Question 130 A vendor invoice PDF contains selectable text, but a developer uses OCR and misreads decimal points in already-machine-readable totals. Why does the current approach fail under these conditions? Retrieval chunk size, top-k and prompt overhead jointly consume context; a large per-chunk setting can still overflow the model. Re-OCR can degrade accurate embedded glyphs; modality-aware extraction reduces avoidable numeric errors. Raster-only scans have pixels rather than a selectable text layer, so text extraction alone cannot recover their fields. Flattened table cells lose header associations, causing retrieved values to be assigned to the wrong compliance category. ExplanationCorrect Answer B:Extract from the native PDF text layer first and use targeted OCR only for raster portions, with typed total validation. Other Answers A:Re-OCR can degrade accurate embedded glyphs; modality-aware extraction reduces avoidable numeric errors. C:That mechanism describes another problem: A scanned equipment inspection sheet has no embedded text; a plain PDF text extractor returns empty strings for all checkboxes and labels. It does not explain this case, where re-OCR can degrade accurate embedded glyphs; modality-aware extraction reduces avoidable numeric errors. D:Re-OCR can degrade accurate embedded glyphs; modality-aware extraction reduces avoidable numeric errors.
Question 131 A vendor invoice PDF contains selectable text, but a developer uses OCR and misreads decimal points in already-machine-readable totals. Which observation would establish that the requirement is now met? Verify sampled equipment IDs and checkmarks against the original page images, tracking unreadable fields explicitly. Reconcile native amounts against invoice line totals and inspect a sample of any OCR-processed regions. Ask for thresholds from two different columns and verify the retrieved evidence names the correct row and header. Record actual prompt token totals and verify the final cited evidence is present without truncation across worst-case queries. ExplanationCorrect Answer B:Re-OCR can degrade accurate embedded glyphs; modality-aware extraction reduces avoidable numeric errors. Other Answers A:Reconcile native amounts against invoice line totals and inspect a sample of any OCR-processed regions. C:Reconcile native amounts against invoice line totals and inspect a sample of any OCR-processed regions. D:Reconcile native amounts against invoice line totals and inspect a sample of any OCR-processed regions.
Question 132 A vendor invoice PDF contains selectable text, but a developer uses OCR and misreads decimal points in already-machine-readable totals. What lasting control best prevents recurrence of the identified issue? Classify documents by native text versus image regions and preserve extraction-method provenance for each field. Keep table-aware extraction fixtures and verify structural fidelity on new DOCX template versions. Enforce token-aware packing, alert on truncation and review maximum per-source chunk settings. Detect text-layer absence at ingest and route those pages through approved image-aware extraction and review. ExplanationCorrect Answer A:Re-OCR can degrade accurate embedded glyphs; modality-aware extraction reduces avoidable numeric errors. Other Answers B:This control addresses a different exposure: A DOCX compliance policy has headings and tables; converting it to plain paragraphs drops column labels needed to understand thresholds. It would not prevent this failure, because re-OCR can degrade accurate embedded glyphs; modality-aware extraction reduces avoidable numeric errors. C:That re-OCR can degrade accurate embedded glyphs; modality-aware extraction reduces avoidable numeric errors. D:Re-OCR can degrade accurate embedded glyphs; modality-aware extraction reduces avoidable numeric errors.
Question 133 A DOCX compliance policy has headings and tables; converting it to plain paragraphs drops column labels needed to understand thresholds. What is the most targeted corrective step for this observed failure? Extract from the native PDF text layer first and use targeted OCR only for raster portions, with typed total validation. Parse document paragraphs and table structures with a format-aware library, preserving row/column context in the indexed representation. Serialize each logical table range with column names, row labels, units and associated footnotes before indexing. Run OCR or document understanding on the raster pages, then check confidence and associate extracted marks with page positions. ExplanationCorrect Answer B:The action targets the stated cause: Flattened table cells lose header associations, causing retrieved values to be assigned to the wrong compliance category. Check the repair by: ask for thresholds from two different columns and verify the retrieved evidence names the correct row and header. Other Answers A:Flattened table cells lose header associations, causing retrieved values to be assigned to the wrong compliance category. C:Flattened table cells lose header associations, causing retrieved values to be assigned to the wrong compliance category. D:Flattened table cells lose header associations, causing retrieved values to be assigned to the wrong compliance category.
