BALANGAY-RAG: Adaptive Pre-Generation Retrieval and Abstention for Filipino-English Code-Switched Queries

Abstract

Fixed retrieval depths can waste context or proceed without designated support when language form, evidence availability, and context budgets vary. This controlled pre-generation study evaluates BALANGAY-RAG, combining the BALANGAY-PH benchmark with BALANGAY-RF, a Random Forest controller that selects k = 1, 2, 4, or 8 passages or abstains. The reported experiment contains 44 public-information passages, 147 English, Filipino, and Taglish queries, and 1,764 grouped held-out states. BALANGAY-RF achieved 0.80 feasible-evidence retrieval recall, 0.80 gold-passage precision among proceeded states, a 0.11 gold-absent proceed rate, and 15.49 unconditional mean retrieved words. A separately archived post-review reconstruction and reproduction regenerates 1,764 states and 10,584 policy outcomes from 39 official-source URLs, reproducing the qualitative context-coverage trade-off while yielding different numerical operating points. No downstream generator was evaluated.