AI Engineer
Contents
  1. Source documents and the extraction contract
  2. Text recovery and transcription fidelity
  3. Document structure and reading order
  4. Evidence spans and coordinate mappings
  5. Table structure and contextual meaning
  6. Evidence-bearing fields and records
  7. Extraction routes and their evidence
  8. Correctness across recognition, structure and documents
  9. Uncertainty and accountable review
  10. Corrections and dependent outputs
  11. Context-preserving downstream handoff
  12. Check understanding
  13. Open questions
  14. Selected talks
  15. References
  16. Talk library
← All topics

Document Understanding and OCR

Document understanding turns files into information that other systems can interpret and verify. Recognizing characters is only one step: a correctly read amount can still belong to the wrong invoice row. Reliable extraction preserves source identity, structural relationships and evidence while making incomplete or uncertain results explicit.

Source documents and the extraction contract

Document extraction produces three distinct results: recognized characters, reconstructed relationships and interpreted fields. An invoice needs more than its words: parties, line items and totals must remain distinguishable. The extraction contract specifies what must survive, including the evidence supporting each value.

A born-digital document originates electronically; scans and photographs encode page appearance as images. One PDF can contain both. Optical character recognition, or OCR, recognizes text in images. A PDF text layer supplies extractable characters, sometimes invisibly over a page image. Its presence does not establish complete recognition or preservation of the original typography.

File coverage and submission completeness

Example

Processing available pages does not supply a missing document.

Application inspection policy: inaccessible, unsupported or failed inputs remain unavailable. Successfully processed pages still require reconciliation against expected submission membership.
Read the diagram as text
  • Inspect submission.
  • Process available pages.
  • Unavailable or partial.
  • Reconcile expected documents.
  • Submission complete.
  • Receipt missing.
  • Inspect submission → Process available pages: Readable input.
  • Inspect submission → Unavailable or partial: Inspection fails.
  • Process available pages → Unavailable or partial: Page processing fails.
  • Process available pages → Reconcile expected documents: All available pages processed.
  • Reconcile expected documents → Submission complete: Expected membership satisfied.
  • Reconcile expected documents → Receipt missing: Required receipt absent.

A two-page invoice supplies a running example. Its application contract separates three responsibilities.

  • Required information — Extract invoice INV-1042, its parties, date, currency, line items and stated total. Preserve unresolved fields. These requirements express fitness for purpose.
  • Source identity — Retain original revision r1, distinguish its two pages from rendered images, and record whether opening required repair.
  • Production history — Provenance records origins and production history. Bind derived locations to their source revision; retain extraction-attempt identity. Collection, identity and permitted use explains the broader responsibility.

A manifest records expected membership. For this submission, it requires the invoice and a delivery receipt. Enumerating both invoice pages establishes coverage of that file; it cannot establish that the missing receipt arrived. Document identity and page accounting must therefore remain separate from submission completeness.

Text recovery and transcription fidelity

Native extraction reads encoded text. Image recognition locates text regions and identifies their characters; implementations can combine these operations. A modular converter can reuse native text and coordinates while analyzing rendered images for layout and tables, avoiding unnecessary retranscription.

Deskewing rotates crooked text toward horizontal lines. Noise, uneven backgrounds, stroke thickness, skew and unsuitable borders can disrupt recognition. Segmentation settings distinguish a page from a line or isolated word. Language and dictionary configuration also matter: sentence-oriented corrections can damage identifiers. These are documented Tesseract considerations, not universal resolution thresholds.

Text recovery, PDF construction and reader behavior require different inspections.
Observed problemMechanism and inspection
Duplicated textOCRmyPDF 17.11 fixed removal that missed Form XObjects—embedded page-content containers—and consequently stacked another OCR layer.
Text damaged during archival conversionVersion 17.8 guards against substitution of non-embedded CID fonts, which use character identifiers, corrupting searchability during PDF/A conversion. PDF/A is an archival PDF format.
Incorrect search results after rewritingVersion 17.6 removes an embedded search index made stale by rewriting. Stale search metadata is different from an obsolete transcription.
Searchable PDF appears to contain spacesTesseract's FAQ describes reader-specific interpretation of hidden text. Checking another reader helps distinguish this from recognition failure.
Poor handwriting recognitionThe Tesseract 4-era FAQ attributes this limitation to an engine designed for printed text; it does not characterize every document model.