Question 134 A DOCX compliance policy has headings and tables; converting it to plain paragraphs drops column labels needed to understand thresholds. What is the decisive mechanism behind the observed result? Table values without labels are semantically ambiguous and invite unsupported numeric explanations. Raster-only scans have pixels rather than a selectable text layer, so text extraction alone cannot recover their fields. Re-OCR can degrade accurate embedded glyphs; modality-aware extraction reduces avoidable numeric errors. Flattened table cells lose header associations, causing retrieved values to be assigned to the wrong compliance category. ExplanationCorrect Answer D:Parse document paragraphs and table structures with a format-aware library, preserving row/column context in the indexed representation. Other Answers A:That mechanism describes another problem: A quarterly report has tables whose cell meanings depend on row headings and footnotes; embedding each numeric cell independently loses context. It does not explain this case, where flattened table cells lose header associations, causing retrieved values to be assigned to the wrong compliance category. B:Flattened table cells lose header associations, causing retrieved values to be assigned to the wrong compliance category. C:Flattened table cells lose header associations, causing retrieved values to be assigned to the wrong compliance category.
Question 135 A DOCX compliance policy has headings and tables; converting it to plain paragraphs drops column labels needed to understand thresholds. Which verification is strongest for this particular failure mode? Verify sampled equipment IDs and checkmarks against the original page images, tracking unreadable fields explicitly. Ask for thresholds from two different columns and verify the retrieved evidence names the correct row and header. Query an exceptional metric and verify the returned table chunk includes its unit, row definition and footnote qualifier. Reconcile native amounts against invoice line totals and inspect a sample of any OCR-processed regions. ExplanationCorrect Answer B:Flattened table cells lose header associations, causing retrieved values to be assigned to the wrong compliance category. Other Answers A:Ask for thresholds from two different columns and verify the retrieved evidence names the correct row and header. C:Ask for thresholds from two different columns and verify the retrieved evidence names the correct row and header. D:Ask for thresholds from two different columns and verify the retrieved evidence names the correct row and header.
Question 136 A DOCX compliance policy has headings and tables; converting it to plain paragraphs drops column labels needed to understand thresholds. What change to engineering governance would reduce this risk over time? Make source/version keys mandatory and enforce deterministic chunk IDs in the Delta ingestion schema. Detect text-layer absence at ingest and route those pages through approved image-aware extraction and review. Keep table-aware extraction fixtures and verify structural fidelity on new DOCX template versions. Classify documents by native text versus image regions and preserve extraction-method provenance for each field. ExplanationCorrect Answer C:Flattened table cells lose header associations, causing retrieved values to be assigned to the wrong compliance category. Other Answers A:That flattened table cells lose header associations, causing retrieved values to be assigned to the wrong compliance category. B:Flattened table cells lose header associations, causing retrieved values to be assigned to the wrong compliance category. D:This control addresses a different exposure: A vendor invoice PDF contains selectable text, but a developer uses OCR and misreads decimal points in already-machine-readable totals. It would not prevent this failure, because flattened table cells lose header associations, causing retrieved values to be assigned to the wrong compliance category.
