Common mistakes

Treating parsing as solved because the pipeline does not error. Chunking by token count and calling it semantic. Embedding chunks with no context and wondering why retrieval finds the wrong section of the right document. Letting each team build their own pipeline because "their documents are different" (the formats are different; the pipeline pattern is not). Running the largest available model on every page of every document because it was easiest to set up. Storing nothing but chunks and embeddings, so that when the chunking strategy changes, everything must be re-ingested from source systems that may no longer have the original.

Every one of these is fixable, and the fix is cheaper earlier. A pipeline built around a canonical document model, with deduplication and versioning at the door, structure-first hierarchical chunking, one-time extraction stored as annotations, and data-engineering discipline throughout, is not dramatically more work than the naive version. It is just work that has to be decided on before the first document goes in.