Preserve literal recognition before repair. If the invoice identifier is recognized as INV-1O42, replacing the letter O with zero creates an interpretation requiring evidence. A plausible identifier is not necessarily the printed identifier. Image-based recognition methods belong in Vision AI.

  • Reprocessing changes preservation — Skipping text-bearing pages can leave image text unrecognized. Redo replaces recognized text; force rasterizes content. Text replacement can invalidate structural tags, and force can flatten interactive content.
  • Inspect intermediate artifacts — Compare the rendered page, preprocessed OCR input and recognized text. On mixed pages, the OCR artifact need not contain all page text.

Document structure and reading order

An intermediate representation connects content records with structural relationships before export. DoclingDocument, for example, separates content items from body and furniture trees; furniture includes headers and footers. Parent-child links express containment, while ordered children express sequence. Representation-local pointers do not promise stable identity after regeneration.

Reading order is the meaningful sequence of content. Reading across two columns can join correctly recognized fragments incorrectly.

Same regions, different reading order

Example

The joins change while recognized text stays fixed.

Column-wise

A → B → C → D

Scroll sideways if the figure extends beyond the screen.

02550751000255075100Horizontal position (pixels)Vertical position (pixels)ABCDFirst joinSecond joinThird joinA: Bill toB: Elm LabC: Ship toD: West Depot
  • 1. A
  • 2. B
  • 3. C
  • 4. D
  • 5. First join
  • 6. Second join
  • 7. Third join
Read coordinates and regions as data

X: 0–100 pixels; Y: 0–100 pixels, increasing down. Equal scale on both axes.

A (polygon)

(8, 20); (45, 20); (45, 35); (8, 35)

B (polygon)

(8, 55); (45, 55); (45, 70); (8, 70)

C (polygon)

(55, 20); (92, 20); (92, 35); (55, 35)

D (polygon)

(55, 55); (92, 55); (92, 70); (55, 70)

First join (polyline)

(26.5, 36); (26.5, 54)

Second join (polyline)

(46, 62.5); (54, 27.5)

Third join (polyline)

(73.5, 36); (73.5, 54)

A: Bill to: (26.5, 28)

B: Elm Lab: (26.5, 63)

C: Ship to: (73.5, 28)

D: West Depot: (73.5, 63)

Row-wise

A → C → B → D

Scroll sideways if the figure extends beyond the screen.

02550751000255075100Horizontal position (pixels)Vertical position (pixels)ABCDFirst joinSecond joinThird joinA: Bill toB: Elm LabC: Ship toD: West Depot
  • 1. A
  • 2. B
  • 3. C
  • 4. D
  • 5. First join
  • 6. Second join
  • 7. Third join
Read coordinates and regions as data

X: 0–100 pixels; Y: 0–100 pixels, increasing down. Equal scale on both axes.

A (polygon)

(8, 20); (45, 20); (45, 35); (8, 35)

B (polygon)

(8, 55); (45, 55); (45, 70); (8, 70)

C (polygon)

(55, 20); (92, 20); (92, 35); (55, 35)

D (polygon)

(55, 55); (92, 55); (92, 70); (55, 70)

First join (polyline)

(46, 27.5); (54, 27.5)

Second join (polyline)

(54, 35); (46, 55)

Third join (polyline)

(46, 62.5); (54, 62.5)

A: Bill to: (26.5, 28)

B: Elm Lab: (26.5, 63)

C: Ship to: (73.5, 28)

D: West Depot: (73.5, 63)

Column-wise order preserves each address block. Row-wise order interleaves labels and values. Arrows encode sequence, not containment.
RepresentationWhat it contributes
Plain text or ordinary MarkdownReadable exports need not retain geometry or every structural relationship; keep the richer representation when those connections matter.
hOCRAn HTML-based OCR format. A bounding box is a rectangle locating content; hOCR uses pixel coordinates from the image's top-left corner.
ALTO XMLA layout-and-text schema with ordering and processing references.

