PatentOS™
Platform architecture

Prior Art Search Architecture

Thirteen stages that take an invention from a raw draft to an attorney-ready prior art report — understanding the invention first, searching the world's patent and technical literature in parallel, then reasoning about every reference the way an examiner would.

Back to Prior Art Search

The pipeline

Each stage feeds the next. Retrieval widens through stages 1–6, then narrows through examiner-grade reasoning in stages 7–10, and resolves into decision-ready output in stages 11–13.

  1. Stage 1

    Invention Understanding

    Instead of asking users for keywords, PatentOS first understands the invention. It accepts a draft patent, claims, a technical specification, a PDF, a Word document, drawings read through OCR and Vision, or a simple natural language description.

    Input

    Draft patentClaimsTechnical specificationPDFWord documentDrawings (OCR + Vision)Natural language description

    AI automatically extracts

    Technical fieldNovel featuresIndependent claimsDependent claimsTechnical componentsAdvantagesProblem being solvedSolutionAlternative embodiments

    Modern research shows that natural-language understanding produces better search queries than relying only on manually chosen keywords. ScienceDirect

  2. Stage 2

    AI Claim Decomposition

    A claim such as "autonomous drone inspection system using thermal camera and LiDAR" is automatically broken into functional blocks, and each block is expanded into the vocabulary a search actually needs.

    Functional blocks

    DroneNavigationObstacle avoidanceThermal imagingLiDARInspectionAI defect detectionCloud reporting

    Then generates

    SynonymsCPC classesIPC classesSemantic conceptsEngineering terminologyAbbreviationsIndustry vocabulary

  3. Stage 3

    Multi-Agent Query Generation

    Instead of one search query, PatentOS generates hundreds. Specialized agents attack the same invention from different angles, which dramatically increases recall.

    Agent 1

    Keyword Search

    thermal inspection drone

    Agent 2

    Boolean

    (drone OR UAV) AND thermal AND inspection

    Agent 3

    Semantic

    Autonomous aerial system detecting infrastructure defects using thermal sensing

    Agent 4

    Patent Examiner

    Search as a USPTO examiner

    Agent 5

    Broad Search

    Aircraft, robot, UAV, flying inspection

    Agent 6

    Cross-domain

    Robotics, automotive, space, defense, mining, agriculture

  4. Stage 4

    Global Search Sources

    PatentOS searches patent offices, scientific literature and technical sources simultaneously, rather than treating anything outside patent databases as an afterthought.

    Patent databases

    USPTOEPOWIPOGoogle PatentsJPOCNIPAKIPOUKIPOIP AustraliaCanadian PatentsIndian Patent OfficeRussian Patent OfficeBrazilMexico

    Non-patent literature

    Scientific papersIEEESpringerNatureScienceDirectPubMedUniversity repositories

    Technical sources

    GitHubTechnical standardsISOIECRFCWhite papersOpen-source projectsConference proceedingsDatasheets

    Searching beyond patents is consistently recommended to improve coverage and reduce the chance of missing relevant prior art. Patently

  5. Stage 5

    Hybrid Search Engine

    Never rely only on embeddings. Six retrieval signals run together, so a reference that one method would miss is still caught by another.

    BM25

    Classic keyword retrieval

    Dense embeddings

    Semantic similarity

    CPC classification search

    Patent classifications

    Citation graph search

    Backward citations, forward citations, patent families

    Similar claims search

    Independent and dependent claims

    Similar figures

    Future vision AI

    Research consistently finds that hybrid retrieval (BM25 plus semantic reranking) outperforms either method alone for prior-art search. ScienceDirect

  6. Stage 6

    Vector Database

    Qdrant indexes the full document surface of every reference. Instead of collapsing a patent into one vector, PatentOS stores several embeddings per document so that a claim can match a claim and a problem can match a problem.

    Indexed

    TitleAbstractClaimsDescriptionParagraph embeddingsFigure captionsCPCIPCAssigneeInventorCitations

    Multiple embeddings

    Abstract embeddingClaims embeddingParagraph embeddingsTechnical feature embeddingsProblem embeddingSolution embedding

  7. Stage 7

    AI Reranking

    A wide candidate set is narrowed by a cross-encoder reranker, then by LLM reasoning and novelty scoring, until only the references worth an attorney's attention remain.

