Everything about what the AI reads, what it costs, and what happens to your documents.
ProSearch AI answers questions about design patent law and practice using a library of more than 70,000 real documents: USPTO prosecution histories, PTAB decisions, and federal court opinions. This page explains how it does that — in enough detail that you can judge for yourself how much weight an answer deserves.
What happens when you ask a question
Every question searches the entire library. All 70,000+ documents are searched, every time. Nothing is excluded from consideration.
A smaller set is read closely. Searching finds candidates; the AI then reads the best matches in detail — typically the 20 most relevant documents for a question about the law, or 10 when you’ve also attached your own document. It reads the passages that matched your question, not the complete documents.
So the honest description is: the whole library is searched, and the best matches are read closely. If a relevant document doesn’t surface in the search, the AI never sees it.
You can narrow the search yourself. The “Search in” checkboxes let you scope a question to design patent applications, PTAB decisions, Federal Circuit opinions, statutes and the MPEP, or any combination.
What it costs, and how you know beforehand
Nothing is charged until you approve it. Before any question runs, ProSearch shows you what it will cost at most, along with your current balance. You choose whether to proceed.
You are never charged more than the amount quoted. You’re billed for what the answer actually uses, which is usually less than the quoted maximum. If a question fails, you aren’t charged at all.
Attached documents cost more than questions about the law, because there’s more material to analyze. A question about a legal issue typically runs $1–2. A question with a document attached typically runs $2–3, depending on length.
Longer questions produce longer answers, which cost more. A targeted question (“what are the drawing objections?”) costs less than an open-ended one (“how should I respond to all of this?”).
Credits never expire.
When you attach a document
You’ll be told exactly what will be analyzed before you’re charged. If your document is longer than can be reviewed in one question, you’ll see which pages will be examined visually and which won’t — before you commit.
Two different things happen to an attached document. Its text is read in full. Its pages are also looked at as images, so the AI can examine drawings and annotated figures — but only up to a limit, currently 10 pages per question.
Pages beyond that limit still have their text included. What’s lost is the visual review of those pages. For an office action, that usually means the AI can still discuss what the examiner wrote about a later figure, but hasn’t looked at the figure itself.
Office actions from Patent Center are scanned images with no text layer. ProSearch runs optical character recognition automatically to recover the text. This works well, but OCR is imperfect, and it’s why a scanned office action takes longer to process than an ordinary PDF.
For long documents, split by section. A 23-page office action can’t be fully reviewed visually in one question. Uploading the § 112 rejection and its figures as one question, and the drawing objections as another, produces better answers than one question about everything.
Practical tips. Omit the PTOL-90A transmittal cover sheet — it carries nothing an answer relies on. Upload only the pages your question concerns. If you have a text-searchable version of the document, use it; the analysis is faster and more reliable.
Limits: up to 3 files per question, PDF or TXT.
What happens to your documents
Uploaded documents are never stored. They’re held in memory only for as long as it takes to answer your question, then discarded. They are never written to disk, never saved to cloud storage, and never added to the searchable library. No other user can ever encounter them.
Your questions and answers are saved to your account history, so you can return to earlier research. The text of your uploaded documents is not saved — only the question you typed and the answer you received. If an answer quotes your document, that quoted passage is part of the saved answer.
There is currently no way to delete your own question history. If you need something removed, contact us.
Questions are processed by Anthropic’s Claude model, hosted by Microsoft on Azure infrastructure under our own account. Your material is not sent to Anthropic directly. Microsoft does not use it to train models. Under Microsoft’s standard terms, prompts and outputs may be retained for up to 30 days for abuse monitoring, accessible only to authorized Microsoft personnel.
Our deployment uses Microsoft’s global infrastructure, which means processing may occur outside the United States. Stored data remains in our resource’s US region.
If you’re evaluating this against your confidentiality obligations, those are the facts you need. We’d rather state them plainly than leave you to guess.
What it can’t do
It is not legal advice, and it is not a substitute for reading the documents. Every answer links to its sources so you can verify them. Please do.
It can miss relevant material. The search matches your question against document text. A document that addresses your issue in unusual language may not surface. Asking the same question different ways will sometimes find different material.
It reads passages, not whole documents. The AI sees the parts of each document that matched your question. Context elsewhere in a long opinion may not reach it.
It can be wrong about what it read. It quotes and characterizes documents accurately most of the time, not all of the time. The citations tell you where to check.
Its visual analysis is real but limited. It can examine drawings, read annotated figures, and discuss what an examiner marked up. It is not a substitute for your own examination of the drawings — particularly for the fine distinctions design patent practice turns on, like broken versus solid lines.
