Document search with verifiable sources

The context

A useful document assistant must do more than produce a fluent answer. It needs to extract readable text, retrieve relevant passages and give the reader a practical way to check the source. The prototype explores that workflow for people reviewing complex PDFs.

How I worked

• Prepare the documents. Combined PDF text extraction and OCR to handle text-based and scanned content. Explored document classification and routing before splitting text into chunks for retrieval.
• Find supporting passages. Used FAISS for semantic search and connected retrieved context to Qwen 2.5. Kept extraction, retrieval and answer generation as distinct stages so failures could be investigated separately.
• Make answers reviewable. Built a Gradio interface for document questions, with supporting passages and page references rather than presenting generated answers without context.
• Evaluate the trade-offs. Compared retrieval quality with answer quality and explored how model size and response time affect the experience.

Deliverables & takeaways

The deliverable is a working document-questioning prototype, documented in the linked Python notebook. It brings OCR, classification, retrieval and a user interface into one workflow.

A key lesson was that retrieving a relevant passage does not guarantee a correct answer: the response and its citation still need to support the particular claim. This shaped how I assessed the system and presented its sources.

Scope & limitations

The prototype was developed in an externship; it is not a deployed Pfizer system or a tool for clinical or compliance decisions. Human source verification remains necessary. Structured tables, citation checks and evaluation on independent documents are areas for further work.

Company-specific document examples and sensitive evaluation details are withheld. The technical notebook remains available to show the implementation.