Containment, sequence and association answer different questions. A picture belongs to a page, follows a paragraph and has a caption. Retaining the picture, caption and embedded text separately preserves those distinctions; a generated image description is another artifact.

Layout analysis assigns element roles and regions. Uncertain footnote links or cross-page continuations should remain inspectable candidate relationships, rather than disappearing into an irreversible text join.

Evidence spans and coordinate mappings

An evidence span selects source text. Text-position selectors use inclusive starts and exclusive ends. Store the text version and counting convention with those offsets: changing the representation can invalidate them. A polygon is a multi-sided spatial region useful when a rectangle fits poorly.

Coordinates require an origin, units and a reference image or page. PDF page coordinates use points; rendered images use pixels. Conversion depends on the actual page transform, including cropping and rotation. Normalized coordinates express positions relative to dimensions, but still need the associated frame. The example below uses image pixels throughout.

Locate evidence after crop and resize

Example

Local coordinates change; the supported source region remains identifiable.

Original render

Locate orange evidence inside the blue crop.

Scroll sideways if the figure extends beyond the screen.

01002003004000100200300400x (pixels)y (pixels)CropEvidence(100,200)
  • 1. Crop
  • 2. Evidence
Read coordinates and regions as data

X: 0–400 pixels; Y: 0–400 pixels, increasing down. Equal scale on both axes.

Crop (polygon)

(100, 200); (200, 200); (200, 280); (100, 280)

Evidence (polygon)

(120, 220); (160, 220); (160, 240); (120, 240)

(100,200): (100, 190)

Processed crop

Compare transformed regions with the original.

Scroll sideways if the figure extends beyond the screen.

01002003004000100200300400x (pixels)y (pixels)CropEvidence
  • 1. Crop
  • 2. Evidence
Read coordinates and regions as data

X: 0–400 pixels; Y: 0–400 pixels, increasing down. Equal scale on both axes.

Crop (polygon)

(0, 0); (200, 0); (200, 160); (0, 160)

Evidence (polygon)

(40, 40); (120, 40); (120, 80); (40, 80)

Crop origin (100,200), scale 2. Inverse mapping recovers source positions, not discarded image content. Interpolation changes samples; nonlinear dewarping requires its actual mapping.
p′=s(p−c),p=c+p′/sp'=s(p-c),\qquad p=c+p'/s Here pp is an original pixel position, c=(100,200)c=(100,200) the crop origin, s=2s=2 the resize factor and p′p' the processed position. The inverse locates processed evidence in the original render.

Text editing needs a separate map. In USD 7.50, the amount occupies character span [5,9); after collapsing the doubled space, it occupies [4,8). Retain the source and destination spans. ICU's edit maps support this correspondence, but arbitrary replacements may provide only span-level alignment rather than exact internal character matches.

One field can need several text segments or page polygons. Its literal mention and normalized value remain separate. Located evidence establishes where to inspect, not whether the interpretation is correct.

Table structure and contextual meaning

Table structure organizes cells into rows and columns. Detection locates the table; structure recognition recovers cells, including cells spanning several rows or columns. Assigning text and interpreting header roles are further tasks. Accurate characters alone cannot establish correct records.

The invoice fixture has these source line items. A currency header, USD, qualifies both monetary columns.
PageItemQuantityUnit priceLine amount
1Cable2 each12.5025.00
2Adapter1 each7.507.50

A price needs several relationships

Example

Cell text alone does not identify its meaning.

