HWP and HWPX document search: validating tables and revision history
Evaluate AI search for Korean HWP and HWPX documents by checking tables, footnotes, embedded images, revisions, and source evidence with representative files.
Evaluate AI search for HWP and HWPX documents by whether it extracts the information and conditions your work requires, and lets you verify them in the original document. Accepting a file is only the beginning. Tables, footnote exceptions, and the effective dates of different revisions can make ordinary text search an incomplete measure of quality.
HWP and HWPX are document formats used by Hancom's Hangul word processor, widely associated with Korean-language documents. Public-sector guidelines, internal policies, and older project reports may contain a mix of both formats. Before converting the entire collection, choose representative documents from one workflow and test processing, search, answers, and source review in sequence. The following approach gives document owners and IT teams a shared basis for acceptance testing.
What is the difference between HWP and HWPX?
HWP uses a binary format, while HWPX packages XML files and other components in a ZIP structure. Hancom's explanation of the HWPX format describes body content and formatting alongside images, embedded objects, metadata, and previews. An HWPX extension does not mean that everything inside the document is plain text.
This distinction matters when selecting a processing method. A more open format does not, by itself, ensure accurate retrieval of tables and footnotes or identify the latest version approved by your organization. Assess the ability to read the file structure separately from the ability to preserve its business meaning.
| Input | First check | Test separately |
|---|---|---|
| Original HWP file | Whether that file version can be read | Extraction of body text, tables, footnotes, and text boxes |
| Original HWPX file | Whether document sections and objects are processed | Text in images, table relationships, and reading order |
| PDF converted from Hangul | Whether content and pages match after conversion | Searchable text and links back to the original |
| Scanned image embedded in a Hangul document | Whether embedded images are processed | Agreement between OCR output and the actual image |
| Password-protected or restricted document | Whether an authorized method can read it | Failure reporting and access boundaries |
When a vendor says it supports Hangul files, ask which formats and versions it processes and what it excludes. Check separately whether the filename merely appears in a list, whether the body is searchable, and whether an answer can open the relevant source.
Select files around information that must not be wrong
Picking a few documents at random and requesting summaries may test only the easiest files. First list the information employees actually need, then select documents containing it. Start with a narrow workflow, such as policy guidance for one department, and include:
- Documents with answers in ordinary paragraphs, and others with answers only in tables.
- Merged table cells, tables continuing onto another page, and footnotes.
- Official notices or test results pasted into the document as images.
- Current and previous versions of the same policy, plus an unapproved draft.
- Files that open successfully but contain material that is difficult to read or access-restricted.
For each document, record its original location, owner, relevant workflow, approval status, and effective status. In the expected-answer sheet, specify required values, units, conditions, and supporting passages rather than prescribing one complete answer sentence. The same collection can then support comparisons between vendors.
Before a bulk import, also agree on how processing failures will be reported. Distinguish whole-file failure, missing objects, completed text extraction, and readiness for search. The number of connected files is not necessarily the number that can be used successfully.
Test relationships between table rows, columns, and conditions
Consider a hypothetical test document created for acceptance testing. An expense table has rows for domestic and international travel, with columns for transportation and accommodation. A footnote requires separate approval under certain conditions. This is an illustrative test, not an actual company's reimbursement policy or a customer case.
If you ask for the accommodation allowance, is a numerically correct answer enough? Omitting the domestic or international category, the unit, or the approval exception can change its practical meaning. Write questions that test those elements separately.
| Example test question | Evidence to record in the expected-answer sheet | Failure condition |
|---|---|---|
| “What is the domestic travel accommodation allowance, and what is its unit?” | Value and unit from the correct row and column | Uses the international allowance or omits the unit |
| “When is separate approval required?” | Footnote and the item it applies to | Gives only the general rule and misses the exception |
| “Which employee group does this table apply to?” | Section heading and the scope stated in the body | Reads the table alone and broadens its applicability |
| “Has the guidance in the embedded image been included?” | Relevant wording inside the image | Reports completion despite failing to read the image |
If a table is converted to text, check whether labels remain associated with their values. Joining numbers from several rows into one line may preserve every character while losing the relationships between them. When headings, table descriptions, and footnotes become separate chunks, test whether search retrieves the necessary context together.
