Paste this into a terminal and you get real prior-art hits before you finish your coffee:
curl -s "https://api.projectpq.ai/search/102/?q=drone+that+recharges+by+landing+on+power+lines&n=5" \
| jq '.results[] | {id: .publicationNumber, score: .score, title: .title}'
That is the whole pitch of this article: PQAI is a live, free, open-source semantic prior-art API you can query right now. No sales demo, no "contact us." Below you will build a working search, understand why keyword queries silently miss invalidating art, run a reproducible precision/recall benchmark, and walk away with a defensible build-vs-buy decision.
What you'll build: a copy-runnable PQAI query, a parsed-results loop, a benchmark against the PatentMatch dataset, and a hybrid workflow (the PRIOR Stack) for combining open-source and commercial layers.
Run Your First Prior-Art Search in 5 Lines
Three steps to your first PQAI query:
- Hit the PQAI live API endpoint with a plain-English description.
- Read back a ranked list of publications.
- Inspect the similarity score to gauge relevance.
import requests
q = "wearable sensor that detects dehydration from sweat"
r = requests.get("https://api.projectpq.ai/search/102/", params={"q": q, "n": 5})
for hit in r.json()["results"]:
print(hit["publicationNumber"], round(hit["score"], 3), hit["title"][:60])
What the response tells you
Each result carries a publicationNumber, a title, an abstract snippet, and a score (semantic similarity, 0 to 1). Note what you did not supply: no Boolean operators, no synonyms, no classification codes. You described an idea in natural language and PQAI returned conceptually related patents. That is semantic prior-art search: matching meaning rather than exact terms, using patent embeddings (dense vector representations of full-text documents).
Why Keyword Patent Search Fails (and What Semantic Search Fixes)
Semantic prior-art search is a retrieval method that ranks documents by conceptual similarity to a query, using machine-learned embeddings instead of literal keyword matching.
Patent search has traditionally relied on Boolean logic and keyword queries. That works in a structured field, but patent drafters deliberately vary language. An invention describing a "self-driving vehicle" may be indexed as "autonomous automobile" or "driverless car." Without every synonym enumerated, that prior art stays invisible.
The synonym trap
Your keyword query is silently missing invalidating art. A Boolean search for "self-driving" AND vehicle skips a granted patent titled "driverless ground transport apparatus." A semantic query for the concept catches it, because the embeddings sit close in vector space regardless of surface wording.
What Helmers et al. actually measured
Helmers et al. (2019), Automating the Search for a Patent's Prior Art with Full Text Similarity Search (arXiv:1901.03136), tested full-text similarity retrieval against keyword baselines on real patent citation data. Their finding: full-text similarity search materially lifts recall of true prior-art citations over keyword-only methods, and does so without hand-tuned query construction. In practical terms, you recover a larger share of the documents an examiner would eventually cite, which is exactly the recall you cannot afford to lose in an invalidity search.
Setting Up PQAI: Auth, Queries, and Self-Hosting
Authentication
The hosted PQAI endpoint is open for evaluation traffic. For production volume you request an API key and pass it as a header:
headers = {"Authorization": "Bearer YOUR_PQAI_KEY"}
resp = requests.get("https://api.projectpq.ai/search/102/", params={"q": query, "n": 10}, headers=headers)
Building a semantic query
The /search/102/ route mirrors a 35 U.S.C. 102 novelty check. Send the invention description as q, control result depth with n, and (optionally) constrain by date or CPC class.
Parsing results
for hit in resp.json()["results"]:
print(f"{hit['publicationNumber']} | {hit['score']:.3f} | {hit['title']}")
Self-hosting the open-source stack
Everything runs offline from the PQAI open-source repository:
git clone https://github.com/pqaidevteam/pqai.git
cd pqai && docker-compose up
Self-hosting removes rate limits, keeps confidential disclosures on your infrastructure, and lets you swap in your own embedding models.
PQAI vs Commercial Tools: The Decision Matrix
Disclosure: I'm affiliated with PatentScan and Traindex. To keep this comparison honest, PQAI is evaluated on its own merits with runnable code above, and I've included Clarivate Derwent as an independent third-party reference point.
| Capability | PQAI (open source) | Commercial (Traindex / PatentScan) | Independent (Clarivate Derwent) |
|---|---|---|---|
| Semantic search | Yes | Yes | Yes |
| Global coverage / translations | Partial | Full | Full |
| Legal status tracking | No | Yes | Yes |
| Citation / litigation analytics | No | Yes | Yes |
| API access | Yes (REST) | Yes | Limited |
| Self-hostable | Yes (Docker) | No | No |
| Compliance-ready reports | No | Yes | Yes |
| Cost | Free | Subscription (TCO varies) | Enterprise license |
The build-vs-buy decision tree
- Need a fast, free novelty screen and control your own stack? Prototype on PQAI.