The page-2 price belongs to Adapter, means unit price and inherits USD. Both page fragments belong to logical table T1 under the fixture's continuation assumption.
Read the diagram as text
  • 7.50.
  • Adapter.
  • Unit price.
  • USD.
  • Table T1.
  • Page-1 fragment.
  • Page-2 fragment.
  • 7.50 → Adapter: Row identity.
  • 7.50 → Unit price: Column meaning.
  • Unit price → USD: Currency qualifier.
  • 7.50 → Page-2 fragment: Located in.
  • Page-1 fragment → Table T1: Part of.
  • Page-2 fragment → Table T1: Continues.

A cell may have several qualifying headers. Explicit associations retain row identity, column meaning and currency together. Attaching the correctly read 7.50 to Cable instead of Adapter produces a wrong record without changing a character.

For this fixture, page 2 continues the same table. Cross-page representations can group such rows, but automatic continuation still needs verification. Repeated headers, changed units and footnotes must be reconciled before joining unfamiliar fragments; documented multipage support is not evidence of reliable stitching.

Annotation rules must preserve different cell meanings.

  • Blank versus zero — An empty cell is not the printed numeric value zero.
  • Ditto versus unreadable — A repetition mark needs an identified antecedent; unreadable content remains unresolved.

Evidence-bearing fields and records

Key-value extraction associates a label with its value; repeated line items group related fields into records. These associations differ from containment. Textract illustrates the distinction with CHILD relationships for contained words and VALUE relationships between keys and values.

The Adapter price retains separate observations and interpretations.
ComponentValue
Literal mention7.50
Normalized interpretationUSD 7.50 per item; decimal point convention; currency supplied by the header
EvidenceRevision r1, page 2, price cell and qualifying header
Use explicit application states after inspection; processing failure is a separate outcome.
StateMeaning
AbsentExpected field is not present in inspected content.
IllegibleContent exists but cannot be read.
AmbiguousSeveral interpretations remain, as with an unqualified 03/04/2026 date.
ConflictingSupported values disagree.
ExtractedA supported interpretation is recorded.

Selection marks, such as checkboxes, encode visual state. Preserve selected or unselected status alongside the associated question and response option. No detection is not an unchecked mark: requested feature filtering can omit marks entirely.

Tderived=2(12.50)+1(7.50)+0.00=32.50 USDT_{\mathrm{derived}}=2(12.50)+1(7.50)+0.00=32.50\ \mathrm{USD} For this fixture, quantities are item counts, prices are USD per item, tax is zero, and there are no discounts or additional charges. Round to cents. The stated total is USD 32.00: flag the discrepancy without overwriting it.

Schema validation checks representation; semantic validation checks contextual rules. Neither independently proves factual correctness. Structured Outputs and Tool Calling owns these interface mechanisms. Here, reconciliation exposes a conflict while preserving both the stated and calculated totals.

Extraction routes and their evidence

Choose routes by document conditions and required evidence.
RouteSuitable conditionEvidence and limitation
Native parsingUsable embedded textRetains encoded text; image content may require separate recognition.
Modular conversionText and layout need separate recoveryExposes intermediate regions and cells. Docling's original pipeline matches predicted cells to native PDF text, including borderless tables.
Template extractionStable, defined layoutsUses visual layout cues; template changes reduce accuracy.
Image-to-structureTask-specific page-image interpretationGenerates structured candidates; field coordinates and calibrated confidence are not automatic.

A vision-language model processes visual and language information. OCR-free means no separate OCR dependency: Donut generates structured information from document images, but still learns to read them. Removing a separate recognition stage does not remove recognition errors.

Generated structure still needs evidence

Example

An integrated extraction route retains an acceptance boundary.

Application policy for image-to-structure output: generated fields proceed only when required support and checks are satisfied.
Read the diagram as text
  • Page image.
  • Structured candidates.
  • Evidence and validation gate.
  • Accepted.
  • Unresolved.
  • Page image → Structured candidates: Model generates.
  • Structured candidates → Evidence and validation gate: Candidates submitted.
  • Evidence and validation gate → Accepted: Required checks pass.
  • Evidence and validation gate → Unresolved: Support missing or checks fail.