Material pasted as an image may need different processing from the document's body text. Compare the original against results for poor scans, small print, and areas obscured by stamps. For collections with many such files, apply the checks in the OCR document search guide as well.
Distinguish the most recently saved file from the policy in effect
The file with the newest modification time is not necessarily the policy currently in force. Someone may have saved an old policy again, or uploaded a revision that takes effect next month. The word “final” in a filename is not a substitute for approval status either.
For policy search, organize document IDs, revision numbers, approval status, effective dates, withdrawal status, and responsible departments. If a file does not contain this information, its owner needs a separate way to maintain it. Do not rely on AI to infer an approval decision that is not recorded.
During testing, ask about both current requirements and a past date. “Which rule applies now?” and “Which rule applied when this work was completed last year?” may require different sources. Define both the purpose of retaining historical documents and the basis for answering current questions.
After adding a revision, check whether the new document is searchable, whether the previous version is incorrectly presented as current, and whether a question about the past retrieves the appropriate history. Extend the exercise to deletion and revoked access using the RAG document updates and deletion guide.
Reviewers must be able to find the evidence again
A filename attached to an answer is difficult to review if there is no passage supporting its conclusion. Check whether document IDs, versions, section headings, table names, and quoted passages lead the reviewer back to the evidence. Opening the source should reveal the cited content in the same document.
If the system provides page numbers, define what those numbers refer to. Use representative files to check whether locations differ between the original Hangul display, print settings, and converted PDF. When a converted document is the evidence view, record which original version produced it. Regenerate a PDF and check whether existing citations now point somewhere else.
Public-sector documents and internal policies often require a person to confirm effective dates and exceptions. A useful acceptance criterion is therefore “the reviewer found the same condition in the source,” rather than merely “the answer reads naturally.” The RAG PoC evaluation guide explains how to assess retrieval, answers, and citations separately.
Check where conversion and search actually run
Keeping the original on an internal server does not establish where conversion, OCR, or answer generation takes place. Organizations with restrictions on external data transfer should identify the locations of the HWP/HWPX processing software, converted files, extracted text, and model calls separately. If file conversion is being added, also review the operating environment, usage permissions, and required licenses.
Search access must match the original document's access boundaries. A collection account's ability to read a file does not mean that every employee may view it. Test restricted document titles, answers, citations, and source access with accounts that have different permissions, then repeat the tests after access is revoked.
In an air-gapped environment, include the procedures for bringing in and updating document processors and models. Installation alone is not enough for sustainable operation: check whether failed files can be reprocessed and processing problems diagnosed after deployment.
Frequently asked questions
Will converting every HWP file to HWPX solve AI search problems?
A format conversion does not establish table meaning, effective dates, or user permissions. Convert representative files first, compare the retained content and objects, and test search and source review. Preserve the originals and record their relationship to the converted versions.
Would converting everything to PDF make search simpler?
That may be a suitable choice, depending on the tools and document structures involved. Check for missing objects, usable text and tables, and a process that keeps converted files synchronized with revisions to the originals. Compare conversion approaches using the same documents and questions.
Can Wissly connect directly to our HWP and HWPX files?
Provide the file versions, object structures, storage systems, and intended deployment environment so the processing scope can be reviewed. Test file registration, body and table extraction, image OCR, and source review separately, and agree on any necessary conversion or integration. Do not assume that every Hangul document will be processed in the same way.
Bring questions as well as file counts to your consultation
Wissly supports a workflow for searching connected internal documents by content, checking source evidence, and drafting work documents. When assessing HWP and HWPX use, prepare the formats and versions of representative files, their table and image structures, examples of missing content, and actual employee questions.
A Wissly deployment consultation can examine document processing, required conversion, search, and source review in the context of your storage and security requirements. The objective is a workflow in which employees can find the applicable conditions and return to the same evidence, beyond simply registering a large collection of Hangul files.
