Local AI has a freshness problem. Open-weight models you run on your own machine are private and free of per-token fees, but their knowledge is frozen at whatever moment their training data was collected. Ask one about this week's news and you get confident nonsense, because the model has no way to know that its answer is stale. heise online's latest heise+ how-to tackles the obvious fix: giving the model a way to look things up.
The recipe is retrieval-augmented generation, or RAG, applied to live web results rather than a fixed document set. In practice it means wiring a local runtime such as Ollama or LM Studio to a search tool, letting the model issue queries, then feeding the retrieved pages back into the context window alongside the user's question — ideally with an instruction to cite what it used and to say when sources disagree. More advanced setups let the model run several searches in a loop, refine its query, and only then answer, the same pattern cloud assistants have used since they gained web access.
The trade-off is the interesting part. The moment your question travels to a commercial search API, the privacy argument for running a model locally weakens considerably. That is why the do-it-yourself route usually pairs a local model with self-hosted search infrastructure — a metasearch engine such as SearXNG querying public engines without a personal profile, or a private index — and why the resulting answers are only as good as the pages the search step returns. Retrieval fixes staleness; it does not fix a bad source.
Practically, the setup is now within reach of a single desktop or a spare GPU server: an open-weight model with reliable tool-calling, a small embedding model for ranking, and a search backend. The lesson goes beyond the tutorial: treat citations as the acceptance test. A grounded local assistant that names its sources and admits ignorance is more useful than one that sounds certain about a world it has never seen.
