I Built a Free Patent Search Engine with 3.5M US Patents — No Login, Powered by SQLite FTS5

I'm a patent lawyer who started coding in December 2025. Today I'm launching a free patent search engine covering 3.5 million US patents (2016–2025).

Try it here: patentllm.org

No login. No ads. No API keys.

Why I built this

Existing patent search tools fall into two camps:

• Google Patents — simple but no filtering, no relevance tuning
• Commercial tools (PatSnap, Derwent) — powerful but $10K+/year

I wanted something in between. Fast full-text search across titles, abstracts, and claims — with CPC classification filtering and smart relevance ranking — for free.

The tech stack (intentionally boring)

Component	Choice	Why
Database	SQLite (74GB, single file)	Zero ops. No Postgres. No containers.
Full-text search	FTS5 + BM25	Built into SQLite. Handles 3.5M records without breaking a sweat.
Ranking	70% BM25 + 30% recency	Newer patents get a boost. Patent relevance isn't just keyword matching.
Web framework	FastAPI + Jinja2	Server-side rendering for SEO. Every search result has a unique URL.
LLM (tagging)	Nemotron 9B on RTX 5090	Classified all 3.5M patents into 100 tech tags. Locally, zero API cost.
Hosting	Chromebook + Cloudflare Tunnel	Yes, really.

Why FTS5 instead of vector search?

Everyone's doing vector search with embeddings these days. I went the opposite direction.

For patent search, exact phrase matching matters. When a patent attorney searches for "solid-state battery electrolyte," they need documents containing those exact terms — not semantically similar documents about "energy storage solutions."

FTS5 gives me:

• Exact phrase matching with quotes
• Boolean operators (AND, OR, NOT)
• BM25 relevance scoring out of the box
• Sub-second queries on 3.5M records
• Zero external dependencies

I added a local LLM (Nemotron 9B) on top for query expansion — it converts natural language like "self-driving car obstacle detection" into FTS5 queries like "autonomous driving" AND "obstacle detection". Best of both worlds.

BM25 weight tuning

Default BM25 weights treat all fields equally. That doesn't work for patents. I tuned the weights based on how patent attorneys actually search:

• Title: 10.0 — if your keyword is in the title, it's almost certainly relevant
• Assignee: 5.0 — searching by company name should rank those patents highest
• Abstract: 3.0 — the abstract is the patent's elevator pitch
• Claims: 1.0 — claims are legally important but verbose; lower weight prevents noise

This produces results that match how an experienced patent searcher would rank them manually.

The data pipeline

All data comes from USPTO PatentsView (public domain, Creative Commons):

1. Download bulk TSV files (titles, abstracts, claims, CPC classifications)
2. Parse and merge into SQLite
3. Build FTS5 virtual table across title + abstract + claims
4. Run Nemotron 9B over all 3.5M records for tech tag classification (took ~48 hours on RTX 5090)
5. Serve via FastAPI

The entire database is a single 74GB SQLite file. Backup is cp. Migration is scp.

What you can do with it

• Keyword search across 3.5M patents with BM25 ranking
• Filter by CPC section (Physics, Electricity, Chemistry, etc.) with group drill-down
• Filter by assignee (Apple, Samsung, etc.)
• Filter by date range
• View patent details — abstract, claims, tech tags, CPC classification
• Link to Google Patents for full text and legal status

What's next

I'm building an AI analysis layer on top (prior art search, FTO analysis, competitive landscape) — that will be the paid tier at ai.patentllm.org.

The search will stay free forever.

Feedback welcome

I'd love to hear from:

• Patent attorneys / IP professionals — does the ranking feel right?
• Engineers interested in patent data — what features would you want?
• Anyone who's built search at scale with SQLite — tips welcome

patentllm.org — give it a try and let me know what you think.