Problem: Blindly chunking every PDF leads to bad retrieval. Solution: A small LLM pass (Traffic Controller) that routes documents to semantic collections (Finance vs Tech) and selects chunking strategies (Table-Aware vs Standard) before ingestion.
It's a simple pattern using Pydantic to force structured decisions from local LLMs (Ollama).
Repo: https://github.com/2dogsandanerd/smart-router-kit Part1: https://github.com/2dogsandanerd/smart-ingest-kit