Text embeddings connect large language models (LLMs) to external knowledge in retrieval and retrieval-augmented generation (RAG), but most embedding methods, benchmarks, and tools were designed for short inputs. Long documents violate this assumption because relevant evidence may be local, distributed across sections, query-dependent, multilingual, or recoverable only through relations among distant passages. As a result, truncation loses evidence and fixed-size chunking often destroys the structure that makes it useful. This survey treats long-document representation as a distinct problem and examines how documents can be encoded as reusable, indexable structures for retrieval and RAG. We develop a representation-first taxonomy that separates embeddings, representations, and retrieval systems, covering single-vector, chunk-based, contextual, hierarchical, multi-vector, learned-sparse, hybrid, long-context, and summary- or graph-based approaches. We also synthesize the role of supervision, synthetic data, and distillation, showing that distillation is central to modern short-text embedding pipelines but remains underdeveloped for long documents. Across benchmarks, domains, and deployment settings, we find that no representation family dominates across quality, faithfulness, latency, cost, and index size. The survey therefore provides a practical map of long-document retrieval methods and argues that scalable RAG requires matching the representation to the evidence pattern, rather than relying on larger context windows, leaderboard scores, or chunk-and-pool pipelines alone.
Roberto Martínez-Cruz (Wed,) studied this question.