Let’s start with a positive outcome.

A publisher, a reputable one with big journals, peer review pipelines, and all the jargon, attempted to introduce a generative AI tool to assist in training new editors. They introduced the tool as a friendly onboarding assistant, suggesting that users could ask it about manuscript triage. However, what they received instead was a digital hallucination machine that provided confident incorrect answers, fabricated committee names, and policy advice with the legal standing of a fortune cookie.

Then, someone suggested “guardrails,” and the entire situation underwent a significant transformation.

The publisher implemented hard-coded inputs, restricted model access to a limited corpus of trusted PDFs and HTML documents, and applied some prompt engineering to the system, essentially creating a laminated cheat sheet. As a result, the chatbot improved, albeit not in terms of intelligence, but rather in reducing the likelihood of incorrect responses. In this context, such progress is considered acceptable.

We are all expected to applaud here, recognizing the importance of restraint and control.

However, this success story has become a kind of parable, and like most parables, it carries a lesson and a potential misinterpretation. The lesson is that uncontrolled AI poses a danger. The misinterpretation, on the other hand, suggests that control is the ultimate objective.

The issue with guardrails lies in their inability to provide guidance on the desired direction. They prevent users from veering off the intended path, but they fail to offer a sense of direction or enhance their understanding of the situation. Essentially, they merely prevent crashes.

Enter RAG.

A Side Note from a Non-Publishing Colleague

(A brief interlude: a few weeks ago, a colleague of mine, who is not in publishing and is not even involved in anything XML-related, was discussing an internal project. The topic involved wrangling legacy system outputs into coherent documentation for customer support. You know, the kind of task that makes seasoned systems architects weep.)

I remarked, “That’s going to be a nightmare to train an LLM on.”

He responded, “Yeah, I’m thinking RAG.”

And that’s it. No preamble, no TED Talk, just RAG casually mentioned, as if we all agreed that it was the obvious next step. Perhaps it is now. Perhaps we have already surpassed the point where RAG needs to justify its existence.

Retrieval-Augmented Generation, or RAG, is a concept that challenges the notion of AI models possessing complete knowledge. Instead of relying solely on the model’s internal knowledge, RAG introduces a collaborative approach by providing it with curated chunks of knowledge in real time. This collaborative environment is akin to having an AI with a reading list, utilizing generative text backed by actual documents owned by the user.

This collaborative approach makes RAG particularly effective in addressing a critical challenge faced by the publishing industry: ensuring that AI models possess a deep understanding of the content they generate.

Now, let’s explore some practical examples of how RAG can benefit individuals and organizations involved in journaling.

Example 1: Production Challenges
Imagine a publishing company facing a complex production process. They receive structured content from various vendors, each with its own schema format, and a style guide that is often inconsistent and outdated. Additionally, the editors have their own preferences and rejection patterns that can be difficult to decipher. In such scenarios, RAG can provide a solution. It can compare the input content against the actual rules and guidelines defined by the publishing company, eliminating the need for manual intervention and ensuring consistency in the final output.

Example 2: Author-Editor Collaboration
Authors often struggle to determine if their work aligns with the journal’s scope and guidelines. Instead of searching the journal’s website, reading conflicting scope statements, and deciphering complex rejection patterns, authors can leverage RAG. By providing the AI with the last 500 accepted articles and the editors’ real comments as context, RAG can offer insights and recommendations, helping authors make informed decisions about whether their work is suitable for publication.

Example 3: Image Alt-Text Generation
Another area where RAG can make a significant impact is in the generation of alternative text (alt-text) for images. Traditional methods of generating alt-text often result in generic descriptions like “A graph” or “Line graph.” However, RAG can provide more meaningful and contextually relevant alt-text by identifying the relevant caption, data, and surrounding article. This allows for a more accurate representation of the image’s content, ensuring that it is accessible to people with visual impairments and provides valuable information for search engines.

Publishers possess what RAG desires.

RAG thrives on high-quality, well-structured, clean, and taxonomized content. Does that sound familiar?

XML-first workflows, controlled vocabularies, and authoritative metadata—these are elements that publishing has been developing for decades, often without a clear destination in mind. Could this be the moment they finally find their purpose?

RAG doesn’t merely tolerate your structure; it actively craves it.

⸻

Guardrails vs. Maps

This isn’t an anti-guardrail rant. Guardrails are essential; they prevent accidents.

However, they’re insufficient if your objective is to reach a destination. RAG provides the model with direction, purpose, and the ability to state, “Based on the following three documents, I believe this is the case.”

This shift eliminates the question of whether the AI will hallucinate. Instead, it focuses on whether we provided it with the appropriate material to work with from the outset.

In our next discussion, we’ll explore how to construct the system around the AI. We’ll delve into the practical aspects of process modeling and orchestration when we transition from playing with demos to deploying a real system.

Safety is important, but useful AI is where the real challenge lies.

#ssp2025