Low-Resource SLM Models for Elastic Semantic Search
Semantic search over domain-specific corpora requires an effective embedding model and infrastructure. Elasticsearch’s native semantic_text field and ELSER sparse-vector inference require a commercial Enterprise subscription, inaccessible to most academic institutions. This paper documents the licensing barrier and presents a reproducible manual pipeline achieving comparable semantic search with free Elasticsearch components and open-source small language models (SLMs), on a four-node Raspberry Pi 4 cluster (8 GB RAM, three-node Elasticsearch) over 1088 USGS documents. Five models were evaluated —ELSER v2, .multilingual-e5-small, all-MiniLM-L12-v2, all-mpnet-base-v2, and msmarco-MiniLM-L12-cos-v5—from 35 screened candidates, plus a BM25 lexical baseline. Available process memory, not compute, is the binding constraint: Elasticsearch’s footprint consumes 5–6 GB of the 8 GB. The three 384-dimensional models reindexed the corpus in 2.1–2.2 h; the 768-dimensional all-mpnet-base-v2 took 10.1 h. On-disk size is unreliable for provisioning: .multilingual-e5-small expands from 1.4 GB on disk to 3.5 GB at runtime (2.5×). Retrieval quality was assessed with Precision@10, MRR, MAP@10, and nDCG@10 over 15 human-judged queries rather than raw similarity scores; embedding-based retrieval outperforms BM25, reaching significance for two of five configurations. Msmarco-MiniLM-L12-cos-v5 offers the strongest quality-per-resource trade-off among the 384-dimensional models for 8 GB ARM deployments. Pipeline and configuration artefacts are documented and reproducible.