> Bron: https://neuralex.nl/en/aanpak
> How Neuralex reduces hallucination and makes it measurable: hybrid retrieval, reranking and citation verification, with an evaluation harness that reports the hallucination rate as a hard number.

Approach & technology

# No answer without a source. And the source is verifiable.

A language model that sometimes makes things up is worthless for serious work. Our entire pipeline is built to prevent that — and to make it demonstrably true. This isn't a marketing promise; it's how the system is put together.

The RAG pipeline

## From question to verified answer

Six steps, one principle: the model may only speak about what it can substantiate.

1.  Step 1

    ### Question

    The user asks a question in natural language.

2.  Step 2

    ### Hybrid retrieval

    Vector (semantic) and keyword (FTS) search together surface the best candidate sources.

3.  Step 3

    ### Reranking

    A reranking model puts the truly relevant passages on top — noise drops away.

4.  Step 4

    ### Generation

    The language model formulates the answer, based solely on the retrieved sources.

5.  Step 5

    ### Citation verification

    Is every claim actually in the source? If not, it gets cut.

6.  Step 6

    ### Answer + source card

    The result comes with a clickable reference to the original.

**Continuous feedback.** An evaluation harness with a fixed set of checked questions (golden set) re-measures quality and the hallucination rate with every change. Regressions become visible before they hit production.

Principles

## Four choices that decide everything

### Citations by design

Sources aren't bolted on afterwards — they're built into the foundation. An answer without a verifiable source doesn't make the cut.

### Your data, your place

For each part of the pipeline we agree where it runs and why. You do not get a promise, you get an overview of which part sits where.

### Layered AI = low cost

RAG and vector search are nearly free; a local model handles the heavy lifting; an expensive API is only called in where it's truly needed.

### Measured, not hoped for

An evaluation harness with a golden set objectively measures quality and hallucination rate. We steer by the numbers and keep them below market leaders.

Under the hood

## Mature, proven technology

No black-box platform locking you in, but a transparent stack we keep in our own hands — and one you can run yourself if you want to.

pgvectorFTS / hybrid retrievalRerankingOllamaLlama · Qwen · GemmaCitation verificationFastAPIAstroCaddyCloudflaren8nschema.org · llms.txt

How we work

## From idea to production — no pilot graveyard

01

### Explore

We map out your process, data and goal. What can AI already take over, and where's the biggest payoff?

02

### Build

A working prototype on your own data — RAG, agent or model — with citations and evaluation built in.

03

### In production

Hosted where we agreed, monitored and maintained. No pilot that disappears into a drawer.

Deeper questions?

## Want to know if it works on your data?

The honest test is a trial setup on a slice of your own data. We show you what the retrieval surfaces, what the model makes of it, and what the source card looks like.

[Bekijk de showcases](/showcase) [See a live product](https://rechtspraak.neuralex.nl)

---
Volledige (opgemaakte) versie: https://neuralex.nl/en/aanpak