Routing can differ by page or region. Preserve surrounding headers when cropping, and retain disagreements between routes. In an extraction-first workflow, classifying a table as a picture sends it down the image path; downstream processing does not automatically restore the missing table relationships.

  • Coverage versus processing cost — Optional OCR and image descriptions add work. Disabling them is appropriate only when their missing output is acceptable. The OpenRAG demonstration presents focused models and a document-specific VLM without a measured quality or latency comparison.
  • Shared acceptance requirements — Evaluate every route against required fields and supporting evidence. Model architecture belongs in Vision AI; a simpler pipeline is not itself evidence of better extraction.

Correctness across recognition, structure and documents

CER=Sc+Dc+IcNc,WER=Sw+Dw+IwNw\mathrm{CER}=\frac{S_c+D_c+I_c}{N_c},\qquad \mathrm{WER}=\frac{S_w+D_w+I_w}{N_w} Character error rate and word error rate use edit alignment. S,D,IS,D,I count substitutions, deletions and insertions; NN counts reference characters or words. Subscripts identify the counting unit.

For reference INV-1042 and prediction INV-1O42, one substitution among eight characters gives CER 12.5%. Treating each identifier as one word gives WER 100%. The differing scores describe counting units, not different errors.

Insertions are outside the reference-length denominator, so error rates can exceed one. Empty references need an explicit policy. Specify character units and word segmentation before comparing results.

Different checks expose different failures.
CheckFailure it can expose
Reading sequenceCorrect words joined across unrelated columns.
Table relationshipsCorrect amounts attached to wrong rows or headers. GriTS separately evaluates topology, content and location.
Evidence localizationA field's text or page anchor points to the wrong supporting content.
Field-value matchingMissing, extra or incorrectly normalized values under an explicit matching rule.
Complete-document acceptanceAny required field, relationship or evidence check remains unsatisfied. Failed, partial and unresolved submissions remain in accounting.

Normalization changes the evaluated task. Case folding, punctuation removal or whitespace changes can erase differences that raw scoring counts. Apply declared transformations consistently to reference and prediction; removing decimal punctuation is inappropriate when amount fidelity matters.

For exact field-value matching, precision is correct predicted pairs divided by all predicted pairs; recall is correct pairs divided by required reference pairs. A wrong value contributes an unmatched prediction and a missed reference pair. Abstention removes a prediction, not the requirement.

Uncertainty and accountable review

Recognition confidence concerns detected words; field confidence concerns associations. Neither alone establishes document acceptance. Missing important keys can require review regardless of scores.

An engine-specific score is not automatically a correctness probability. Calibration compares assigned probabilities with observed frequencies for a defined event. High-confidence structural mistakes also require validation and auditing.

Review is a route, not a result

Example

Missing evidence and audits matter alongside scores.

Application policy. Alternative extraction is a bounded handoff endpoint; it must return through checks before acceptance.
Read the diagram as text
  • Document checks.
  • Accept.
  • Targeted review.
  • Alternative-attempt endpoint.
  • Request resubmission.
  • Unresolved or rejected.
  • Document checks → Accept: Complete, passes, no audit selected.
  • Document checks → Targeted review: Reviewable concern or audit sample.
  • Document checks → Alternative-attempt endpoint: Alternative justified; budget remains.
  • Document checks → Request resubmission: Required source unavailable.
  • Targeted review → Accept: Resolved; all checks pass.
  • Targeted review → Unresolved or rejected: Cannot resolve.
  • Document checks → Unresolved or rejected: Unsupported; no recovery available.
  • Review record — Retain source context, proposed value, failed check, reviewer decision and unresolved issues.
  • Completion boundary — Abstention withholds an automatic result. Straight-through processing completes a case without manual review. Sending work to a reviewer completes neither extraction nor resolution.
automatic coverage=AN,accepted error rate=EA\text{automatic coverage}=\frac{A}{N},\qquad \text{accepted error rate}=\frac{E}{A} For NN evaluated submissions, AA are automatically accepted and EE of those fail the reference acceptance criteria. Accepted error rate is undefined when A=0A=0. Abstentions remain in NN.

