Article content
**Query date: 2026-09-22.** This guide is about collecting what an Office Action cites. It is not an IDS filing guide. For the later decision about which references to attach, and how to handle non-patent literature (NPL), see IDS preparation: what to attach.
Start with the OA, not with a guessed number
When a reviewer says “this OA has twelve citations”, the first job is to make a checkable inventory of those twelve items. A citation can be in the opinion text, a cited-reference table, an IDS page, a PTO-892 form, or an official search-report annex. A list copied by hand from one page is therefore not a complete collection method.
Without a citation tool, the safe manual sequence is:
1. Make a copy of the original PDF and record its filename and query date. 2. Read the cover and identify the office, application and document type. 3. Search the PDF for headings such as `PTO-892`, `Notice of References Cited`, `FOREIGN PATENT DOCUMENTS`, `NON-PATENT DOCUMENTS`, `search report`, and the local equivalent. 4. Transcribe every row in document order, retaining the printed country, number, kind code, year and source page. 5. Resolve each patent identifier in the responsible office or a reliable patent database. Keep the printed citation beside the normalized value; normalization is not proof that the source has that publication. 6. Download the original patent PDFs one by one, rename them with the case and citation relationship, and keep a separate list of anything not retrieved.
This is slow but auditable. Do not stop after finding the first searchable number. A continuation, an older grant, an English family publication and an NPL row may all be separate entries.
How the parser finds the citation blocks
The OA pipeline first extracts PDF text, keeps page and block positions, and uses both the cover text and filename as country evidence. Strong US markers include `United States Patent and Trademark Office`, `USPTO`, `PTOL-326`, `PTO-892`, `PTO/SB`, and `Office Action Summary`. A filename such as `P900033USI` is a useful fallback hint, not a replacement for the text.
It then scans in layers rather than trusting one regular expression:
| Layer | What it does | Why it matters |
|---|---|---|
| Country-aware extractor | Reads the likely office's number and kind-code forms | Avoids treating a date or an application number as a patent publication |
| Supplemental sweep | Looks for OCR-degraded US, foreign and PTO-892 shapes | Recovers rows where spaces, punctuation or table cells were lost |
| Context and deduplication | Preserves first-seen block and drops shorter OCR truncations | Keeps the official order without counting one citation twice |
For an IDS form, the country and number may be separate table cells. The IDS-specific extractor therefore unmerges fused rows and requires country evidence. It handles, for example, an eight-digit US grant with `B1` or `B2` and year, an eleven-digit US publication with `A1` or `A2` and year, and foreign rows such as `104704658 CN A`. Concrete OA examples include `US 2024/0039765 A1`, `CN 104704658 A`, `特開2017-165727`, and `TW I606827`. These examples show why a bare digit string is not enough. For a country-by-country number and kind-code reference, use patent number examples and kind codes rather than copying a second format guide here.
US, JP and CN documents are not interchangeable
The US path recognizes PTO-892 tables, `Notice of References Cited`, IDS markers such as `PTO/SB/08`, and the separate `NON-PATENT DOCUMENTS` section. While scanning a PTO-892 page it masks the NPL section before applying patent extraction, so a journal date or an ISBN is not promoted to a patent number. An IDS marker also closes a preceding PTO-892 block; the two sources remain distinguishable in the result.
JP documents often put citations in a numbered official list and may print an imperial-era identifier. The regional NPL extractor looks for positive bibliography signals such as `[online]`, journal or volume/page data, ISBN, or a publisher-like entry. It does not infer “NPL” merely because the row contains Japanese text.
CN search reports need a different boundary. The locator finds the search-report section, collapses letter-spaced CJK headings, and masks the examiner query or search-record portion. A search query log can contain `/pn` family lists, but it is not itself cited art. A citation table that follows the report marker is retained. This distinction is why copying every number from a CN PDF produces both false positives and missing references.
The result keeps where each item came from (the opinion text, a PTO-892 table, an IDS page or a search report) together with the number exactly as printed. That lets a reviewer answer “where did this item come from?” instead of seeing only a normalized identifier.
Reconcile a twelve-reference OA before downloading
For a real docket, use a reconciliation sheet rather than relying on the number shown in the result heading. Give every candidate a stable row number and fill these columns:
| Check | Question |
|---|---|
| Source block | Is it from the rejection text, a search report, PTO-892 or IDS? |
| Printed form | What exactly did the office print, including spaces, slashes and kind code? |
| Normalized value | What value will the resolver search, and which country supplied the evidence? |
| Match | Does the returned title, date and kind code agree with the OA row? |
| Artifact | Is there an original PDF, an English family PDF, neither, or a corrupt response? |
| Next action | Download, retry, manual lookup, or IDS review only? |
Suppose rows 1–7 occur in the opinion text, rows 8–10 in a PTO-892 table, row 11 on an IDS page and row 12 is a journal article. The correct result has twelve source rows, but only eleven patent candidates may enter the patent download path. If row 10 appears once in the text and once in PTO-892, deduplication can merge the identifier while retaining both source memberships. If row 12 has a DOI but no patent number, it remains one NPL row; it does not reduce the patent count by pretending that the DOI lookup failed.
