PROJECT DEMO · CSE 573

Semantic
Patent
Analysis

Retrieval System Using NLP and RAG

Group 21 · Members

Ananth Sundararajan Sekar

Ann Liu

John T.S Gondo

Krishnabalaji Budhiraju

Melisa Torres Mejia

Vikram Vinod

Group 21

Ananth Sundararajan Sekar · Ann Liu · John T.S Gondo · Krishnabalaji Budhiraju
· Melisa Torres Mejia · Vikram Vinod

Problem Definition

Patent Analysis
Has Exceeded
Human Capability

The Scale Problem

Over 10 million active patents are currently in force, with thousands of new filings submitted every week. No human team can read, compare, and analyze this corpus at speed.

Current Limitations

Keyword-based search cannot capture semantic meaning.
Different wording for the same innovation is routinely missed.
Rigid classification systems lack cross-domain flexibility.

Downstream Impact

Missed Prior Art Semantically equivalent patents go undetected.
Incomplete Assessment Lack of tools for thorough landscape analysis.
High Barrier Hard for non-experts to navigate the space.

System Architecture

End-to-end pipeline from raw patents to conversational AI — fully operational

Offline Pipeline

Data Ingestion

Google Patents via SerpAPI. 1000+ text patent documents stored. Key tech areas: AI, Biotech, Telecom.

Preprocessing

Cleaning, tokenization, and normalization of raw patent text. Remove special characters.

Embedding (SBERT)

Each patent encoded as a dense 768-d semantic vector using Sentence-BERT (all-mpnet-base-v2).

Clustering & Index

Patents grouped via HDBSCAN; FAISS vector index built. UMAP used for 2D dimensionality reduction.

Online Pipeline

1. Query Embed

User question encoded into the same 768-d vector space.

2. FAISS Retrieval

Top-k most similar patents retrieved in milliseconds.

3. RAG Response

GPT generates a grounded, evidence-backed conversational answer.

RAG System Diagram

System Architecture

Algorithms

End-to-end pipeline from raw patents to conversational AI — fully operational.

01

Data Collection

Google Patents via SerpAPI • 50 patents • 5 domains • 2015–2024

02

Preprocessing

spaCy NLP • tokenization • lemmatization • boilerplate removal

03

SBERT Embeddings

all-mpnet-base-v2 • 768-dim vectors • doc-level + 2,928 RAG chunks

04

UMAP + HDBSCAN

Dimensionality reduction + density clustering • ARI=1.0 • Purity=1.0

05

KeyBERT Labels

Auto-generated cluster names • TF-IDF + KeyBERT • 5 clusters labeled

06

FastAPI Backend

4 REST endpoints • /clusters • /novelty • /rag/chat • /health

07

React Frontend

Patent Intelligence Dashboard • Clustering • Novelty • RAG Chat tabs

DATA SET

Sources & Scope

Primary Source

Google Patents via SerpAPI. Full patent text including detailed descriptions, claims, and background sections.

Corpus Size

500–1,000 documents stored as parsable text files. All patents are U.S.-issued, spanning 2015–2024.

Tech Focus Areas

AI & Machine Learning Deep learning architectures, model optimization, and inference systems
Semiconductor Design Advanced node processes and chip architectures
Renewable Energy Solar, wind, and energy storage technologies
Biotech & Gene Therapy CRISPR, mRNA delivery, and gene-editing platforms
Telecommunications & 5G Millimeter wave, MIMO, and network slicing

Collection Strategy

Phase 1

Metadata Collection

Weeks 1-2

Query SerpAPI using technology-specific search terms. Rate-limited retrieval.

Phase 2

Full Text Collection

Weeks 2-3

Extract IDs and query Google Patents API. Implement exponential backoff.

Phase 3

Data Validation

Week 3

Verify completeness, remove duplicates, and create a data catalog.