It has a knowledge cutoff. The underlying model was trained on data with an end date, and the library is updated periodically. For anything recent, check the primary source.
Technical detail — for those who want to know precisely how this works
Retrieval
Search runs against an Azure AI Search index over the document corpus, using keyword matching with semantic reranking — not vector embeddings. Semantic reranking reorders what keyword search found; it does not retrieve documents that share no vocabulary with the query.
To widen that net, questions about the law use multi-query retrieval: your question is sent to Claude, which returns three alternative keyword-dense phrasings using design patent terms of art. All four queries (yours plus three rephrasings) run against the index. Results are pooled, deduplicated by document, scored by the best result per document, and the top 20 documents are passed forward. Each contributes up to 10 highlighted passages.
If semantic ranking fails, the system retries with plain keyword search. If the rephrasing call fails, it proceeds with your original question alone. If nothing is retrieved, it says so rather than answering from general knowledge.
Questions with an attached document use a simpler retrieval path: a single query, no rephrasing, no semantic reranking, top 10 documents.
Content-type filters are applied as index filters on document filename prefixes before ranking.
Generation
Model: Claude Opus 4.8, deployed through Azure AI Foundry (Global Standard) on our own Azure resource.
System prompt: The model is instructed on an authority hierarchy — Supreme Court opinions as controlling; Federal Circuit as binding in patent matters; district court opinions as persuasive only; ITC determinations as non-binding outside the Commission; PTAB decisions as non-precedential unless designated; prosecution documents as not precedent but potentially relevant; statutes and CFR rules as binding law; the MPEP and USPTO guidance as examiner guidance without force of law. It is instructed to name the deciding tribunal, never to present persuasive material as controlling, and to flag reasoning that predates a controlling doctrinal shift — specifically that LKQ v. GM (2024) overruled the Rosen/Durling framework.
Answer length is capped at 5,000 tokens for corpus questions and 8,000 for document questions.
Citations
The model is asked to cite inline as [document|"verbatim quote"], where the quote is 5–12 consecutive words copied from that document’s retrieved excerpt.
Citations are constrained to retrieved documents. Before a citation becomes a link, the document name is checked against the list of documents actually retrieved for that question. An invented document name never becomes a link.
Quotes are not verified against source text. The quote is used to position and highlight the passage when you open the source. In testing, roughly three of four links land on the intended passage; the remainder open the document without a highlight.
If the answer cites nothing, no sources are listed — you’ll see a note saying the answer rests on your attached document rather than the library.
Document processing
Uploaded PDFs are processed for text and embedded images. When extracted text averages fewer than 200 characters per page, the file is treated as a scan and sent to Azure Document Intelligence for OCR; whichever result is longer is used.
When embedded images can’t be trusted — no extractable raster images, or decoded images that fail a brightness check — the PDF is attached natively so the model renders the pages itself. USPTO office actions from Patent Center take this path: every page is a 2550×3300 bitonal CCITT G4 scan with no text layer, and these do not decode reliably.
Rendered pages cost roughly 1,480 tokens each, which is what constrains the page limit.
Current limits per question: 3 files; 60,000 characters per file and 100,000 combined; 10 extracted images per file and 10 combined; 10 natively rendered PDF pages combined. When a limit binds, you’re told which document was affected and how.
These sit inside a hard constraint: our deployment is limited to 40,000 uncached input tokens per minute, and a single question must fit within it.
Pricing
Charges are computed from metered token counts returned by the model: input, output, cache-read, and cache-write tokens are priced separately. The rate table lives in one place and is used for both quoting and charging, so the two cannot diverge.
Uploads are quoted exactly. The complete request payload is assembled, measured with the model’s token-counting endpoint, and priced. The payload is then held server-side for 15 minutes and sent unchanged — the tokens counted are the tokens billed.
Corpus questions are quoted conservatively. Building the real payload would require a Claude rephrasing call the user might never confirm, so instead a deliberately high ceiling is quoted based on measured averages. Actual usage is charged, capped at the quote.
Billing occurs only after a complete, successful answer. Every failure path exits before the charge. If a charge fails to post, the answer is still delivered.
Data handling
Uploaded files are read into memory, processed, and discarded. No filesystem write, no blob storage, no database record of file contents. Question and answer text is stored per user in our database with no expiry and no user-facing delete.
Between the estimate and your confirmation, the assembled request is held in server memory only, bounded by count and size, cleared on use, on a 15-minute expiry, and on process restart.
AI-generated results may be incomplete or inaccurate and are not legal advice. Verify all results independently. Design Patent Pro LLC is not a law firm and does not provide legal services.