    Reranking flow

    Top 1000 candidatesCross-encoder rerankerLLM reasoningNovelty scoringTop 100

    Research indicates reranking substantially improves relevance after the initial retrieval stage. arXiv

  8. Stage 8

    Patent Examiner AI

    PatentOS imitates examiner reasoning. For every patent, the AI answers which features are disclosed and which are missing, then scores novelty, inventive step and legal risk.

    Does this disclose

    • Feature A
    • Feature B
    • Feature C
    • Feature DMissing

    Novelty score

    91%

    Inventive step

    Moderate

    Legal risk

    High

  9. Stage 9

    Claim Mapping

    PatentOS automatically maps every claim element to the exact prior art evidence that discloses it — down to the paragraph, the claim or the page. This becomes one of the strongest features for patent attorneys.

    Example claim mapping: claim elements and their prior art evidence
    Claim elementPrior art evidence
    Feature 1Patent X paragraph 12
    Feature 2Patent Y claim 5
    Feature 3IEEE paper page 8
    Feature 4No evidence found

  10. Stage 10

    AI Debate

    PatentOS launches several specialized AI agents that argue the invention from opposing sides, so weaknesses surface before an examiner or an opponent finds them.

    Examiner Agent

    Attempts to reject the claim.

    Inventor Agent

    Defends novelty.

    Patent Attorney Agent

    Suggests claim amendments.

    Litigation Agent

    Identifies invalidity risks.

    This adversarial, multi-agent approach reflects emerging research directions for improving patent analysis. DOI.org

  11. Stage 11

    Risk Dashboard

    The evidence collapses into the metrics an IP team actually decides on.

    Novelty scoreInventive step probabilityPrior art densityTechnology maturityFiling riskLitigation riskCompetitive overlapWhite-space opportunities

  12. Stage 12

    Interactive Patent Landscape

    The same corpus becomes a map — who is filing, where the clusters are, and where nobody has filed yet.

    CompetitorsPatent clustersTechnology evolutionCitation networksFiling trendsGeographic coverageWhite-space areas

  13. Stage 13

    Automatic Prior Art Report

    Everything above is generated as a professional report, exportable to PDF, DOCX or Excel.

    Report includes

    • Executive summary
    • Search methodology
    • Databases searched
    • Search queries
    • Relevant patents
    • Relevant non-patent literature
    • Claim-by-claim analysis
    • Novelty assessment
    • Inventive step analysis
    • Recommended claim amendments

AI technology stack

The components behind the pipeline, end to end.

LLM

Qwen3 72B, aligned with the current PatentOS architecture

Embeddings

Patent-specialized embedding model plus a general-purpose embedding model

Vector database

Qdrant

Keyword search

Elasticsearch/OpenSearch (BM25)

OCR

PaddleOCR

Vision

Multimodal model for figures and diagrams

Backend

FastAPI + Python

Database

PostgreSQL + Supabase

Inference

vLLM

Frontend

Next.js + React

Long-term vision

An AI Patent Examiner, not an AI patent search

Rather than positioning PatentOS as “AI Patent Search,” position it as an AI Patent Examiner. The platform would understand inventions, search across worldwide patent and technical literature, explain why each reference matters, map evidence directly to claim elements, estimate novelty and inventive-step risks, recommend claim amendments, and produce attorney-ready reports.

  • Understand inventions
  • Search across worldwide patent and technical literature
  • Explain why each reference matters
  • Map evidence directly to claim elements
  • Estimate novelty and inventive-step risks
  • Recommend claim amendments
  • Produce attorney-ready reports

This approach aligns with current best practices in AI-assisted prior-art search while providing a differentiated workflow for enterprise IP teams. Lexology

See the pipeline run on your invention

Bring a draft, a claim set or a plain description — PatentOS takes it through all thirteen stages and returns an attorney-ready prior art report.