DOCSIFTER / CHINESE DOCS REVIEW

Protect code and structure.
Sift out the text issues.

DocSifter takes its name from sift: separate what must stay untouched before reviewing what can change.

Code blocks, SQL, API parameters, and product terms stay intact. Typos, awkward phrasing, and formatting issues are traced to their source lines.

01 / LOCAL PASSRules + local small model02 / DEEP REVIEWVerify edits · Check context03 / FEEDBACK LOOPRecord false positives · Reuse

Scroll to see the review pipeline unfold

01 / STRUCTURE FIRST

Separate code from prose.
Review only what readers read.

SQL blocks, configuration, and comments inside code blocks stay protected.
Markdown and AsciiDoc markup are separated from prose,
so only readable text enters review.

CODE AND CONFIG STAY OUT · MARKUP STAYS INTACT · LINE NUMBERS STAY

02 / RULES + LOCAL REVIEW

Rules handle known issues.
The local model handles context.

Rules handle deterministic terminology and wording first.
The local pass checks context-dependent typos.
Every suggestion points back to its original line.

RULE: log-in → log in · CONTEXT: databse → database

03 / LARGE MODEL, ON DEMAND

Verify the suggested edit.
Then check across lines.

First, verify each first-pass suggestion and whether it changes the meaning.
Then read across lines for contradictions, terminology drift, and broken references.

OPTIONAL · SELF-HOSTED OR OPENAI-COMPATIBLE
SET FILE SCOPE AND BUDGET CAPS · REMOTE FAILURES DO NOT BLOCK LOCAL REVIEW

04 / FALSE-POSITIVE FEEDBACK

Mark a false positive once.
Filter it on later runs.

False positives marked in the report are stored locally
and filtered from later runs. Export the records
for evaluation baselines or offline fine-tuning.

STORED LOCALLY · REUSED ON LATER RUNS · EXPORTABLE FOR TEAM DATASETS

05 / LOCAL FIRST · CONTROLLED COST

Keep sensitive docs local.
Run the baseline without cloud APIs.

Rules and local review handle routine batch checks. Once the model is available, the baseline runs offline.
Without a remote endpoint, document content stays on the machine.

0Cloud API bill for
local baseline reviewHardware and power excluded
Deep review follows backend pricing

Enable deep review only when needed, with file-scope and budget limits.

06 / DOCS-AS-CODE

Build a full baseline.
Guard every pull request.

Run local checks from the CLI, batch reviews in the Web UI, and incremental checks through GitHub webhooks.
Every suggestion keeps its source, rule ID, and original line. The tool finds issues; people make the call.

View DocSifter on GitHub OPEN SOURCE · MIT LICENSE

LOCALREVIEW ENGINE

connection.adocASCIIDOC

== Database connection
Same prose, different markup.[source,sql]
----
SELECT id FROM orders;
----

PROTECTED CONTENTSTRUCTURE PROTECTED

STRUCTURE PROTECTED

SQL AND CODE BLOCKS
COMMENTS INSIDE CODE BLOCKS
API PARAMETERS, PATHS, AND CONFIG KEYS

connection.mdDOCUMENT EXCERPT / UTF-8

##Database connection

Please **log-inlog in** to the console.

```sql

-- Query orders; keep this code unchanged

SELECT id FROM orders;

```

SQL + BLOCK COMMENT / KEPT INTACT, NOT REVIEWED

Connect to the databsedatabase and check sync status.

Retries automatically after failure.

Requires a manual retry after failure.

LINE NUMBERS KEPT · PROTECTED CONTENT STAYS OUTWAITING FOR PROSE

MARKUP / SEPARATED FROM PROSE

MD ## ** ```sql

ADOC == [source,sql] ----

L03 / RULE ENGINETERM STANDARD

~~log-in~~→log in

L11 / LOCAL SMALL MODELCONTEXTUAL TYPO

~~databse~~→database

Review.

OPTIONAL / DEEP REVIEWON DEMAND

Verify suggested edits.
Check across lines.

01 / SECOND-PASS REVIEWL11 · SUGGESTED EDIT

~~databse~~→database

Use the surrounding text to verify the edit and whether it changes the technical meaning.

02 / CROSS-LINE CONSISTENCYSEPARATE SWITCH

L12Retries automatically after failure.

L24Requires a manual retry after failure.

CONTRADICTION FOUND

Check the implementation to confirm the actual retry behavior.

FLAG THE ISSUE AND EVIDENCE; DO NOT REWRITE FACTSCONTEXT

HUMAN FEEDBACKSTORED LOCALLY

Reject a false positive.
Filter it next time.

L11 / PRODUCT NAMEREJECTED BY REVIEWER

~~DocSifter~~ → Documentation SifterMARKED FALSE POSITIVE

01 / FINDFlag the issue

02 / REJECTConfirm false positive

03 / STOREKeep it locally

04 / EXPORTEvaluate or fine-tune

REUSED ON LATER RUNS · EXPORTABLE FOR TEAM DATASETSLOOP

00 / DEFINE THE BOUNDARYSCROLL TO SIFT · PRODUCT STORY

PRODUCT FILM / 02:31

Watch DocSifter
review a document.

From structure protection and local review to optional deep review and false-positive feedback. See how the complete workflow fits into everyday documentation work.

LANDSCAPE DEMOPLAY ON DEMAND · AUDIO CONTROLS

DOCSIFTER / ENGLISH PRODUCT DEMO02:31 · 1920 × 1080

Let the tool handle mechanical checks.
Keep judgment with the team.

DocSifter
Built for Chinese technical documentation,
with CLI, Web UI, and GitHub pull-request workflows.

Protect before correcting

Keep code blocks, configuration, and comments inside code blocks out of review, while preserving markup.

Keep routine review local

Rules and local review run the baseline. Once the required resources are available, this layer works offline.

Review difficult cases on demand

Use optional deep review to verify edits and find cross-line conflicts, with scope and budget limits.

Reuse false-positive decisions

False positives are stored locally, filtered from later runs, and available for export.

Protect code and structure.Sift out the text issues.

Separate code from prose.Review only what readers read.

Rules handle known issues.The local model handles context.

Verify the suggested edit.Then check across lines.

Mark a false positive once.Filter it on later runs.

Keep sensitive docs local.Run the baseline without cloud APIs.

Build a full baseline.Guard every pull request.

Watch DocSifterreview a document.

Let the tool handle mechanical checks.Keep judgment with the team.