Embedding Generation — References

What a vector depends on besides its text, and the evidence behind the encoding conventions a build pipeline must apply identically on both sides.

Encoding conventions that must match

• Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks (Reimers & Gurevych, 2019), arXiv:1908.10084 — pooling over token representations and the siamese training that makes the resulting vectors comparable, which is why pooling is a property of the index rather than of the call site.
• Dense Passage Retrieval for Open-Domain Question Answering (Karpukhin et al., 2020), arXiv:2004.04906 — the two-tower construction with separate query and passage encoders, the origin of the asymmetry that instruction prefixes now express.
• Matryoshka Representation Learning (Kusupati et al., 2022), arXiv:2205.13147 — representations whose leading dimensions can be truncated, and therefore a dimension that must be chosen identically for stored vectors and for queries.

Measuring whether a re-encode is worth it

• MTEB: Massive Text Embedding Benchmark (Muennighoff et al., 2022), arXiv:2210.07316 — the task taxonomy and per-task tables, a shortlisting instrument rather than a migration justification.
• BEIR: A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models (Thakur et al., 2021), arXiv:2104.08663 — evidence that transfer between domains is uneven, which is the reason a general improvement does not imply an improvement on your corpus.

Implementations

• Sentence-Transformers documentation — the authority on pooling strategies, normalisation and prompt handling, and the place most reproducibility mismatches between a build job and a query path originate.
• Text Embeddings Inference — a serving component for open-weight encoders, including batching behaviour relevant to build throughput.

No verified reference is recorded for configuration-identity hashing, for parallel-collection migration procedures, or for source-to-vector reconciliation. These are operational practice, and the published literature addresses the models rather than the pipelines that apply them.