Guides · Choosing an approach

Do you need RAG, or just better search?

Short answer

If people know which document they need and struggle to find it, that is a search problem and RAG is an expensive way to solve it. RAG earns its cost when the answer is spread across several documents, or when the question is about what the documents mean rather than where they are.

Find vs answer
The distinction that decides the whole question
Citations
Non-negotiable: every answer names document, version and clause
Retrieval
Where quality is won — not in the model
Updated August 2026 · Written for people scoping a project, not for search engines

What RAG is, without the jargon

Retrieval-augmented generation means the system searches your own content first and then writes an answer using only what it found, with the sources attached. The model does not answer from memory; it answers from your documents, which is what makes the output checkable and keeps it current when the documents change.

That is the whole idea. Everything else — chunking strategies, embeddings, rerankers, hybrid search — is implementation detail in service of one thing: putting the right passages in front of the model before it writes.

The question that decides it

Ask what people are actually doing when they fail today. If they know the document exists and cannot locate it, the problem is retrieval and findability, and a well-configured search over well-tagged content will be faster, cheaper and easier to trust.

If they can find the documents and still cannot answer the question — because the answer requires reading four contracts and comparing their termination clauses — that is the case RAG was built for. The system reads all four and shows you where each part of its answer came from.

01

Search is enough when: the answer lives in one known document, the vocabulary is consistent, and the corpus is well structured.

02

RAG earns its cost when: the answer spans documents, the same concept is worded differently across them, or the user cannot name what they are looking for.

03

Neither will help when: the content is out of date or contradictory. Both approaches faithfully surface a wrong answer that is sitting in your files.

Where quality actually comes from

The common assumption is that answer quality tracks the model. In practice it tracks retrieval. If the right passage is not in the context, no model recovers; if it is, most current models produce a good answer.

That is why the work concentrates on the unglamorous end: how documents are split, how versions are handled so a superseded policy does not answer as current, how permissions are enforced so a retrieval never returns something the asker may not see, and how the system behaves when it finds nothing relevant. That last one matters more than it sounds — a system that says "I could not find this" is trustworthy; one that produces a plausible paragraph anyway destroys the trust in a week.

Citations are the product

For anything used in a professional or regulated context, the answer is not the deliverable — the answer plus its provenance is. Document, version, clause. A lawyer, an auditor or a claims handler cannot act on an unattributed paragraph, and will not.

It also changes the failure mode from dangerous to harmless. With citations, a wrong answer is caught in seconds by the person checking the source. Without them, it is repeated in a client email.

What it takes to run one

01

A corpus somebody owns. Someone has to decide what is authoritative and what is superseded. That decision cannot be automated.

02

Permissions that already exist. Retrieval must inherit your access model. Building one during the project is a much larger job.

03

An evaluation set. Thirty to fifty real questions with known good answers, written before launch. It is how you tell tuning from guessing.

04

A refresh path. Documents change. If reindexing is manual, it stops happening within a quarter.

Follow-up questions

What people ask next.

Technically yes, and it will faithfully answer from the outdated files too. Deciding what is authoritative is a prerequisite, not something the system resolves for you.

Still unsure whether your process is worth automating?

Bring us the process. We take it apart with you at no charge and give you a straight answer, including when the answer is no.