read first
Local AI for PDFs: check the document before buying more memory
A poor PDF summary can look like a model-size problem. Yet table values may have become jumbled during extraction, or a scanned page may never have been read correctly. Follow one document through the pipeline before choosing new hardware.
Start by copying one page
Copy a representative page containing both prose and a table into a text editor. Check reading order and whether numbers remain beside the correct labels. A scan without usable text may require OCR or a visual-input path. A later model change cannot recover text that never reached it. For financial figures or dates, verify extraction before judging fluent summaries. Use a public document with a similar layout for initial setup rather than exposing sensitive originals unnecessarily.
A whole-document summary is not a single lookup
Summarizing an entire argument differs from locating one clause. Retrieval can select relevant passages rather than sending every page on each request. But if retrieval misses the essential passage, a capable answer model can still fail. Combining chapter summaries can also lose cross-chapter exceptions. Define the question and inspect what source material actually reaches the model. A large document collection does not, by itself, require the maximum context setting.
Long documents expose the wait before the answer
A model can feel quick in chat yet take much longer to begin after a document. File extraction, input processing and answer generation contribute different delays. Our prefill comparison estimates model input processing; it does not reproduce the entire PDF extraction pipeline. In your application, also time upload through completion and compare it with a request whose text is already prepared. Knowing where the wait occurs helps distinguish a preprocessing problem from an inference problem.
Keep three questions with known answers
A request to summarize the document is hard to score. Ask for a date, a table value and something the document does not contain. Verify the first two against the original and check whether the third receives an honest unknown. Request a location, but open it rather than trusting a plausible page number. Repeat this small set when changing models or retrieval settings. It is not a universal quality score, but it is a firmer basis than judging whether the prose sounds more fluent.
Buy for the way you use documents
If you read whole long documents daily, input processing and memory headroom both matter. If you retrieve a few passages, retrieval accuracy and answer quality may be the first priorities. Storage for the collection is not the same as inference memory. Validate extraction and answers on representative material, then compare hardware at that input length. If privacy motivated the project, check OCR, embeddings, retrieval and external tools too: owning a local machine does not make every connected step offline.