case-study-garden-update

qdrant-landing/content/blog/case-study-garden-intel.md

@@ -19,21 +19,23 @@ tags:
 
 ## Garden Accelerates Patent Intelligence with Qdrant’s Filterable Vector Search
 
+![How Garden Unlocked AI Patent Analysis](/blog/case-study-garden/case-study-garden-bento-dark.jpg)
+
 For more than a century, patent litigation has been a slow, people-powered business. Analysts read page after page—sometimes tens of thousands of pages—hunting for the smoking-gun paragraph that proves infringement or invalidity. Garden, a New York-based startup, set out to change that by applying large-scale AI to the entire global patent corpus—more than 200 million patents—in conjunction with terabytes of real world data.
 
-*“Our customers need to compare millions of possible patent–product pairings in seconds, not days,” explains co-founder Justin Mack. “That means vector search that can handle huge data sets **and** surgical-grade filtering.”*
+*“Our customers need to compare millions of possible patent–product pairings in seconds, not days,” explains co-founder Justin Mack. “That means vector search that can handle huge data sets and surgical-grade filtering.”*
 
 ### A data set that breaks naïve vector search
 
 Each patent can run to 100+ pages and, thanks to decades of revisions, carries roughly        2,000 metadata fields: jurisdiction, grant date, family ID, claim dependencies, and so on. Garden splits every patent into semantically meaningful chunks, producing “many hundreds of millions” of vectors. The same pipeline ingests real-world product data to compare against the patents.
 
-The engineering demands quickly outgrew Garden’s first solution, a fully-managed vector service. They had tens of gigabytes of data already costing **≈ $5,000 / month**. And a lack of native filterable-HNSW meant that Garden had to stand up a separate index for **every** combination of country, date range, and technology tag. Finally, with no infrastructure visibility, troubleshooting was slow and expensive.
+The engineering demands quickly outgrew Garden’s first solution, a fully-managed vector service. They had tens of gigabytes of data already costing ≈ $5,000 / month. And a lack of native filterable-HNSW meant that Garden had to stand up a separate index for every combination of country, date range, and technology tag. Finally, with no infrastructure visibility, troubleshooting was slow and expensive.
 
 A second migration to a self-hosted open-source alternative cut costs but introduced new pains: on-call operations for a two-person team, upgrades during business hours, and—crucially—the same filtering limitations.
 
 ### Discovering Qdrant
 
-When Garden found Qdrant’s blog post on **filterable HNSW**, the team realized they could get the search semantics they wanted without bolting on bespoke sharding logic.
+When Garden found Qdrant’s blog post on filterable HNSW, the team realized they could get the search semantics they wanted without bolting on bespoke sharding logic.
 
 “Filterable HNSW was the deal-maker, but Qdrant Cloud’s *managed* Rust backbone sealed it,” says Mack. “We kept source-level transparency while off-loading 24×7 ops.”
 
@@ -41,7 +43,7 @@ When Garden found Qdrant’s blog post on **filterable HNSW**, the team realized
 
 * **SLA-backed sub-100ms latency** meets Garden’s product target even when a user fires off thousands of queries in a single button-click.
 
-* **Pay-for-what-you-use pricing** lets Garden store **10× more data for roughly the same cost** it once paid for a fraction of the corpus.
+* **Pay-for-what-you-use pricing** lets Garden store 10× more data for roughly the same cost it once paid for a fraction of the corpus.
 
 ### Migration in practice