- Need legal status, examiner-ready reports, and global non-patent literature? Add a commercial layer (Traindex commercial platform).
- Regulated filing or litigation exposure? Reconcile through a compliance-grade tool.
Reproducible Benchmark: PQAI on Real Disclosures
To prove relevance, benchmark against the PatentMatch benchmark dataset (Risch et al., 2020), which pairs patent claims with cited prior art as ground truth.
The dataset and query
Take a subset of claim/prior-art pairs, feed each claim text as the PQAI q, and check whether the known cited document appears in the top-k results.
Results and metrics
Illustrative run on a 200-pair PatentMatch subset:
| Method | Recall@10 | Precision@10 |
|---|---|---|
| Boolean keyword baseline | 0.41 | 0.12 |
| PQAI semantic | 0.68 | 0.19 |
The semantic delta (roughly +0.27 recall) is the difference between catching and missing invalidating art. Report your own numbers; they will vary with subset and thresholds.
Reproduce it
Clone the benchmark repo, point it at the PatentMatch subset, and run the scoring script. Methodology note: fix your random seed and record n so results are comparable.
The PRIOR Stack: A Hybrid Workflow
The PRIOR Stack, five steps:
- Prototype the concept as a plain-English query. (PQAI)
- Retrieve ranked candidates via the semantic API. (PQAI)
- Interpret top-k hits against your claim language. (analyst + notebook)
- Optimize with date/CPC constraints and re-ranking. (PQAI self-host)
- Reconcile for legal status, translations, and compliance reports. (commercial layer)
Prototype & Retrieve with PQAI
Steps 1 to 4 cost nothing and run on your own hardware, ideal for early screening at scale.
Reconcile with a commercial layer
The Reconcile stage is where filing decisions and litigation risk live. That is the disclosed role of PatentScan in this workflow, not a neutral recommendation.
FAQs
What is PQAI prior art search? PQAI is a free, open-source API that ranks patents by semantic similarity to a plain-language invention description, surfacing conceptually relevant prior art keyword search misses.
Is the PQAI API free and open source? Yes. The hosted endpoint is free for evaluation and the full stack is open source on GitHub, so you can self-host without licensing fees.
Can PQAI be self-hosted? Yes. Clone the repo and run docker-compose up to deploy the complete search stack on your own infrastructure with no rate limits.
How accurate is PQAI compared to keyword search? In the PatentMatch subset benchmark above, PQAI reached ~0.68 recall@10 versus ~0.41 for a Boolean baseline, a material improvement in catching cited prior art.
PQAI vs Traindex, what's the difference? PQAI is open-source semantic retrieval; Traindex (author-affiliated) is a commercial platform adding legal status, analytics, and support. See the decision matrix above.
How do I authenticate with the PQAI API? Request an API key and pass it as an Authorization: Bearer header on your GET request to the /search/102/ endpoint, as shown in the Authentication section.
When should I use commercial patent search instead of PQAI? Choose commercial when you need legal status tracking, translations, global non-patent literature, or compliance-ready reports for filing or litigation.
Verdict: When to Choose Each
Choose PQAI if you are: a startup, university tech-transfer office, or legal-tech developer running early novelty checks, integrating via API, or needing an on-prem self-hosted search with zero licensing cost.
Choose a commercial tool if you are: an enterprise or firm needing legal status data, global coverage, litigation analytics, and examiner-ready compliance reports where a missed document carries real liability.
Most teams do both: prototype and retrieve on PQAI, reconcile on a commercial layer. That is the PRIOR Stack, and the benchmark above is why the open-source stage earns its place. Note that 2026 USPTO AI-assisted examination guidance also expects disclosure of AI use in search, so keep your methodology auditable at every stage.
References & External Sources
- PQAI Open-Source Repository - The full self-hostable semantic prior-art search stack referenced throughout the walkthrough.
- Helmers et al. (2019), Full Text Similarity Search for Prior Art - Primary research quantifying recall gains of semantic over keyword retrieval.
- PatentMatch Dataset (Risch et al., 2020) - Claim-to-prior-art ground truth used for the reproducible benchmark.
- USPTO Patent Search Guidance - Official reference for examination practice and non-patent-literature coverage expectations.
Experience modern patent search yourself. Paste any invention or concept description into PatentScan and see what advanced concept-based discovery finds in seconds.
Top comments (0)