A Sovereign Retrieval-Augmented Generation Pipeline on a Vector-Enabled RDBMS: Local Synthetic Judge, Fail-Closed Verification, and Fine-Tuned Compact Filter Models

Author: Aleksandr A. Dmitriev (ORCID 0009-0000-6164-8474), ABEL BLEO AI LAB (independent research laboratory). Developed by the author outside of employment duties. CC BY-NC-ND 4.0. We describe a fully local (sovereign) retrieval-augmented generation (RAG) pipeline for corporate and scientific document collections under strict data-control requirements: PostgreSQL/pgvector storage (1024-d embeddings, HNSW), local embedding and generative models, and three reliability mechanisms — (1) a provenance-tracked local synthetic judge for regression evaluation of retrieval quality; (2) fail-closed answer verification; (3) specialized compact filter models (under 1B parameters) fine-tuned with LoRA on the system's own operation logs. A fine-tuned 0.6B noise filter achieves recall 1.00 on true signals with 79% false-positive suppression on a 132-item holdout, outperforming the non-tuned baseline on both metrics and reducing the operator-facing stream 3.6-fold. Three negative results essential for reproduction are reported: cascading filters is inadmissible under asymmetric error costs; inference input format must match the training format; disabling chain-of-thought degrades compact models. The full training-to-deployment cycle runs on commodity hardware (Apple M1 Pro) with no external compute. Preprint of article 3 in the ABEL BLEO AI LAB publication series. Extends the NIDUS Architecture Specification v1.0 (10.5281/zenodo.21041851) and Methodology Report v1.1 (10.5281/zenodo.21374543).