How do you validate PDF, table, and image extraction before RAG indexing?

A RAG index can contain every visible word and still misrepresent the source. OCR may put a warning under the wrong procedure. A spreadsheet extractor may return the number 2 without its column heading or unit. A diagram’s labels may be searchable even though its arrows were never represented. Accept extraction against the original document before splitting or indexing it.

This guide starts after a source revision has passed the source admission gate. It checks whether a particular parser run preserved the parts of that source needed for the intended questions. The chunking trial then compares boundaries on a verified extraction; the retrieval trial tests indexed passages; the citation audit checks claims after generation. A good score at either later stage cannot recover a fact that parsing discarded.

Freeze source regions before looking at parser output

Choose a controlled source revision and record its ID, checksum or version, format, and the parser or OCR route with its version and settings. Select decision-bearing regions from the original, not from what the parser happened to return. Cover the formats your workflow really ingests: selectable-text and scanned PDF pages, rotated or two-column layouts, tables with merged headers and units, footnotes or captions, and images whose visual relationships answer a question. Include a region with no usable text so a blank extraction cannot silently pass.

For each region, a knowledgeable reviewer records the exact fact and relationship a future answer needs: an amount with its unit, a negation attached to the right step, a cell under its row and column headers, or an arrow connecting named components. Give the location in the original: page and bounding area for a PDF or image; worksheet and cell or range for a spreadsheet. Store original files and raw parser output in approved storage. Put only synthetic IDs and controlled location codes in the browser worksheet below.

Set blocking rules before inspecting results. A typo in irrelevant decoration may be repairable later; a wrong value, missing “do not,” swapped header, lost figure relation, or untraceable source location blocks that region from grounding a production answer. Test a few unseen files after tuning the parser so the acceptance decision is not fitted only to the examples used for repair.

Compare five things with the original

Gate	Compare the extracted representation with	Block when
Critical text	Numbers, signs, units, dates, names, negations, and selection marks in the source region.	A changed or absent token can reverse the intended decision.
Reading order	Headings, columns, warning boxes, captions, and the procedure or figure they qualify.	Correct words are attached to the wrong section or step.
Table relationship	Row and column labels, merged headers, cell spans, units, titles, and related notes.	A value cannot be tied to the right entity, condition, or unit.
Visual relationship	Diagram arrows, legend symbols, callouts, and any caption needed to interpret them.	The output lists labels but omits the relation that the question asks about.
Provenance	Original page and region, or worksheet and cell range, plus source revision.	A reviewer cannot get back to the specific original evidence.

Review the exact representation that will be indexed, not only a visually attractive preview. Preserve the raw structured output as a diagnostic record. Azure’s OCR skill can return plain text and layoutText with line locations; the latter helps a reviewer see where words came from, but a box is not proof that the words were correctly grouped. For supported inputs, Azure Document Intelligence Layout v4.0 exposes page, paragraph, and table-cell positions, including row and column indices and spans; its XLSX text extraction does not provide table analysis. When the TABLES feature is requested through Amazon Textract’s table-analysis route, its blocks and relationships represent cells, merged cells, titles, and footers. These structures are useful comparison material, not independent truth about the original.

Check the chosen product’s actual limits. Azure’s Blob indexer extracts textual content and metadata from supported files, while images need enrichment for more search utility. Its supported file and extracted-character limits, plus configurable error tolerance, mean that an indexer run’s success is not a content-accuracy certificate. Even a high OCR confidence does not prove that a warning belongs to the right step or that a table header governs the right value.

Treat unsupported elements as gaps, not empty facts

One parser need not handle every format. In Azure Document Intelligence Layout v4.0, Office text can be extracted, but embedded images in Office files are not supported; its XLSX input does not receive table analysis. That is a limitation of this route, not a claim that all spreadsheet parsers lose tables. The same documentation says figure and table bounding regions exclude associated captions and footnotes. Those words may still appear elsewhere in the overall output; verify that the indexed representation links them to the right figure or table. Its Markdown option renders tables as HTML in v4.0 to preserve merged cells and multirow headers; a convenient text view still needs comparison with the original cell relationships.

If a required region is unsupported or fails, keep it out of the candidate answer corpus until it is repaired. A native spreadsheet reader may preserve worksheet cells and merged ranges; a scanned page may need a different OCR or layout route; a diagram may require a reviewed structured description linked to the original image. Recheck the new indexed representation against the same marked regions. Do not invent a diagram relationship from visible labels, let an image description approve itself, or try to fix a missing table header by adding more chunk overlap.

Once critical regions pass, freeze the extraction output and its parser configuration for the chunking comparison. If only part of a source passes, document the allowed question scope and block questions that depend on the failed region. A parser change, new file layout, or newly added image requires another extraction check before the source’s chunks are called ready.

Example one: a warning moves to another procedure

In this fictional test, page six of a scanned HelioDrive D9 maintenance manual has two columns. In the left column, a boxed warning says isolate breaker Q4 before removing the inspection cover. The right column starts a different replacement procedure. A plain OCR text stream contains “Q4,” “inspection cover,” and the warning sentence, but places it after the right-column heading. A keyword search can find every word while a later answer attaches the safety condition to the wrong procedure.

The reviewer marks the warning box, its left-column section heading, and the cover-removal step on the original page. The acceptance row checks the words, their sequence and section association, and a location that opens the same page area. The run fails if the warning remains attached to the replacement step, regardless of OCR confidence or a correct-looking final answer. Repair the layout route or supply a reviewed structure that binds the warning to cover removal, then rerun the same region and an unseen two-column page.

Example two: the spreadsheet loses headers and a diagram

In another fictional test, a Frostline R7 cold-chain workbook has a Routes sheet. Merged headings distinguish Chilled and Ambient routes; beneath them, separate columns give maximum temperature in °C and hold time in hours. The Bay C row says that its Chilled route has a two-hour limit. An embedded schematic uses an arrow and legend to show which valve reaches the bypass branch. Two intended questions ask for Bay C’s Chilled hold time and for the bypass valve.

A flat text export lists Bay C, 2, and several temperatures without retaining the merged heading or units. After the embedded schematic is exported separately, an image OCR route finds the valve labels but never represents the arrow. Neither question has complete extracted evidence. For the table, compare the original sheet, row, merged header, unit, and cell coordinates with a representation produced by a route that actually preserves them. For the schematic, create or obtain a reviewed description tied to the embedded image, or exclude that question. Azure Layout v4.0’s ability to extract XLSX text cannot be reported as successful XLSX table analysis or embedded-image understanding.

Decide whether the extraction can be admitted

For each sampled region, the worksheet records original location, required fact and relation, parser route, output location, observed mismatch, severity, owner, and pass / repair and rerun / exclude. Record a positive control that remains usable beside a blocked region; otherwise excluding an entire file can conceal a too-broad fix. The release owner accepts the extraction only when every predefined blocking region passes or its dependent question class is explicitly removed. Keep a controlled pointer to the original and raw output so another reviewer can reproduce the comparison.

Boundary

This is a representation-fidelity check, not a judgment that the original is authoritative, licensed for ingestion, current, or accessible to a requester. The source gate decides admissibility; access and update checks test the serving path. Extraction acceptance also does not certify retrieval rank, answer faithfulness, or citation support. Those need their own tests after the original material has survived parsing.