Lower accepted error can result from accepting fewer documents rather than improving extraction. Report coverage alongside accepted error, unresolved outcomes and review workload. Compare thresholds on labeled cases representative of the intended document population.

Corrections and dependent outputs

A source revision changes the document; an extraction attempt changes its processing; a reviewer correction changes an interpretation. Preserve all three separately. Record parser, model, configuration and schema versions. Revision fields alone do not implement a correction lifecycle.

Use application identities to reconnect reviewed elements after regeneration. A corrected cell association invalidates dependent records, even if its text is unchanged. Recheck affected outputs without erasing the original observation or reopening unrelated decisions.

Correction invalidates dependent output

Example

History remains; eligibility changes.

1 / 3 · Original

Retain observation.

Application policy. Replacement validation covers the corrected mapping; the invoice-total conflict remains unresolved.
Read the diagram as text
  • Cell C17.
  • Original: Cable.
  • Reviewed: Adapter.
  • Export e1.
  • Previously eligible.
  • Now stale.
  • Export e2.
  • Mapping validated.
  • Cell C17 → Original: Cable: Observed.
  • Cell C17 → Reviewed: Adapter: Corrected.
  • Original: Cable → Export e1: Produced.
  • Export e1 → Previously eligible: Status.
  • Export e1 → Now stale: Status.
  • Reviewed: Adapter → Export e2: Rebuilds.
  • Export e2 → Mapping validated: Status.
  1. Original. Retain observation. Active: Cell C17, Original: Cable, Export e1, Previously eligible. New: Cell C17, Original: Cable, Export e1, Previously eligible.
  2. Corrected. Withdraw old eligibility. Active: Cell C17, Original: Cable, Export e1, Reviewed: Adapter, Now stale. New: Reviewed: Adapter, Now stale.
  3. Rebuilt. Validate replacement mapping. Active: Cell C17, Original: Cable, Export e1, Reviewed: Adapter, Now stale, Export e2, Mapping validated. New: Export e2, Mapping validated.
  • Controlled replacement — Compare regenerated outputs with retained review decisions before restoring eligibility. Reproducible lineage supplies the broader versioning model; processing-history metadata does not guarantee this application policy.

Context-preserving downstream handoff

A chunk is a selected portion of document content; an index is a structure used to find content. Their implementation belongs in Search and Retrieval. Structure-aware handoff can retain headings and captions and repeat table headers so independently consumed rows keep their meaning.

The application export contract carries more than text.
PayloadRequired companion
Passage or recordHeadings, headers, units and relevant footnotes
Derived informationSource revision, extraction version, evidence selectors, unresolved states and applicable access metadata

Resolve exported content to evidence

Example

Searchable derivatives remain connected to original content.

Application contract: retain context and version-specific mappings. A summary locates an entity; the original entity supplies inspectable evidence.
Read the diagram as text
  • Exported record or summary.
  • Headers, units, qualifiers.
  • Version-specific evidence mapping.
  • Original revision and page.
  • Exported record or summary → Headers, units, qualifiers: Interpreted with.
  • Exported record or summary → Version-specific evidence mapping: References.
  • Version-specific evidence mapping → Original revision and page: Resolves to.

Generated descriptions remain model interpretations, separate from extracted captions and original pictures. Enrichment can add useful searchable language without replacing the evidence it describes.

Open questions

  1. Reliable multipage continuation remains difficult when repeated headers resemble new tables or qualifiers change. Wrong joins corrupt whole records. Progress requires source-annotated unfamiliar layouts where continuation and separation decisions preserve row identity and units.

