How Do You Tell a Real RAG Engineer From a Prompt Engineer?

Retrieval-augmented generation has moved from a research pattern to a standard part of most document-heavy products in the space of about three years. The original formulation, published by Lewis and colleagues in 2020, described combining a parametric language model with a non-parametric retrieval index so that answers could be grounded in a corpus rather than in model weights alone. What has changed since is not the idea but the labour market around it: companies now hire specifically for it, and most have no established way of doing so.

Where to hire turns out to be the easier half of the problem. The routes are knowable and are compared further down this page. The harder half is telling, in a 45-minute conversation, whether the person in front of you has actually built a retrieval system or has been adjusting prompts and calling it one. That is where most RAG hires go wrong, so it is where this page starts.

Ask what they measured, and how. A prompt engineer changes the instruction and judges the result by reading it. A RAG engineer changes retrieval and judges it against a labelled evaluation set with numbers attached. Candidates who cannot describe their eval set have been tuning prompts and calling it retrieval work.

Four questions separate the two reliably in interview.

"Walk me through your evaluation set." The strongest signal available. A real answer describes how many questions, where the ground truth came from, who labelled it, and what the disagreement rate was between labellers. A weak answer describes eyeballing outputs.

"What did you try that made retrieval worse?" Anyone who has genuinely tuned a pipeline has made it worse at least once and can say so precisely. A candidate whose every change improved things has not been measuring.

"How do you handle a question the corpus cannot answer?" Retrieval systems fail most visibly when they confidently answer from irrelevant context. Good candidates talk about abstention thresholds and about surfacing uncertainty rather than about better prompts.

"What is in your chunking logic that is specific to this document type?" Generic answers indicate generic work.

None of this requires you to be an AI specialist yourself. All four questions are answerable by a non-specialist listening for whether the candidate reaches for measurements or for adjectives.

What Does a RAG Engineer Actually Need to Be Good At?

The core skill is measurement, not model choice. A capable RAG engineer can tell you whether a change to chunking, embeddings, or reranking genuinely improved answer quality, because they built an evaluation set before they started tuning. Vector database experience is the most over-weighted line on most job descriptions.

The work breaks into four areas, and candidates are rarely equally strong in all of them.

Ingestion and chunking. Getting documents into a form that retrieves well. This is far less glamorous than it sounds and matters far more than most teams expect. Splitting on fixed token counts destroys tables and clause structure; splitting on document structure requires actually parsing the format. Most retrieval failures that get blamed on the model originate here.

Embedding and index strategy. Choosing a representation, deciding whether hybrid keyword-plus-vector retrieval is warranted, and setting up the index so it can be rebuilt without downtime. Familiarity with a particular vector database is the easiest of these to learn on the job.

Reranking and grounding. Narrowing a candidate set to what actually answers the question, and constraining generation so the answer stays inside the retrieved evidence. This is where hallucination is contained or is not.

Evaluation. Building a labelled set of questions with known good answers, and measuring retrieval quality against it. This is the skill that separates engineers who can improve a system from engineers who can only change it.

For the distinction between this role and adjacent AI roles that teams frequently confuse it with, see AI agent developer vs RAG vs ML engineer, which covers where each one belongs rather than repeating the definitions here.

What Are the Ways to Hire a RAG Engineer?

There are four places to hire a RAG engineer: employ one directly, contract one through a freelance marketplace, buy the outcome from an AI consultancy, or place one through a managed remote workforce. They differ most in who employs the person and who absorbs the loss when the retrieval pipeline underperforms.

The question "where do I hire one" usually turns out to be four different questions wearing the same coat. Who employs the person. Who directs their work day to day. How the money is structured. And who is left holding the problem when the system answers wrongly. Those four answers vary far more between routes than the headline rate does.

Here is how the routes compare on the dimensions that actually differ.

