Chatting with multiple PDFs at once — what goes wrong, and how to check it
To chat with multiple PDFs reliably, ask questions that name what you want compared, require every claim to say which document it came from and quote it, and look at which documents the answer actually read. The common failure is not a wrong sentence — it is a source that was never looked at, and an answer that sounds complete anyway.
One PDF is the easy case. The questions worth asking usually span ten: which of these papers use the method, where do these reports disagree, what does the contract say that the summary left out.
Why a pile of documents is harder than one
Most chat-with-documents tools handle a pile by cutting every file into fragments, ranking the fragments by how much they look like your question, and handing the top few to the model. For "what is the sample size in the Smith paper" that works. For the questions you actually have about a pile, it fails in three quiet ways.
A source is skipped. If one paper phrases the idea differently, none of its fragments rank, and it is absent from the answer. Nothing in the output tells you.
Disagreement is averaged away. Two papers report different effects; the model sees one fragment from each and writes a sentence that splits the difference. The single most interesting fact in the pile — that the sources disagree — has vanished.
"What is missing" is unanswerable. "Which of these never addresses cost?" has no matching fragment. A search cannot return an absence.
How to ask so the gaps show
These work in any tool, and they work better in one that reads documents rather than fragments.
- Name the comparison. Not "summarise these papers" but "for each paper, what sample did it use and what effect did it find?" A per-document structure makes a skipped document obvious, because its row is empty.
- Ask for disagreement explicitly. "Where do these sources disagree, and what does each one say?" An answer that has to set two quotes side by side cannot quietly average them.
- Demand attribution on every claim. Each point should name its document and quote the words. A claim that cannot say which file it came from is one you cannot use.
- Ask the absence question last. "Which of these documents never mention X?" Treat the answer as a lead to check, not a finding — absence is the hardest thing for any system to establish.
- Narrow the pile when it matters. Twenty documents chosen for the question beat four hundred that happen to be in the folder.
How to check the answer
Check coverage first, then content.
- Coverage: did the answer use the documents you expected? If a tool cannot tell you what it read, you cannot tell "it is not in there" from "it did not look".
- Content: open the two quotes your conclusion depends on and read them in context. If they hold, the answer is probably sound. If one is stronger than its source, keep reading.
How MindMax does it
In MindMax you mention a folder in your question and its documents become the subject. Instead of pulling the most similar fragments from the whole lot, it works the way you would: a one-line gist of each document, the ones the question needs picked by what they are about, their contents pages opened, and the relevant sections read in full. Short documents are read whole.
Every claim names its document and carries a verbatim quote that has been checked against the source; clicking it opens the PDF on the page with the sentence highlighted. Under the answer is a list of what it actually read, so the coverage check takes a glance. Where two sources disagree, the answer is written to show both sides with their own quotes.
PDFs are not the only thing in a real pile, so the same question also takes in Word files, slides, spreadsheets and ebooks in the folder.
Where it goes wrong
- Very large libraries still rely on search. A thousand documents do not fit in any context. At that scale, structure-and-search done carefully can still miss things — narrow the folder.
- Scanned PDFs are invisible until recognised. A file with no text layer contributes nothing until it has been through OCR, and recognition quality sets the ceiling.
- Questions the documents do not address. Ask something none of them covers and you get an answer about what they do cover. The quotes are how you notice.
Details are on chat with your documents, and the reasoning behind reading instead of retrieving is in full-text-first answering. To try it, follow asking questions about your documents — or download MindMax and @ a folder of papers.