Chat With Your Medical PDFs: How It Actually Works

What 'chat with your PDF' actually means

The phrase "chat with your PDF" has become a marketing staple. Dozens of tools claim to do it. But the range of quality behind that phrase is enormous — from tools that can barely extract text from a simple document to tools that genuinely understand the structure of complex, image-heavy professional literature.

For most business use cases, basic text extraction is sufficient. A contract is mostly paragraphs of text. A financial report has some tables. A meeting transcript is pure text. Extract the words, index them, and you have a searchable document chatbot. This is what most "chat with your PDF" tools do, and for business documents, it works well enough.

Medical documents are different. A clinical guideline has treatment algorithms embedded as figures. A journal article has outcome data in tables that are as important as the text describing them. A textbook chapter has diagnostic images — histopathology slides, radiological scans, surgical photographs — that are essential to understanding the clinical content. A tool that ignores these elements is ignoring a significant fraction of the medical knowledge in the document.

This article walks through how Medevidex handles medical PDFs — not as a sales pitch, but as a technical explanation of what happens between uploading a document and asking your first question. Understanding the process helps you understand both the capabilities and the limitations of the tool.

Step one: upload your documents

The process starts with uploading. You select a PDF — a guideline chapter, a journal article, a textbook section — and upload it to Medevidex. The document is stored in a private, isolated environment. No other user can see it. No staff member can access it. It is yours.

You can organise documents into collections before or after upload. Collections work like folders, but with an important difference: when you ask a question, you can scope your query to a specific collection. This means your bladder cancer guidelines and your prostate cancer guidelines stay separate, and answers come from the collection you choose — not a blended mix of everything.

Upload is straightforward — select files, choose a collection, and the processing begins automatically. You can upload multiple documents at once, and processing happens in the background. You do not need to wait for one document to finish before uploading the next.

Step two: AI processes and indexes the document

This is where the engineering matters, and where most tools fall short with medical content. When Medevidex receives a PDF, it runs the document through a multi-stage ingestion pipeline that was built specifically for the complexity of medical literature.

Text extraction and OCR. The first stage extracts text from each page. For digitally-created PDFs, this is straightforward. For scanned documents — which are common in medicine, especially older textbook chapters and archived journal articles — the system uses optical character recognition to convert page images into searchable text. The OCR engine was selected specifically for its accuracy on medical terminology, multi-column layouts, and mixed text-figure pages that are typical of clinical literature.

Figure and table extraction. This is where Medevidex diverges from general-purpose document tools. The pipeline identifies figures, tables, and clinical images on each page, extracts them as discrete elements, and indexes them alongside the surrounding text. A treatment algorithm figure is not treated as decoration — it is recognised as a distinct piece of clinical content that may be the answer to a question about management pathways.

Caption and context binding. Figures and tables in medical documents have captions — "Figure 3: Management algorithm for non-muscle-invasive bladder cancer" — and these captions are bound to their corresponding visual elements. When you later ask about a management algorithm, the system can retrieve both the figure and its caption, along with the page reference.

Semantic chunking. The extracted text is split into meaningful segments — not arbitrary blocks of 500 characters, but semantically coherent chunks that respect heading boundaries, paragraph structure, and clinical context. A section about surgical technique stays together as a chunk. A paragraph about drug dosing is not split in the middle of a dosing table. This semantic chunking is what allows the AI to retrieve relevant passages that make sense when read in isolation.

Embedding and indexing. Each chunk of text, each figure caption, each table description is converted into a mathematical representation — an embedding — that captures its meaning. These embeddings are stored in a vector database. When you later ask a question, your question is also converted into an embedding, and the system finds the chunks whose meaning is closest to your question. This is semantic search — it finds relevant content even when the exact words differ.

Step three: ask questions and get cited answers

Once your documents are processed, you can start asking questions. The interface is a chat — you type a question in natural language, and the system responds with an answer that cites specific documents, pages, and passages.

Behind the chat interface, several things happen in sequence. Your question is converted into an embedding. The system searches your document collection for the most relevant chunks. Those chunks — the actual text from your PDFs — are assembled into a context window along with your question. The AI then generates an answer based on that context, citing the specific chunks it used.

Each citation in the response links back to the source document and page. You can click through to see the original page in context — the full text, the surrounding figures, the original layout. This is the verification step that makes the tool clinically useful: you do not have to trust the AI's interpretation. You can check it against the source in seconds.

