Nederland.ai — Advice Tools

Nederland.ai needed conversational advice tools embedded directly into their public website — not a generic chatbot widget, but domain-specific tools that answer questions grounded in their own content. The constraints were tight: cheap to run, fast first-response, and able to cite the source documents behind every answer instead of hallucinating.

The team had tried wiring up ChatGPT's API directly to a static system prompt and quickly hit the predictable failure modes — hallucinations on questions outside the training data, no ability to cite sources, no domain grounding. They needed a retrieval-augmented setup but didn't want to manage a full vector database.

We built a Python service on Flask that sits between the website and OpenAI. The website hits our endpoint; we resolve the question against a FAISS index built from Nederland.ai's source documents; we hand the relevant chunks plus the user's question to the ChatGPT API as a single retrieval-augmented prompt; we return the answer with source citations.

FAISS keeps the vector search in-memory and fast — no Postgres extension, no managed vector DB, no per-query fees. The index rebuilds on a schedule when source documents change. The Flask layer handles rate-limiting, prompt construction, caching of common questions, and serves a simple JSON API the website embeds via JavaScript.

Deployed as a self-contained Python process on Nederland.ai's existing infra. Total cost is dominated by ChatGPT API calls — the FAISS layer adds essentially nothing.

Multiple advice tools now embedded across the Nederland.ai website, each grounded in its own document corpus. Answers cite their sources. First-response time under 2 seconds. The team can add new advice tools by dropping a new document set into the indexer — no engineering work per tool.

Key design choice that paid off: keeping FAISS in-process instead of running a separate vector database. Lower latency, lower cost, simpler deployment, fewer moving parts to break.