Web search

Web search brings current information from the internet into the chat: prices, news, the state of the law, details about other companies. This page explains how it is used, which building blocks it consists of and which of them you choose yourself during setup.

The search framework comes from LibreChat. What we add is the preconfiguration.

How it is used in the chat

There is a switch for web search in the input field. If it is on and the search is fully set up, the colleague decides for itself when to search. You do not have to write “search the internet”, a question about the current state of something is enough.

The answer is then built on the pages found, with source references. The stored system prompt demands that explicitly.

The three building blocks

A working web search consists of more than a search engine.

Building block	Task	Required
Search engine	Delivers hits, meaning links and short text excerpts	yes
Scraper	Fetches the full text of the pages found	yes
Reranker	Sorts the findings by how well they fit	no

Without a scraper the colleague would only have headlines and two liners, and no reliable answer can be built from those.

Web search is completed during setup

A search engine is already configured on delivery. You choose the scraper yourself, because which provider fits depends on how you handle data residency; a reranker is optional and not set out of the box.

Plan this step in firmly: the search tool is only offered to the model once a scraper is configured.

For admins: configuring scraper and reranker

Both are maintained in the admin panel under “Konfiguration → Funktionen” (configuration, features tab), in the two cards for web search. There you choose the search provider, the scraper and optionally a reranker, plus the SafeSearch level. LibreChat supports several providers for scrapers and rerankers; which ones qualify is listed in the LibreChat documentation.

Keys never belong in the configuration in plain text. Enter them as a reference to a variable and store the value in the place provided for it.

Which scraper fits the EU positioning

The intended European route is a self hosted scraper. Then fetching full text never leaves your infrastructure.

The hosted variants of the scraper providers are set up quickly but do not process EU resident. Anyone using them should name that just as clearly as the non EU model cards.

So that the colleague searches by itself

Two more adjustment points decide whether a search happens:

1. Role permission. In the role editor under “Werkzeuge” (tools), the permission “Web-Suche” (web search) for the roles concerned.

2. Card flag. The search tool has to be active on the model cards you offer. Only then is it attached to the card and the model decides for itself when to use it.

3. Afterwards, check in a fresh chat.

How much text reaches the model

How much text from the search results goes to the model is limited and adjustable. More means more context for the answer and more tokens per request, so higher cost. Less means a thinner basis.

Data protection in web search

The search runs in your instance, but the search query itself goes out: it is sent to external search services. What happens in the process and what does not:

Goes outside	the search term, plus the pages fetched by the scraper
Does not go outside	who searched. The identity of the person stays with you
Stays with you	the conversation, the documents, the answer

Wording for training and privacy notices

Put precisely, the statement reads: storage and inference run where you operate the instance; the search query itself goes to external search services, and it does so without user identity. Anyone describing web search in a privacy notice is best served by taking that distinction over verbatim.