Four routes to hiring a RAG engineer, compared on employment, cost structure, and where risk lands.
Route Who employs the engineer Time to hire Cost model Who carries the risk if it does not work
In-house hire You. Full legal employer, including benefits and termination obligations. Not published; varies by market and seniority Salary plus employer costs, plus any recruiting fee You. Severance, rehiring, and the sunk ramp time are all yours.
Freelance marketplace Nobody. The engineer is an independent contractor. Not published; depends on the individual's availability Hourly or per-milestone, usually plus a platform fee Split and often unclear. Contract terms govern, and remedies are typically limited to withheld payment.
AI consultancy The consultancy. You buy an outcome, not a person. Not published; scoping usually precedes any start date Project fee or retainer, often milestone-linked The consultancy, but only to the limit written into the statement of work.
Managed remote workforce The provider is the legal employer abroad. You direct the work. F5 Hiring Solutions: shortlist in 7-14 business days (self-reported commitment, not an independent measurement) All-inclusive weekly rate covering salary, HR, equipment, and management Shared. The provider carries employment risk and replacement; you carry the technical direction.

Three of the four rows say "not published" under time to hire, and the fourth is our own published commitment rather than an independently measured benchmark. Both of those things are deliberate. There is no credible industry benchmark for how long it takes to source a RAG engineer through direct employment, a marketplace, or a consultancy. Figures circulate, but they trace back to vendor marketing rather than to survey data. Ask each provider for their own number, including us, and hold them to it. A self-reported service commitment and a measured industry average are different kinds of claim, and this table should not be read as if the F5 Hiring Solutions row were the latter.

If you already know you want the managed route and want the detail on sourcing from India specifically, that is covered separately in hire RAG engineers from India. The comparison here is deliberately route-level. For the same four-route comparison applied to AI hiring generally rather than to retrieval specifically, see how to hire AI experts: four routes compared.

What Does Each Route Cost and How Long Does It Take?

Only two of the four routes have publishable figures. US employment cost can be benchmarked to a BLS occupation. F5 Hiring Solutions publishes its rate and its shortlist timeline. Marketplace and consultancy pricing is negotiated per engagement and is not published, so no number is given for either here.

BLS does not publish an occupation called RAG engineer, or indeed AI engineer. The closest published occupation is Software Developers, and it is used here as a proxy rather than as a measurement of the role.

US benchmark and F5 Hiring Solutions rate. The BLS median is May 2025 OEWS. The fully-loaded figure is an estimate: median x 1.4265 (BLS ECEC, Dec 2025, USDL-26-0505).
Basis Occupation / role Figure
US median, BLS OEWS May 2025 Software Developers, SOC 15-1252 (proxy) $135,980 per year
US fully loaded (estimate, x1.4265) Software Developers, SOC 15-1252 (proxy) Approximately $193,975 per year
F5 Hiring Solutions, all-inclusive AI engineers, LLM engineers, AI/ML engineers From $600 per week
F5 Hiring Solutions, all-inclusive AI solution architects From $800 per week
Freelance marketplace Contractor, hourly or per milestone Not published; negotiated per engagement
AI consultancy Project or retainer Not published; scoped per statement of work

Two cautions on reading that table. The proxy is a proxy: SOC 15-1252 measures software developers as a whole, and retrieval engineering sits at the higher end of that distribution rather than at its median. And the 1.4265 multiplier is an estimate derived from BLS Employer Costs for Employee Compensation, not a figure any individual employer will match exactly. It is offered so the arithmetic is visible, not as a precise forecast of your own cost.

Which Route Fits Which Company?

Direct employment fits companies where retrieval quality is the product. Marketplaces fit bounded evaluations. Consultancies fit fixed deliverables with a defined end. A managed remote workforce fits an ongoing role you want to direct yourself without becoming an employer in another country.

Hire in-house when retrieval is the product. If answer quality is the thing customers are paying for, the person improving it should be inside the company, in the standups, hearing the support tickets. Do not outsource your core differentiator to any of the other three routes.

Use a freelance marketplace for bounded questions. "Is our corpus even retrievable" is a good contract. "Own our retrieval quality indefinitely" is not. Contractors are well suited to a defined evaluation with a deliverable at the end, and poorly suited to a system nobody owns after they leave.

