Millions of Chinese companies still run core systems on JDK 8, and when they try to add AI they hit an awkward fact—not one of the mainstream options works, not LangChain4j (the Java version of LangChain), not Spring AI. A hands-on tutorial published this week on Juejin concludes that bypassing frameworks to call big-tech SDKs directly is the only viable path right now.
What this is
The author's scenario is textbook: enterprise internal knowledge bases still rely on MySQL LIKE fuzzy matching, while the boss demands employees query them "just like using ChatGPT."
The tutorial progresses in four layers—get the API working, add caching (so high-frequency Q&A doesn't double-bill), concurrency control plus rate limiting and circuit breaking, and streaming output (typewriter effect). Each step is required for any enterprise AI rollout—none of it is cosmetic.
The key judgment hides in the framework selection comparison table: LangChain4j's older compiled artifacts require JDK 11, newer versions demand JDK 17; Spring AI has required JDK 17 since birth; Dify requires separate deployment and data leaves the perimeter (compliance won't pass). All three mainstream paths hit a wall, leaving only direct calls to Java SDKs from Alibaba, Baidu, and other big-tech vendors.
Industry view
Pragmatists will endorse this approach: direct SDK calls are flexible, don't drag in a full framework stack, and are easier to debug when things break. The tutorial's basic dialogue integration took just 40 lines of code—the barrier to entry is extremely low.
But we see a deeper problem—what's marketed as the "Java AI framework ecosystem" is essentially useless for China's installed enterprise base. Banking, government, and traditional manufacturing core systems are largely stuck on JDK 8; AI transformation can't be driven by LangChain4j or Spring AI, so every company reinvents the wheel on its own.
The counter-argument is equally valid: skipping frameworks means every enterprise must rewrite caching, rate limiting, and circuit breaking from scratch—long-term maintenance costs are severely underestimated. The tutorial's viewpoint hides architectural hidden liabilities; three to five years from now, legacy systems will become even more bloated "legacy system plus bare SDK patchwork," and far harder to clean up.
Impact on regular people
For enterprise IT: The real money in AI rollout isn't in greenfield demos—it's in retrofitting legacy systems. Next year's budgets will clearly shift from "AI innovation showcases" to the main line of "AI transformation of legacy systems."
For individual careers: Engineers who can wire AI onto legacy systems are scarcer than pure-play AI engineers who only know how to call the ChatGPT API. Solid Java fundamentals plus LLM SDK integration is the most pragmatic skill combination inside enterprises.
For consumers: These retrofits are nearly invisible to consumers, but they determine whether banking apps, government hotlines, and customer service systems actually become smarter. Honestly, most of the time they won't—at least not for the next couple of years.