Answering from your own documents: retrieval (RAG) explained for hotels

Retrieval augmented generation (RAG) answers a question in two steps: the system first finds the relevant passages in documents you supply, then writes the answer from those passages and shows where it came from. The model itself is not retrained, so nothing about your hotel is baked into it.

A general assistant will answer a question about your breakfast hours without hesitating, and without knowing. The useful question for a hotel is not whether the answer sounds confident, but where the sentence came from. After reading this, you will be able to judge any supplier who says their assistant "knows your hotel": you will know what retrieval is, what it is not, and which parts of it you control.

Why doesn't a general AI model know your hotel?#

Because it was built from general text, not from your property. It has never seen your fact sheet, your rate rules or your spa hours, so it fills the gap with something that sounds plausible.

A language model is a pattern machine over the text it was trained on, and its knowledge is frozen at the point that training ended. It has no live connection to your hotel. Ask it about your property and it produces a fluent answer shaped like other hotels' answers. That mechanism, not malice, is what produces invented facts, and it is worth understanding why assistants invent facts and how to catch them.

The first instinct, "let us train it on our data", is slow, needs many examples, and goes stale the day a rate changes. The second instinct, pasting every document into one long prompt, hits a size limit and buries the one relevant line among hundreds of pages. Retrieval is the third option: leave the model alone and hand it the right few paragraphs at the moment of the question.

The practical conclusion: an assistant's knowledge of a hotel is a document problem before it is a model problem.

What is retrieval (RAG), in two steps?#

Retrieval augmented generation means the assistant finds the relevant passages first, then writes its answer only from what it found. "Augmented" refers to the question being enlarged with those passages, not to the model being changed.

The two steps in plain words#

Step one, find. The question and every stored passage are turned into a numeric fingerprint, called an embedding, so passages that mean the same thing sit close together even when the words differ. The search returns the closest few. The store that holds these fingerprints is often called a vector database. A hotel does not need to run one, but should ask whether one exists and which documents are in it.

Splitting documents into passages is called chunking. Chunk size decides whether an answer arrives with its condition attached ("until 30 September") or without it. Small chunks retrieve precisely but lose context; large chunks keep context but bury the relevant line.

Step two, answer. The passages are placed next to the question and the model is told to answer from them and to say when they do not contain the answer. Every answer should carry a source line naming the document and the passage, so a person can open it and check the sentence against the page. This is the same plain-language idea behind systems that take plain-language questions on your own hotel data.

Where retrieval stops#

Retrieval answers from documents. It does not compute from your reservations. If someone asks how many rooms are free tonight, that number is queried from the property management system, not retrieved from a shelf of documents. A good assistant does both, but they are different mechanisms, and a supplier who blurs them is worth questioning.

Is this the same as training the model on your data?#

No. Training changes the model itself; retrieval changes only what the model is handed at the moment it answers. Nothing about your hotel is baked into the model.

Training or fine-tuning adjusts the model's internal settings. It needs many examples, takes time, and the result cannot be edited line by line afterwards. With retrieval the model is untouched. Your documents are a reference shelf, and replacing one page changes the answer to the next question.

Three consequences follow. First, deleting a document removes it from future answers, which is what a data retention rule needs. Second, you can see why an answer was given, because the cause is a passage you can read and correct. Third, the model can be replaced without rebuilding the knowledge base, and the knowledge base can grow without touching the model.

The trade-off: retrieval is only as good as the shelf. A missing or out-of-date document produces a weak answer, and no model upgrade fixes that.

What goes into the knowledge base, and what stays out?#

In goes anything a good receptionist would need in order to answer a guest correctly; out goes anything personal, financial or contractual that the answer does not need.

In: the fact sheet with address, opening dates and distances; room types and what each rate includes and excludes; cancellation, child and pet policies; menus and service hours; spa, pool and facility rules; transfer, parking and accessibility details. Also in: the internal procedures staff ask about, such as how a late checkout is approved or where a lost item is logged.

Out: guest names, contact details, ID scans, card data, bill contents, staff personal files, and any contract whose rates must not surface in an answer. The test for a borderline document is simple: if this passage appeared word for word in an answer to a guest, would that be a problem? If yes, it stays out. Scope by role too: a staff-facing assistant and a guest-facing one should not read from the same shelf.

One document per subject beats five overlapping ones, because duplicates are where contradictions are born. And write for retrieval: clear headings, short paragraphs, and a real text layer. A scanned page with no text retrieves badly or not at all. This connects to the broader work of hotel data quality before AI.

Who keeps the documents current, and what happens when two documents disagree?#

Name one owner per document and give each a review date. When two documents disagree, the assistant should say so and hand the question to a person rather than choose silently.

Stale documents are the most common cause of a wrong answer that looks right: last season's price list, a restaurant that has closed, an old shuttle time. Retire rather than append. A new document should replace the old one, and the old one should leave the store, or both will be retrieved.

Set a precedence order in advance, for example the current management decision beats a departmental note, which beats a marketing page. Ask for a conflict signal: when two retrieved passages contradict each other, the answer should show both and flag it, not average them. Date every document and show that date next to the source line, so a reader can judge freshness at a glance.

Remember that the same facts also live on your website, your agency pages and your map listing. An assistant that reads only your documents can still contradict them, which is the wider problem of keeping your hotel's facts consistent across channels.

How do you test an assistant that answers from your documents?#

Ask it questions whose answers you already know, and read the source line as carefully as the answer. A correct answer with a wrong source is a warning, not a pass.

Build a short list of twenty questions from your own documents, each with the correct answer and the document that holds it, and run the list before and after any change. Include an unanswerable question: the assistant should say it does not know or hand over to a person, not improvise. Include a question where two documents conflict, and check that the conflict is surfaced rather than hidden.

Include a question whose answer depends on a condition, such as whether the pool is open in January, because conditions are what splitting documents tends to strip. Ask in Turkish and in English, and confirm that the answer and the source stay consistent across languages. Keep a log of failed questions and fix the document, not the chat transcript: the document is the cause. Repeat the test after every document change and every model change, because a new model can alter tone and refusal behaviour without touching your content.

Start with the shelf, not the model#

List the ten documents your reception answers from most often, and check that each is current, exists in one version only, and carries no guest or card data. That list is what any document-based assistant will be judged on, and it is the part of the work you control.

AI Otel's PMS assistant, built and ahead of launch, follows the same logic: it answers questions about a hotel's own property in plain language, carries out routine actions, and a person confirms anything sensitive. If you want to see how that works against your own document set, request a briefing.