Question 137 An ingestion notebook embeds policy chunks directly but does not store their source document IDs, making outdated vectors impossible to revoke. Which engineering action most directly addresses this specific requirement? Write chunk_id, document_id, source_version, content and timestamps to a governed Unity Catalog Delta table before indexing. Parse document paragraphs and table structures with a format-aware library, preserving row/column context in the indexed representation. MERGE approved new chunk rows by stable IDs and version while removing only the superseded department's rows. Separate the latest substantive reply from quoted history and deduplicate repeated quoted blocks before embedding. ExplanationCorrect Answer A:The action targets the stated cause: Without stable provenance keys and versions, incremental refresh and deletion cannot identify which indexed chunks belong to a superseded source. Check the repair by: replace one policy document and confirm the Delta change updates only its chunk rows with valid lineage. Other Answers B:Without stable provenance keys and versions, incremental refresh and deletion cannot identify which indexed chunks belong to a superseded source. C:Without stable provenance keys and versions, incremental refresh and deletion cannot identify which indexed chunks belong to a superseded source. D:Without stable provenance keys and versions, incremental refresh and deletion cannot identify which indexed chunks belong to a superseded source.
Question 138 An ingestion notebook embeds policy chunks directly but does not store their source document IDs, making outdated vectors impossible to revoke. What architectural constraint is most directly responsible? Model-generated metadata is untrusted and a direct write can corrupt lineage or authorize documents incorrectly. Flattened table cells lose header associations, causing retrieved values to be assigned to the wrong compliance category. A full table overwrite discards unrelated tenants' chunks and can break downstream indexes. Without stable provenance keys and versions, incremental refresh and deletion cannot identify which indexed chunks belong to a superseded source. ExplanationCorrect Answer D:Write chunk_id, document_id, source_version, content and timestamps to a governed Unity Catalog Delta table before indexing. Other Answers A:Without stable provenance keys and versions, incremental refresh and deletion cannot identify which indexed chunks belong to a superseded source. B:Without stable provenance keys and versions, incremental refresh and deletion cannot identify which indexed chunks belong to a superseded source. C:That mechanism describes another problem: A pipeline overwrites an entire shared document-chunk table when only one department's handbook is refreshed. It does not explain this case, where without stable provenance keys and versions, incremental refresh and deletion cannot identify which indexed chunks belong to a superseded source.
Question 139 An ingestion notebook embeds policy chunks directly but does not store their source document IDs, making outdated vectors impossible to revoke. What validation evidence would most specifically confirm the resolution? Replace one policy document and confirm the Delta change updates only its chunk rows with valid lineage. Compare tenant row counts and hashes before and after one department refresh; unrelated tenants must remain unchanged. Ask for thresholds from two different columns and verify the retrieved evidence names the correct row and header. Inject a malformed ACL and duplicate chunk ID in a test batch; verify both are rejected before production commit. ExplanationCorrect Answer A:Without stable provenance keys and versions, incremental refresh and deletion cannot identify which indexed chunks belong to a superseded source. Other Answers B:Replace one policy document and confirm the Delta change updates only its chunk rows with valid lineage. C:Replace one policy document and confirm the Delta change updates only its chunk rows with valid lineage. D:Replace one policy document and confirm the Delta change updates only its chunk rows with valid lineage.
Question 140 An ingestion notebook embeds policy chunks directly but does not store their source document IDs, making outdated vectors impossible to revoke. What prevention measure is most targeted to this underlying cause? Make source/version keys mandatory and enforce deterministic chunk IDs in the Delta ingestion schema. Keep table-aware extraction fixtures and verify structural fidelity on new DOCX template versions. Apply schema constraints, scoped service-principal privileges and a staging-to-production promotion gate. Use scoped Delta merges with audit logs and tenant-isolation regression tests. ExplanationCorrect Answer A:Without stable provenance keys and versions, incremental refresh and deletion cannot identify which indexed chunks belong to a superseded source. Other Answers B:Without stable provenance keys and versions, incremental refresh and deletion cannot identify which indexed chunks belong to a superseded source. C:This control addresses a different exposure: A team lets a model write raw chunk text and arbitrary permissions directly into a production Delta table without type or access checks. It would not prevent this failure, because without stable provenance keys and versions, incremental refresh and deletion cannot identify which indexed chunks belong to a superseded source. D:That without stable provenance keys and versions, incremental refresh and deletion cannot identify which indexed chunks belong to a superseded source.
