A cluster of releases is separating the document-agent stack into parsing, multimodal representation, and evaluation.
LlamaIndex's ExtractBench makes extraction quality a distinct test surface, while WeMM-Embedding and related releases focus on how mixed document content is represented for retrieval and downstream use. Read together, they make one-model document processing look increasingly incomplete.
The practical architecture is routed: classify the document, choose an extraction path, validate a schema, embed what must remain searchable, and evaluate failures on representative files. The roundup remains an implementation radar, not proof that one stack is best; each component still needs task-specific testing and provenance.