Clustering Results — SOTA

UMAP + HDBSCAN on SBERT

Perfect separation across all 5 technology domains

1.000

Adjusted Rand Index

1.000

Normalized Mutual Info

1.000

Cluster Purity

0.795

Silhouette Score

Baseline vs SOTA Performance

Source: mindframes.app

5 Technology Clusters

C0 Semiconductor

C1 Biotechnology

C2 Renewable Energy

C3 AI & Machine Learning

C4 Telecom 5G

0 noise points · 10 patents per cluster · Raw HDBSCAN noise reassigned

NOVELTY SCORING — LIVE DEMO

Score a candidate patent against 50 prior-art documents using cosine similarity

Formula

\text{Novelty Score} = 1 - \text{max\_similarity}(\text{candidate}, \text{corpus})

Solar Energy + AI

0.823

Solar panel efficiency improvement using AI

Relatively novel

Nearest: CN112054588B (Wind-solar hybrid)

CRISPR + AI Combo

0.741

A new method for CRISPR gene editing using AI optimization

Relatively novel

Nearest: AU2015101792A4 (CRISPR systems)

Endpoint: POST /api/novelty/score · Frontend: Novelty Scoring tab · Indexed: 50 patents

Corpus Statistics

Avg Novelty 0.280

Median 0.232

Min near-duplicate

0.010

Max very novel

0.633

UI & Data Visualization (1/2)

Patent Intelligence Dashboard

UI & Data Visualization (2/2)

System Capabilities &
Evaluation Metrics

Development Roadmap

Project
Plan

End-to-end pipeline from raw patents to conversational AI.
9-week execution strategy.

Week 1-2 Feb 18 - Mar 3

Data Collection & Preprocessing

Lead: John

Deliverables: Raw dataset, cleaned data, preprocessing pipeline.

Week 3-4 Mar 4 - Mar 17

Embedding Generation

Lead: Krishna Balaji

Deliverables: Document & chunk embeddings, embedding analysis.

Week 5 Mar 18 - Mar 24

Clustering & Labeling

Lead: Ananth

Deliverables: Cluster assignments, quality metrics, labels.

Week 6 Mar 25 - Mar 31

RAG Development

Lead: Krishna Balaji

Deliverables: FAISS index, RAG pipeline, query interface.

Week 7 Apr 1 - Apr 7

Novelty & Visualization

Lead: Vikram, Ananth

Deliverables: Similarity scoring, Streamlit dashboard.

Week 8 Apr 8 - Apr 14

Testing & Evaluation

Lead: Melisa

Deliverables: Test results, evaluation report, user study.

Week 9 Apr 15 - Apr 21

Documentation & Report

Lead: Ann

Deliverables: Final report, presentation, documentation.

Demo

Live Demonstration

Transitioning to the live system demo.

Thank You!

Any Questions?

We welcome feedback. Reach out to any member of Group 21 for further discussion on methodology, evaluation design, or potential collaboration.

Contact

CSE 573 · Spring 2026
Project 9 · Group 21

Repository

Full codebase, data catalog, and technical report available upon project completion.
github.com/ananthsekar007/cse573-semantic-project

Methods Recap

SBERT HDBSCAN UMAP FAISS RAG KeyBERT

Semantic Patent Analysis

Patent Analysis Has Exceeded Human Capability

The Scale Problem

Current Limitations

Downstream Impact

System Architecture

Offline Pipeline

Data Ingestion

Preprocessing

Embedding (SBERT)

Clustering & Index

Online Pipeline

1. Query Embed

2. FAISS Retrieval

3. RAG Response

RAG System Diagram

System Architecture

Algorithms

Data Collection

Preprocessing

SBERT Embeddings

UMAP + HDBSCAN

KeyBERT Labels

FastAPI Backend

React Frontend

DATA SET

Sources & Scope

Primary Source

Corpus Size

Tech Focus Areas

Collection Strategy

Clustering Results — SOTA

UMAP + HDBSCAN on SBERT

Baseline vs SOTA Performance

5 Technology Clusters

NOVELTY SCORING — LIVE DEMO

Solar Energy + AI

CRISPR + AI Combo

Corpus Statistics

Patent Intelligence Dashboard

UI & Data Visualization (2/2)

System Capabilities & Evaluation Metrics

Development Roadmap

ProjectPlan

Data Collection & Preprocessing

Embedding Generation

Clustering & Labeling

RAG Development

Novelty & Visualization

Testing & Evaluation

Documentation & Report

Demo

Thank You!

Any Questions?

Contact

Repository

Methods Recap

Semantic
Patent
Analysis

Patent Analysis
Has Exceeded
Human Capability

System Capabilities &
Evaluation Metrics

Project
Plan