Why medical documents need special treatment

General-purpose document chatbots were built for contracts, reports, and business correspondence. These are text-dominant documents where the words carry nearly all the information. A figure in a business report is usually a bar chart that restates what the text already says.

Medical documents are structurally different, and the differences matter for clinical accuracy.

Figures contain treatment algorithms. In clinical guidelines, the management pathway is often presented as a flowchart figure. The text describes the rationale, but the figure is the actionable summary — the thing you actually follow when making a decision. A tool that skips figures misses the most clinically useful element on the page.

Tables contain trial outcomes. The primary endpoint, hazard ratio, confidence interval, and p-value of a landmark trial are in a table, not in a paragraph. When a colleague asks "what was the hazard ratio in KEYNOTE-045?", the answer is in a table. A tool that cannot read tables cannot answer the question.

Clinical images are diagnostic, not decorative. A histopathology image showing muscle invasion in a bladder biopsy, a CT scan demonstrating lymphadenopathy, a cystoscopy image showing a papillary tumour — these are not illustrations. They are clinical evidence. In a textbook, they teach pattern recognition. In a case report, they support the diagnosis. A tool that treats all images as decorative clip art fundamentally misunderstands medical literature.

Practical examples: what you can ask

The value of the tool becomes concrete through examples. Here are the kinds of questions that work well — and the kinds of answers you get back.

Staging systems. "What is the TNM classification for muscle-invasive bladder cancer according to the 2024 EAU guideline?" The system retrieves the relevant staging table from your uploaded guideline, cites the page, and presents the classification. You can click through to see the original table in context.

Treatment algorithms. "What is the first-line treatment recommendation for cisplatin-eligible metastatic urothelial carcinoma?" The system finds the management algorithm figure and the supporting text, cites both, and synthesises an answer that includes the recommendation grade. The figure itself is retrievable — you can see the actual flowchart from your guideline.

Drug dosing and schedules. "What is the dose and schedule for gemcitabine-cisplatin in the neoadjuvant setting?" The system locates the dosing table or text passage, cites the source, and presents the regimen. No ambiguity about which guideline the dosing came from.

Comparative questions across documents. "How does the EAU recommendation on neoadjuvant chemotherapy for MIBC compare to the AUA recommendation?" If both guidelines are in your collection, the system retrieves the relevant passages from each and presents them side by side, cited to their respective source pages.

Trial-specific questions. "What was the primary endpoint of the CheckMate 274 trial?" The system finds the results table and the primary endpoint description from the paper you uploaded, cites both, and gives you a specific, verifiable answer.

What the tool does not do

Honesty about limitations is more useful than marketing claims. Here is what Medevidex does not do, and why these boundaries exist.

It does not search public databases. If you want to discover new papers on a topic, you need PubMed, Google Scholar, or a tool like Consensus. Medevidex only searches documents you have uploaded. This is a feature, not a limitation — it means every answer comes from a source you have vetted — but it means you are responsible for curating your own library.

It does not replace reading. The tool accelerates retrieval and cross-referencing, but clinical understanding requires reading the full text, understanding the methodology, and applying judgment. An AI-generated summary of a trial is not the same as reading the trial.

It does not provide clinical recommendations. Medevidex retrieves and cites what your documents say. It does not tell you what to do with a specific patient. Clinical decision-making involves factors — patient preference, comorbidities, local expertise, resource availability — that no document retrieval tool can assess.

The citation standard: document, page, passage

A citation that says "Source: EAU Guidelines 2024" is not a citation. It is a gesture in the direction of a citation. A 200-page guideline contains dozens of topics, hundreds of recommendations, and thousands of data points. Telling you the answer came from "the guideline" does not help you verify anything.

Medevidex citations include three elements: the document name, the page number, and the specific passage or chunk that was used to generate the answer. This means you can go directly to page 47 of the EAU guideline, read the paragraph the AI cited, and confirm that the answer accurately reflects what the source says. The verification loop takes seconds, not minutes.

Getting started

The workflow is simple: upload a PDF, wait for processing to complete, and start asking questions. There is no configuration, no setup, and no learning curve beyond what you already know from using any chat interface.

Start with one document you know well — a guideline chapter or a key article. Ask it a question you already know the answer to. Check the citation. If the answer matches your understanding and the citation takes you to the right page, you have a tool you can trust. Build from there.

Medevidex is free to start. Upload your first document and see how it handles the complexity of medical literature — the figures, the tables, the clinical images that general-purpose tools miss.

Read more

How to Use AI to Review Medical Literature Faster · Why Medical AI Needs Page-Level Citations · Organising Your Medical Library With AI