Kimi Docs is the document workflow inside Kimi for creating, editing, reviewing and converting Office files. Moonshot’s current documentation says Kimi can upload an existing Word document, insert comments and revision suggestions, handle headings, spacing, citations and lists, and return an original-format file such as DOCX or PDF for local editing.
Those capabilities need separate checks. Correcting 16 known errors is a factual task. Preserving editable tables and hyperlinks is a document-integrity task. Word comments are a traceability task. A clean PDF is a rendering task. One attractive preview cannot prove all four.
We built a compact synthetic DOCX with 16 intentional defects and a ground-truth ledger. The file uploaded and completed the entry UI’s analysis stage on August 4, 2026. Kimi then displayed a high-demand alert before review began. No corrected DOCX or PDF was produced, so the benchmark remains not run.
Independent guide: Kimi AI Guide is not affiliated with Moonshot AI. Product statements were checked against official Kimi pages on August 4, 2026. Our attempt used one Free account and one synthetic file. We report the access blocker, fixture and method; we do not invent correction or export results.
Kimi Docs at a glance
| Capability | Official description | What still needs independent checking |
|---|---|---|
| Word creation and editing | Long-form documents with professional formatting | Factual accuracy, style stability and local editability |
| Review | Comments and revision suggestions in existing Word files | Which issues are detected, whether comments anchor correctly and whether correct content survives |
| Structure | Heading hierarchy, paragraph spacing, citations and lists | Styles, bookmarks, links and tables after download and reopen |
| PDF work | Create, re-edit, extract and compare PDF content | Visual fidelity, text selection, page order and export integrity |
| Entry routes | Dedicated Docs page, mobile Docs mode or general Agent mode | The route, model, plan and date used in a test |
| Download | Original-format .docx, .xlsx or .pdf files | Non-zero valid package, no repair prompt, no clipping or missing glyphs |
The current Kimi Docs and Kimi Sheets overview documents the workflow and file outputs. Kimi’s Docs use-case library includes manuscript review with comments, contract review, version comparison, translation and PDF design. These are Moonshot’s examples, not results from our fixture.
Kimi Docs versus the general files guide
Our Kimi files guide explains upload limits, extraction strategy and long-document risk across file types. This page has a different search intent: it evaluates editing an existing Word document while preserving its structure, plus the DOCX and PDF exports. Uploading and summarizing a file is not the same task as making precise, traceable revisions to it.
How to review a Word document with Kimi
- Open Kimi Docs and sign in.
- Upload the source DOCX and wait until the UI finishes initial analysis.
- Name the authority inside the file or provide a separate reference packet.
- Define which changes are allowed: targeted correction, comments, tracked revisions or full rewriting.
- State preservation requirements for headings, tables, links, citations, page furniture and accurate text.
- Ask Kimi to expose unknowns and conflicts rather than guessing.
- Download the editable DOCX. Export PDF only as a separate deliverable.
- Reopen both files and inspect text, tables, hyperlinks, comments, revisions and every rendered page.
A review prompt that protects the source
Review the attached document using [named appendix or source packet] as the
authority. Correct only supported factual, numerical, calculation, unit,
grammar, consistency, citation and hyperlink errors. Preserve accurate content,
styles, headings, editable tables, working links and page layout. Do not invent
an owner, date, threshold or citation. Add concise comments for material changes
when supported. Return a clean editable DOCX and list unresolved issues.If you also need a redline, ask for it as a second file. A clean copy and a review copy serve different purposes and should not be conflated.
Our 16-error DOCX fixture
The source is a fictional operations-readiness memo. It contains no real organization, customer, employee or operational record. Appendix A acts as the internal authority for product name, reporting dates, ticket counts, service-level figures, resolution times, escalations and channel totals.
Error map
| Error family | Examples in the fixture | Count |
|---|---|---|
| Spelling and grammar | sucessfully, subject–verb agreement, duplicate word, missing article | 4 |
| Numbers and calculations | Wrong ticket count, SLA rate, growth rate and table total | 4 |
| Dates and units | Q2 ending in July instead of June, minutes instead of hours in prose and table | 3 |
| Meaning and consistency | Ambiguous confirming party and inconsistent product name | 2 |
| References | ISO/NIST citation mismatch, wrong NIST year and broken ISO hyperlink | 3 |
| Total | 16 |
Three editable tables, the memo masthead, heading hierarchy and official ISO and NIST references must remain intact. Deleting the sentence or table row that contains an error does not count as a complete correction.