Use a consultancy when you want the outcome, not the person. If the deliverable is a working pipeline handed over with documentation, and you have people who can maintain it afterwards, this is a reasonable purchase. The failure mode is the handover: a system built by people who left, maintained by people who did not build it.

Use a managed remote workforce for an ongoing role you direct. This is where F5 Hiring Solutions operates. F5 Hiring Solutions is the legal employer abroad, handles HR, payroll, and equipment, and assigns the professional full-time exclusively to one client. You direct the work as you would any team member.

Worth stating plainly: this route is the wrong choice in at least two situations. If the work requires a US-resident attestation for regulatory reasons, no offshore arrangement solves that. And if you need someone for three weeks, a full-time placement is the wrong instrument entirely; a marketplace contract fits better.

What F5 Hiring Solutions actually places

There is no standalone RAG engineer role on the F5 Hiring Solutions roster, and it would be misleading to imply a RAG bench that does not exist. Retrieval work sits inside four roles that F5 Hiring Solutions does place: AI engineers, LLM engineers, AI and ML engineers, and AI solution architects. Candidates are matched to the retrieval work specifically rather than assigned by title, and F5 Hiring Solutions delivers a shortlist in 7-14 business days from a database of 85,500+ candidates, with a zero-cost replacement in 7-14 days at any time.

Deciding

The four routes are not ranked, because the right answer depends on a question only you can answer: is retrieval quality something you intend to own permanently, or something you need built once. Companies that answer "own permanently" and then buy a project consistently regret it, and companies that answer "need it built once" and then hire a permanent seat carry a cost they did not need.

If the answer is an ongoing role you want to direct without setting up an entity abroad, book a call and we will scope it: https://calendly.com/joel-f5hiringsolutions/f5

Frequently Asked Questions

Where is the best place to hire a RAG engineer?

There is no single best place. Direct employment suits companies where retrieval quality is the product. A freelance marketplace suits a bounded evaluation or prototype. An AI consultancy suits a fixed deliverable. A managed remote workforce suits an ongoing role you want to direct yourself without becoming the legal employer abroad.

What does a RAG engineer actually need to be good at?

Chunking and embedding strategy, retrieval evaluation, reranking, and grounding. The hardest part is measurement: knowing whether a change to chunk size or a reranker genuinely improved answer quality, or only moved a benchmark. Vector database familiarity matters far less than the ability to build an evaluation set.

How do you tell a real RAG engineer from a prompt engineer?

Ask what they measured. A prompt engineer changes the instruction and judges the output by reading it. A RAG engineer changes retrieval and judges it against a labelled evaluation set with recall and precision numbers attached. If nobody can describe their eval set, they have been tuning prompts.

Should we hire a RAG engineer or an AI agent developer?

A RAG engineer builds systems that retrieve and ground answers in your documents. An AI agent developer builds systems that decide and act across tools. Many teams describe an agent problem and hire for retrieval, or the reverse. Establish which failure you actually have before writing the job description.

How long does it take to hire a RAG engineer?

It depends entirely on the route, and most routes do not publish a timeline. F5 Hiring Solutions delivers a shortlist in 7-14 business days. For direct employment, freelance marketplaces, and consultancies, ask each provider for their own figure rather than relying on a general benchmark, because none is published.

Can one RAG engineer handle a whole retrieval system?

For a single corpus with a defined user group, usually yes. Once you have multiple document types, access controls per user, and answer quality tracked as a business metric, retrieval work tends to split into pipeline engineering and evaluation. That is the point at which a second seat is worth considering.

Is a US-based RAG engineer worth the premium?

It depends on whether the work requires being in the room. Retrieval engineering is measurable remotely: the evaluation set is the supervision. Where US employment genuinely matters is regulated work requiring a US-resident attestation, or systems where the engineer must sit with clinicians or traders daily.

What does F5 Hiring Solutions place for retrieval work?

F5 Hiring Solutions places AI engineers, LLM engineers, AI/ML engineers, and AI solution architects who build retrieval systems, starting at $600 per week all-inclusive, with AI solution architects from $800. There is no separate RAG job title on the roster; retrieval work sits inside those roles.

Sources