  2. Corrections need durable correspondence across length-changing text edits and regenerated representations. Exact internal alignment may be unavailable, making automatic replay unsafe. Progress means identifying which reviewed spans still map uniquely and explicitly withholding ambiguous transfers.

  3. Automatic acceptance under document drift remains unresolved when structural mistakes receive high scores. Lower acceptance can conceal rather than fix failures. Progress requires joint evidence on accepted error, coverage and review burden across new layouts, including audits of apparently successful cases.

Selected talks

Explore the talk library →

Follow the curated reading path through the speakers and demonstrations behind this entry.

Explore more talks

The rest of the library, beyond the curated path. Cited talks support this entry; reviewed transcripts were processed in full. Metadata candidates have not been reviewed as sources or verified as topic members.

1 matching talks

TalkSpeakerEventYear
Tools for the Next Generation of AI Engineers

Transcript reviewed

Stefania DrugaAI Engineer Summit 20252025

References

Coverage and source review
Processed transcripts
5 processed in full · 4 in the curated path
Automated source review
Passed
Metadata candidates
0 unreviewed; not verified topic membership
Corpus version
1bd8e407b26a07b33815594e1b2db5f41827119a2b3cb6fbf240f9fc571fc767

Automated review checks source support; it is not publication approval.

A synthesis of selected conference talks and technical references. Citations link to the source material; they do not imply that every talk on this subject is included.

  1. W3C Technique PDF3: Ensuring correct tab and reading order in PDF documents

    Meaningful sequence and a concrete multicolumn failure despite visible text being present.

  2. OpenCV: Geometric Image Transformations

    Geometry underlying resizing, rotation, perspective correction and nonlinear image remapping.

  3. Microsoft Document Intelligence: Invoice data extraction

    Invoice-specific motivation and a concrete distinction between source recognition and semantic field extraction.

  4. PyMuPDF: OCR — Optical Character Recognition

    Recognition definition, image-only pages, mixed content and the distinction between displayed appearance and searchable text.

  5. Structuring the Unstructured: Advanced Document Parsing for AI Workflows

    The invoice example presents structured extraction of selected fields such as bill number, total price, and sender into a Pydantic format.

  6. PyMuPDF: Appendix 3 — Assorted Technical Information

    Page-to-image geometry and the distinction between successful opening and unmodified source integrity.

  7. W3C: Web Annotation Data Model

    Evidence-span definition, representation-specific offsets and spatial selectors; sections 4.2.4–4.2.7.

  8. VoiceVision RAG - Integrating Visual Document Intelligence with Voice Response

    The demonstrated ingestion pipeline renders each PDF page as an image and retains document and page metadata alongside it before embedding.

  9. Docling Technical Report

    Sections 3.1–3.3; a concrete modular extraction route separating recognition, structure and assembly.

  10. Tesseract: Improving the quality of the output

    Document-image defects, preprocessing and configuration-dependent recognition behavior.

  11. OCRmyPDF: v17 release notes

    Version-specific upstream explanations of duplicated OCR layers, font-related text corruption and stale search metadata.

  12. Tesseract FAQ

    Handwriting and searchable-PDF troubleshooting sections of the FAQ framed around Tesseract 4.

  13. OCRmyPDF: Advanced features

    OCR processing modes, mixed inputs, tagged PDFs and partial processing in the inspected 17.11 documentation.

  14. OCRmyPDF: Advanced features — Debugging the intermediate files

    Documented inspection artifacts for separating rendering, preprocessing and recognition failures.

  15. Docling: Docling document

    A concrete intermediate representation connecting content, containment, sequence and spatial information.

  16. hOCR 1.2: OCR Workflow and Output embedded in HTML

    First-use explanation of bounding boxes and hOCR; sections on bbox, cflow, x_wconf and encoding guidelines.

  17. Library of Congress: ALTO — Technical Metadata for Layout and Text Objects

    Brief introduction to ALTO and its explicit ordering and processing-history mechanisms.