Examples of ground truth
| Incorrect source | Expected correction | Basis |
|---|---|---|
| 1,284 Q2 tickets | 1,248 | Appendix A and channel total 610 + 420 + 218 |
| 92.0% SLA | 89.3% | 1,115 ÷ 1,248, rounded to one decimal |
| Ticket volume increased 8.4% | 11.4% | (1,248 − 1,120) ÷ 1,120 |
| Median resolution 7.5 minutes | 7.5 hours | Appendix A unit |
| Q2 through July 30 | Through June 30, 2026 | Defined quarter boundary |
| NIST AI RMF 1.0 published in 2024 | 2023 | Official NIST release date |
ISO link ending in .htmlx | Official .html URL | Official ISO destination |
Exact task submitted to Kimi
Review the attached operations memo. Correct factual, numerical, calculation,
unit, grammar, consistency, citation, and hyperlink errors by using the memo's
Appendix A and its official references as the authority. Preserve all accurate
content, the masthead, heading hierarchy, both tables, working links, and overall
layout. Do not rewrite correct passages unnecessarily. Return a clean, editable
DOCX. If Word comments are supported, add concise comments for material factual
or numerical corrections, but do not let missing comments block the corrected
file.Result: upload succeeded, review did not start
| Check | Observed result |
|---|---|
| Signed in | Yes |
| DOCX uploaded | Yes, 41,728 bytes |
| Entry-stage file analysis | Completed |
| Prompt submitted | Yes, at 18:44:32 PDT on August 4, 2026 |
| Review execution | Did not begin |
| Visible response | High-demand alert before execution |
| Corrected DOCX | None |
| Word comments or tracked revisions | Not testable |
| Table and hyperlink preservation | Not testable |
| PDF export | None |
| Final score | Not calculated |
The source DOCX SHA-256 is 7A6655AAF59443C18BAA2429821A679686B0700FB3D10AC282F36D6CC0B76771. Before upload, the fixture passed 13/13 structural checks and an accessibility audit with no reported high-, medium- or low-severity findings. The fixture’s visual render check was not run because LibreOffice was unavailable in the test environment. We do not present structural QA as a visual-layout pass.
Download the Kimi Docs benchmark pack. It includes the DOCX, complete 16-error ledger, official targets, 100-point rubric, structural reports and raw blocked-run record.
How a completed output will be scored
| Category | Points | Requirement |
|---|---|---|
| Sixteen intentional corrections | 80 | Detect the issue, replace it with the supported correction and preserve the surrounding sentence or table row |
| Accurate-content preservation | 6 | Accurate text, metadata, headings and Appendix A facts remain |
| Editable tables | 5 | Tables remain structured, legible and corrected rather than flattened |
| References and hyperlinks | 4 | ISO and NIST references remain and open the intended official targets |
| Render integrity | 5 | No clipping, overlap, missing glyphs, broken furniture or major style drift |
Comments are recorded separately. They are useful evidence but not part of the core 100 points, because a document can be correctly fixed without comments and heavily commented while still wrong.
DOCX and PDF checks that text extraction misses
Tables
Confirm that rows, columns and cell text remain editable. Check widths, repeating headers, borders and page breaks. A screenshot of a table is not a preserved Word table.
Hyperlinks
Visible link text can look correct while the relationship target remains broken. Inspect and open the actual ISO and NIST targets in the downloaded DOCX and PDF.
Comments and revisions
Count comments, confirm each is anchored to the relevant correction and inspect author/date metadata. If tracked changes are requested, verify insertions and deletions in a Word-compatible editor rather than relying on clean text extraction.
Rendering
Render every DOCX and PDF page. Look for clipped tables, orphaned headings, overlaps, substituted fonts, missing symbols, broken headers and unexpected blank pages. A valid OOXML zip package can still render badly.
Limits and next test
This attempt proves only that the signed-in Docs entry page accepted and analyzed this one DOCX before a capacity blocker. It does not measure correction accuracy, comment quality, Word preservation or PDF fidelity. The next valid run must retain the unmodified source, corrected DOCX, optional redline, PDF export, raw transcript, comments, relationship targets and rendered page images.
See our methodology for stop rules and the Kimi AI Test Lab for other controlled fixtures.