Retrieval-Augmented Generation (RAG) systems for enterprise technical domains require embedding models that can precisely surface exact-identifier content — IP addresses, BGP autonomous system numbers, VLAN identifiers, and infrastructure component names — from dense technical corpora. Most published RAG benchmarks evaluate on general purpose datasets and assume GPU-accelerated inference. We present a controlled three-phase empirical study comparing three embedding architectures — nomic-embed-text (dense, 137M parameters), Jina Embeddings v3 (dense, 570M parameters), and BGE-M3 (hybrid dense+sparse, 570M parameters) — on a real-world GPU-as-a-Service infrastructure corpus comprising 451 chunks across 19 content types, evaluated against 30 cross-domain questions using six openweight language models on fully offline, CPU-only commodity hardware. Our results demonstrate that: (1) embedding model size does not reliably predict domain-specific retrieval quality — nomic-embed-text outperforms both larger models by an average of 6–10 percentage points; (2) hybrid sparse+dense retrieval resolves specific exact-identifier retrieval failures but cannot overcome corpus ambiguity where two chunks are semantically and lexically equivalent; (3)llama3.1:8b, an 8B-parameter LLM, exhibits embedding invariant reasoning performance across all three embedding conditions, suggesting that larger LLMs may exhibit similar resilience to retrieval quality variation; and (4) the pgvector ivfflatapproximate index causes silent retrieval degrada tion on corpora below approximately 1,000 chunks, based on observations at 451 rows. All experiments were conductedfully offline on a 4-core, 16GB RAM mini PC with no GPU.
Eswar Raman (Sun,) studied this question.