Probe, Don't Prompt: A Hidden-State Probe for Metadata Filtering in Multi-Meta-RAG
The study replaces a costly GPT-3.5 prompt for extracting news sources with a cheap, local probe trained on a small open-source model's hidden states. This probe acts as a metadata filter by classifying query tokens into exactly 49 known sources. The method achieves high accuracy and avoids API costs and source drift, proving that shallow probing is an effective replacement for complex prompt engineering in retrieval systems.