Gemma 4 Powers Local Swiss Teletext RAG

This is a strong niche RAG experiment because teletext-style articles are compact, time-stamped, and information-dense, which exposes retrieval quality fast.

• –The local-only architecture is the main selling point: no data leaves the machine, which is useful for sensitive corpora.
• –Hybrid retrieval plus cross-encoder reranking is the right shape for terse news queries, where pure vector search often misses short exact matches.
• –The corpus size is meaningful for a personal/local setup, and the German-language coverage makes multilingual retrieval quality a real issue instead of a toy demo.
• –The repo reads more like a well-documented PoC/tutorial than a polished product, so system prompt tuning and model selection are still the obvious areas to improve.