OCR creates a second reconciliation problem. A table can turn `US 2024/0039765 A1` into `US20240039765 A1`, split `CN 104704658 A` across cells, or truncate a longer number to a prefix. The parser normalizes full-width digits, letters and dashes, preserves the raw text, and drops a shorter candidate when the same country and continuation context prove that it is an OCR truncation. A human should still compare the returned title and date. If country evidence conflicts, keep the row unresolved instead of forcing a country from the digits alone.
When the PDF is scanned or unusually long
Reading the text layer is faster and easier to audit, but a scanned OA may have no usable text. For a citation list, DownloadPatent reads the text of up to the first 150 pages, locates the report section, and uses OCR only for those report pages. If the OA is a scan that cannot be read, the result says so instead of reporting that zero citations were found. That is materially different from an OA with a confirmed empty citation section.
Tables can also cross a page break. The parser merges continuation blocks before discovery, while retaining page and block provenance. A manual workflow should do the same: when a country column is on one page and the number or kind code continues on the next, copy the whole row before deciding that it is incomplete. Do not use a search-engine snippet to fill a missing cell; record the source page and resolve it separately.
The practical stopping rule is simple: stop only when each expected row has a source location, a classification, a normalized lookup value or an explicit unresolved status. “The ZIP contains twelve files” is not that stopping rule. It can be eleven patent PDFs plus one duplicate, or twelve files with one wrong family member. The reconciliation sheet is what makes the handoff to IDS preparation defensible.
Before closing the collection task, make four small checks. First, compare the first-seen order with the OA pages; a later family lookup must not reorder the office's list. Second, count patents and NPL separately, and count unresolved patents separately from confirmed NPL. Third, open a sample from each country path, including at least one foreign publication and one US PTO-892 row. Fourth, save the source PDF, the result list and the failure notice under the same matter reference. Those checks are useful even when every download succeeds because a correct file attached to the wrong citation is still a docketing error.
If a reviewer asks why a row is absent, the answer should be a status and a source location, not “the system did not find it”. `not_relied` or an applicant-disclosed IDS membership can explain why a row is grouped differently, but it does not erase the row from the collected evidence. Keep those meanings separate when exporting a list for counsel. Collection is complete when the evidence is accounted for; attachment selection comes next.
NPL is a classified result, not a failed patent download
NPL means non-patent literature: an article, standard, RFC, book, conference paper or online publication. The extractor accepts positive evidence such as a DOI, ISBN, RFC, IEEE or ISO label, a URL with an access marker, or a year combined with journal, volume, pages, conference or author evidence. Official US tables are read under `NON-PATENT DOCUMENTS`; IDS tables use `NON-PATENT LITERATURE DOCUMENTS`. Regional search-report lists have their own positive markers.
After extraction, a cross-check asks whether the row is actually a patent citation in disguise. If removing the patent number, date, list number and separators leaves only a few characters, the NPL candidate is discarded as a patent row. A row containing digits is not automatically a patent, and a row containing a title is not automatically NPL.
Each result row is marked as a patent or as NPL. An NPL row stays visible for review with its source and any DOI, but it never enters the patent PDF download. Do not turn that state into an empty PDF or a “download failed” count. Whether an NPL publication or relevant portion belongs in an IDS package is the separate decision covered in IDS preparation.
Name the package so a second person can audit it
The package should preserve the cited number and distinguish original documents from family versions. The current filename rules are:
| Artifact | Filename shape |
|---|---|
| Original citation | `${normalized}.pdf` |
| English family publication | `${sourceClean}__EN__${normalized}.pdf` |
| IDS counterpart | `${officialPrefix}${citedNormalized}__${languageTag}_${normalized}.pdf` |
| ZIP package | `${caseOrFileStem}_${country}_${citationLabel}_${count}refs_${YYYYMMDD}.zip` |
The date in the ZIP name is UTC. A case stem can be recovered from supported filenames such as `ACME_P900033USI_...pdf`, but it should not replace the citation's original printed identifier. DownloadPatent currently delivers original PDFs and available English family publications. Google machine translation is an external Google Patents link; it is not a selectable ZIP artifact. The system does not create rebuilt text PDFs or DOCX files.
The tool has its own safety limits: 500 selected items across primary references, annexes and versions, and 100 IDS counterparts. These are DownloadPatent limits, not USPTO submission rules. A firm doing the same work manually or with another service must check that service's limits independently.
A missing file must stay visible
The download stage distinguishes `failed-not-found`, `failed-corrupt`, `failed-no-id`, and service errors. Confirmed not-found items are not given a guessed URL. A transient source error can retain a manual lookup link, but it is still a failure until someone verifies it. The ZIP adds `未能下載清單.txt` only when there are failures, and the notice includes every in-scope patent with no PDF, including an unselected confirmed failure that would otherwise be easy to overlook.
A successful ZIP therefore means “these selected patent PDFs were assembled”, not “every string in the OA became a file”. Before handing it to the attorney, compare the citation count, the NPL list, the failure notice and the source-page order. Then use the IDS guide to decide what should actually be attached.
DownloadPatent applies this same collection and failure accounting to the OA PDF; it does not replace the source office, the attorney's review, or the IDS filing decision.