and applied denoising and blur filters to enhance images for OCR. The deliverables were Colab notebooks and a processed image.

The project processed SDF PDFs with PyMuPDF and other libraries, diagnosed missed text and malformed tables, applied regex and anchor-phrase logic to locate vendor, doc type, product, lot, part number, manufacture and expiration dates, and documented extraction issues and next steps.

I processed a scanned supplier PDF, ran Tesseract, PaddleOCR, and EasyOCR, and compared outputs. I noted layout-aware engines extracted more text and vertical words, found PaddleOCR produced the most comprehensive extracts, and recommended combining tools for edge cases.

The externship brief provided a certificate-style document. I chose a Hugging Face embedding model, used page-based chunking, implemented hybrid retrieval, and ran queries that returned extracted quality-control tests and storage-condition statements from the source.

A sample SDF document was queried for quality-control test methods. I ran three RAG configs changing top-k and threshold settings, logged retrieved chunks, best excerpts, and confidence scores, and concluded configuration C produced the clearest, most relevant answer.