Why does an AI answer better when it reads my own PDFs?

The difference is not that one system is smarter. It is that the two are being asked different questions. "What does the literature say about X" asks a model to reconstruct a body of work from statistical memory, which is a task with no verification step built into it. "What do these forty papers say about X" asks it to read and summarize text that is present, which is a task where every sentence can be traced.

That is why the same underlying technology produces such different reliability in the two settings. Published studies of generated references, where the model was asked to recall, report fabrication rates ranging from roughly a fifth to two thirds depending on the model and the task. Those numbers describe recall, not reading.

The engineering literature names the two settings directly. Lewis and colleagues separate parametric memory, which is what the model absorbed during training, from non-parametric memory, which is an external index the model reads at answer time (Lewis et al., 2020, p. 1). Their system retrieved passages from a December 2018 Wikipedia index of 21 million 100-word documents and then wrote the answer conditioned on those passages. The point is not that the model holds more facts. The evidence can be specific to your question, swapped when the literature moves, and inspected afterwards as the basis for the sentence you got (Lewis et al., 2020, p. 4).

The measured gains are real but conditional. On Natural Questions, RAG-Sequence reached 44.5 Exact Match against 34.5 for closed-book T5-11B and 41.5 for DPR, and on CuratedTrec it reached 52.2 against 46.8 for REALM and 50.6 for DPR (Lewis et al., 2020, p. 6). RAG-Token led WebQuestions at 45.5. The authors report a new state of the art across the open-domain question answering tasks, with the qualification that this did not hold on the T5-comparable TriviaQA split (Lewis et al., 2020, p. 5). Supplying documents helps when the retrieval step actually surfaces the right passage, which is why a tightly scoped library beats a sprawling one.

There is a second benefit that gets less attention. A corpus you assembled is a corpus you can defend. When an examiner asks why a claim rests on the studies it rests on, "these are the papers I selected against these criteria" is an answer. "The tool surfaced them" is not.

The trade is coverage. A system confined to your library cannot tell you about a paper you never added, and it will say so rather than invent one. That silence is the feature: a gap you can see is a gap you can go and fill.

What is retrieval, and why does it reduce fabrication?

Retrieval means finding the relevant passages in a defined set of documents first, then writing the answer from those passages. The model is no longer being asked to remember a literature, only to summarize text in front of it, which removes the situation where invented references are produced.

The reference cannot be invented because it is not being produced from memory at all. It is the document the passage came from, which you put there, and which you can open.

This is also why the size of the corpus matters less than its relevance. Forty well-chosen papers produce better answers than four thousand assorted ones, for the same reason a focused reading list beats a library.

Does retrieval eliminate errors entirely?

No. The passage retrieved can be the wrong one, the summary of it can overstate what it says, and a claim can be attributed to a paper that was merely reporting someone else's finding. Retrieval changes the error from invented sources to misread ones, which is smaller and checkable.

The reported-versus-argued distinction is the one worth watching. Papers spend much of their length describing other people's work, so a passage containing a claim is not evidence that the paper's authors endorse it.

Grounding also does not make fluent prose trustworthy. Smith and colleagues warn that language models produce seemingly correct but false answers, especially on complex questions, and that polished writing creates a halo effect in which readers take fluency for accuracy (Smith et al., 2024, p. 5). A separate paper by Smith, Greaves, and Panch describes model output as probabilistic and variable, prone to generating non-factual narrative detail (Smith et al., 2023, pp. 1-2). Your PDFs constrain what the model reads. They do not audit what it wrote.

Two things follow for how you work. Smith and colleagues advise stating your context and constraints explicitly, because that is what makes a response relevant rather than generic (Smith et al., 2024, p. 2). A curated PDF collection does that by construction, since a seminar reading list or dissertation bibliography is already a constraint on the answer space. They also require subject-matter review of anything generated, and note that summaries are harder to check than citations or code (Smith et al., 2024, p. 7). Note the limit of the transfer: Lewis and colleagues tested Wikipedia, not students' libraries, so the mechanism carries over but the benchmark numbers do not.

The practical consequence is unchanged by any architecture: open the citation, read the passage, and confirm it supports your sentence. Retrieval makes that check take ten seconds instead of an afternoon, which is the actual improvement.