I will build end to end local rag systems and chatbots

A Polish RAG chatbot, built end to end from scraping the articles to a finished answer with three links to the sources it used.

The knowledge base is small on purpose: 141 help articles. At that size I could measure every change on a test set instead of guessing.

Search is hybrid. Keyword matching handles questions where one exact word decides the answer, vector search handles meaning, the two rankings get merged, and a cross encoder picks the best three out of twenty candidates. Accuracy is 58 out of 60 on the test set.

Real questions arrive with typos, so a spell corrector sits in front of the pipeline, guarded by a word frequency check so it does not mangle correct input.

The system also knows when to refuse. Off topic questions get cut cheaply before generation, a second model decides the borderline cases, and a coverage check afterwards catches answers not grounded in the retrieved text.

Retrieval runs locally: embeddings, indexes and the reranker all sit on the server. The public demo calls a hosted model for generation because the VPS has no GPU, but it is the same code with a different config file, and it runs fully offline on Ollama.