> Bron: https://neuralex.nl/en/showcase/rechtspraak > How Neuralex made 3.6 million Dutch court rulings searchable: FTS5 retrieval in milliseconds, an LLM chain with fallback, and answers that back up every claim with a source. Showcase · Neuralex Legal # 3.6 million rulings, searchable with proof The lab's flagship: a LegalTech platform that indexes the complete Dutch case law locally and puts an AI layer on top that **never answers without a source**. Below you'll see exactly how the technology fits together — the result you can see live. [Open the live portal](https://rechtspraak.neuralex.nl) [What the product does](/en/lab#legal) 3.6M court rulings 2.25M with full text 70–300ms search response (FTS5) 16 law books linked Architecture ## From question to substantiated answer Everything runs on our own hardware in the Netherlands. The edge is Cloudflare, the rest is self-built and self-hosted — including the entire search index. The RAG chain ## Five steps, every claim traced to its source The core of every Neuralex product: retrieval first, language model second, and an answer that only exists if the source carries it. 01 ### Question in plain language The user describes a situation or asks a legal question — no jargon or ECLI numbers required. 02 ### Retrieval over 19 GB An FTS5 full-text index searches 3.6 million rulings and 16 law books in 70–300 ms. 03 ### Context with sources The most relevant rulings and statutory articles go to the language model as context, complete with ECLI references. 04 ### LLM chain with fallback A model-agnostic chain: any API can be plugged in — EU-hosted models or fully local — with cost and latency measured per call. 05 ### Answer with proof Every claim in the answer gets a clickable source card: ECLI plus exact quote. No source, no claim. Tech in detail ## What's under the hood ### FTS5 on 19 GB of SQLite The complete Dutch case law in one local full-text index. No external search service, no data that leaves the building — yet responses in milliseconds. ### Provider chain with cost tracking Every AI call runs through an ordered chain of providers. The layer is model-agnostic: any API works, from models hosted exclusively on European servers to a model running fully on-prem. If a provider fails, the next one takes over — with model, tokens and cost logged per call. ### Citation verification The answer is only finished once every claim can be traced back to a real ruling or statutory article. Source cards show ECLI and the exact quote — verifiable in one click. ### Self-healing deployment The app runs under launchd with KeepAlive: if a process crashes or the machine restarts, the portal is back up within seconds. A health endpoint feeds external monitoring. ### Daily data pipeline New rulings from rechtspraak.nl are automatically imported, indexed and linked to statutory articles every night. The knowledge base publishes daily from a queue. ### On-prem & GDPR — and what the demo does The 3.6M rulings sit on our own Dutch hardware and search happens entirely locally: the index never touches an external service. The language model is a separate choice. In the public demo the answering step currently runs through an external API, because that is cheap testing on public case law. For a production deployment handling client data that step moves to EU-hosted or fully on-prem — which is exactly why the layer is built model-agnostic. Built with PythonFlaskGunicornSQLite FTS5pgvectorModel-agnostic LLM layerCloudflare Tunnellaunchdn8n Why this works ## AI you can build on Most AI fails on trust: it makes things up, it leaks data, or it doesn't fit the process. This portal flips those three around — and that's the blueprint for every Neuralex product. ### Citations, verifiable Every answer points to the exact ruling. The pipeline checks that a cited ECLI actually exists and supports the passage; where a claim fails that check it is flagged, not papered over. ### Data stays in the Netherlands On-prem or in a GDPR-proof environment. Sensitive data doesn’t leave the building and doesn’t go to an American cloud. ### Custom-built, not a wrapper No thin layer around a chatbot. Built on the processes, documents and hardware of the domain itself. ### Measurable results An evaluation harness objectively measures quality and hallucination rate. We steer by numbers, not by feel. See it yourself ## The best demo is the portal itself Ask a real legal question and see how every claim is backed by ECLI sources. This is the same RAG blueprint we can build for any domain. [Open rechtspraak.neuralex.nl](https://rechtspraak.neuralex.nl) [RAG for your domain?](/en/contact) --- Volledige (opgemaakte) versie: https://neuralex.nl/en/showcase/rechtspraak