# LLM fine-tuning and RAG | Slash Digital Lab, Bogotá

Source: https://slash-digital.io/en/fine-tuning/

Artificial intelligence RAG when knowledge must stay current and citations matter. Fine-tuning when style, language or a narrow skill doesn't fit in context. Hybrid when both are true.

- Default RAG
- When it fits Adapters · LoRA
- Rule Evaluation > hype

Start a project
See real work

In short
Slash Digital Lab chooses between LLM fine-tuning and RAG by your constraints, not by recipe. If your data changes daily, RAG. If tone or format must be yours, fine-tuning. If compliance requires keeping knowledge in-house, local inference with open models.

Read the full criterion in Fine-tuning vs RAG and how we choose models in How we choose a model. For private, self-hosted AI: how to size your own AI server and open-weight models in 2026.

What we deliver

## What we build

The cheapest path that meets the quality bar; the rest is consultant margin.

01

### RAG in production

Indexing, re-ranking, citations and answer evaluation over your real documents.

02

### Purposeful fine-tuning

LoRA adapters for style, format or narrow skills; no money burned on what RAG already solves.

03

### Honest evaluation

Test sets from your real cases and cost/quality comparison across routes.

04

### Privacy

Data residency, local inference or VPC when compliance requires it.

How we work

## How we decide and build

-

### Constraints

Data, languages, budget and compliance: the frame before the model.

-

### Baseline

We test the best generic model; that's the yardstick everything else must beat.

-

### Experiment

RAG and/or adapters over the baseline with comparable metrics and cost per run.

-

### Decision

The cheapest route that meets the bar, documented. If none exists, we say so.

Scope and deliverables

## What the project delivers

- **Constraint analysis:** Data, language, budget and compliance, in writing.

- **Baseline:** The best generic model evaluated on your cases.

- **Experiment:** RAG and adapters with comparable metrics and cost.

- **Recommendation:** The chosen route with estimated cost and risks.

- **Implementation:** A production pipeline if you decide to continue.

- **Documentation:** The criterion, documented and shareable.

Proof in production

## The criterion, applied

30, fixed prompts per evaluation

3, languages evaluated

2, routes compared: RAG & adapters

100%, decisions backed by evidence

Read more

Peoplesnap

### From smartphone to DOOH mosaic, without a single booth

Peoplesnap is Slash Experience's photobooth product: a white-label web app that turns each guest's phone into the activation. QR, selfie, AI editing (100+ parameters), social share, print and mosaics on event screens. Already live for Mouton Cadet, La French Tech Bordeaux, CEA Marcoule, Denver Pub Crawl, With Agency, We Love Beer and CES Las Vegas 2024.

Learn more →

Frequently asked questions

## The questions we hear often

**When is fine-tuning NOT the answer?**
When the problem is missing or changing knowledge: RAG solves that cheaper and safer.

**Do you work with open models?**
Yes, including local inference with GLM, Qwen, Mistral and others when compliance requires it.

**What does an experiment cost?**
It's scoped in sprints; discovery sets the size. Write to us with your case.

**Will results improve in Spanish?**
We explicitly evaluate your business Spanish; it's one of the lab's axes.

**When do you recommend hybrid?**
When style must be yours and knowledge must be current: adapters for tone, RAG for the data.

**What data do we need for fine-tuning?**
Real examples from your case, reviewed by an expert and anonymized when they contain personal data; your cases also become the test set. First we write down your data, language, budget and compliance constraints, because RAG often solves the problem with less cost and risk.

**Does our data have to leave the company?**
Not necessarily. If compliance requires keeping knowledge in-house, we use open models with local inference, in your VPC or on your own hardware. And client prompts are never published: only the method and findings go to /insights/.

**What costs come after training?**
Serving the model, usage-based or on a dedicated GPU, and maintaining it: when the base model is retired or you change vendor, a new training and evaluation cycle follows. That's why we compare the cost and quality of each route, RAG or adapters, before recommending one.