  18. Structuring the Unstructured: Advanced Document Parsing for AI Workflows

    The PDF image extraction demo associates each source picture with its caption and embedded text elements.

  19. ICU4J: Edits.Iterator

    Offset mapping between literal text and transformed text.

  20. Google Document AI: Document REST resource

    Concrete evidence-bearing field representation, discontinuous support, normalization and extraction history.

  21. PubTables-1M: Towards comprehensive table extraction from unstructured documents

    Table vocabulary, spanning-cell ambiguity and document-specific structural evaluation; sections 1, 3 and 5.

  22. W3C WAI: Tables with Multi-Level Headers

    Authoritative example of table meaning requiring relationships beyond simple containment or column position.

  23. Microsoft Document Intelligence: Custom template document model

    Template-route applicability, layout drift and explicit cross-page grouping.

  24. Amazon Textract: Text Detection and Document Analysis Response Objects

    Concrete graph-like document representation and key-value association.

  25. Amazon Textract: Selection Elements

    Selection marks as visual evidence with structural context rather than ordinary transcribed words.

  26. OWASP Input Validation Cheat Sheet

    Semantic-correctness and result handling: Goals, Input Validation Strategies, and Implementing Input Validation.

  27. OWASP Input Validation Cheat Sheet

    Goals of Input Validation; Input Validation Strategies; Implementing Input Validation. Brief vocabulary bridge to /topics/structured-outputs.

  28. OCR-free Document Understanding Transformer

    Section 2 and Figure 3; an image-to-structure alternative to modular OCR pipelines.

  29. VoiceVision RAG - Integrating Visual Document Intelligence with Voice Response

    In the described extraction-first pipeline, upstream classification determines how the table enters downstream processing.

  30. OpenRAG: An open-source stack for RAG — Phil Nash

    Docling offers focused extraction models or a task-specific vision-language model, while optional OCR and image descriptions add processing cost.

  31. Hugging Face Evaluate: Character Error Rate

    Metric description, Output values, insertion example, and Limitations and bias.

  32. Hugging Face Evaluate: Word Error Rate

    Metric description, formula, inputs, and limitations; OCR application is a sequence-metric inference.

  33. Structuring the Unstructured: Advanced Document Parsing for AI Workflows

    Losing spatial structure can merge unrelated column text and flatten tables into text that no longer supports reliable extraction.

  34. Introduction to Information Retrieval: Precision and recall

    Textbook definitions, equations 38–39. Field-value matching and abstention consequences are a direct application of those set definitions, not a reported extraction experiment.

  35. JiWER: Text Transformations and Alignment

    Word error rate, Character error rate, Alignment, and Transformations sections.

  36. Amazon SageMaker AI: Human Loop Activation Conditions for Amazon Textract

    Document-specific review routing and separation of recognition and field-level signals.

  37. SelectiveNet: A Deep Neural Network with an Integrated Reject Option

    Section 2 Selective Prediction Problem Formulation, section 3 confidence-score approaches, and section 4 SelectiveNet.

  38. Selective Classification: Coverage and Conditional Risk

    Section 2, Problem Setting, selection function, coverage definition, and equation 1; empirical formulas are direct specializations.

  39. Docling: Chunking

    A short downstream handoff example; chunking implementation belongs in /topics/search-and-retrieval.

  40. Structuring the Unstructured: Advanced Document Parsing for AI Workflows

    The demo uses a local Granite vision-language model to generate image descriptions beyond the captions already extracted.

  41. VoiceVision RAG - Integrating Visual Document Intelligence with Voice Response

    Use summaries as searchable proxies, then resolve retrieved summaries back to original entities for answer generation.

  42. Structuring the Unstructured: Advanced Document Parsing for AI Workflows

    Docling combines OCR and layout analysis into a typed document representation that supports structural inspection and multiple exports, including extracted tables.

  43. POC to PROD: Hard Lessons from 200+ Enterprise GenAI Deployments

    The logistics example places a custom classifier before generative AI processing of receipts and shipping documents.