KoboldCpp on Windows: SillyTavern and web search setup

This tutorial covers KoboldCpp setup on Windows, then connects the local LLM backend to SillyTavern and adds web search. The presenter downloads the executable from the project's GitHub releases and selects a GGUF model from Hugging Face. The model file can live separately from the executable. Context size is adjustable; the presenter temporarily uses 2048 while recording but advises against treating that as a general recommendation.

In SillyTavern, the walkthrough selects Text Completion and the KoboldCpp API type, then connects using the default port, 5001. The presenter recommends deriving context size from the backend after a request fails because SillyTavern's context setting is too large. KoboldCpp's own interface also supports scenarios and character imports from PNG files.

Web search requires both the SillyTavern extension and KoboldCpp's Network option. Installing the extension alone returns empty results in the demonstration. After restarting the backend with search enabled, the presenter checks direct /websearch queries in terminal output and uses backticks to trigger searches within chat. Higher visit counts require more context for the model to use the retrieved material.

The presenter reports repeated crashes during recording, but says ordinary use worked without those problems. They also report needing the insecure "no certify" mode on Mac, while generally not needing it on Windows. These are observations from their setup, not guarantees about other systems.