.png)
3 hours ago
2
Last year, I had to leave the database infrastructure field (if you remember some of my previous voices) and dive into RAG application development. At the time, I thought I was finally going to break free from the tedious low-level tech stack and do something “more high-level” in applications. The result? Reality gave me a resounding slap in the face.
You Can Never Escape the Infrastructure Curse
Look around: everyone’s talking about ChatGPT, talking about all sorts of fancy AI applications. Product managers are excitedly drawing various chat boxes, designers are dancing with joy over natural language interactions. It’s like everyone’s saying: Hey, look, we can finally throw away those boring technical details and just talk to AI directly!
Bullshit.
Anyone who’s actually done RAG knows this thing is just infrastructure engineering wearing an AI costume. And it’s way more complex than traditional databases. This field hasn’t fully developed yet, so we see many people trying to conflate methodology with implementation details.
Every day I wonder: did I escape old problems, or did I just earn the privilege of new ones?
RAG is Not Just Retrieval-Augmented Generation
Retrieval: Not as Simple as Just Finding Stuff
Retrieval should be the part we’re most familiar with, right? After all, we’ve been doing SQL and full-text search for decades.
Wrong.
Retrieval in RAG has mutated into a completely different beast. Now we’re dealing with proximity in semantic space, cosine similarity between embedding vectors, finding nearest neighbors in high-dimensional space.
I once thought I had mastered the art of query optimization. Now? I’m debugging why the same question gets different retrieval results in summer versus winter. Why? Because language models’ semantic understanding can be affected by time, context, and even seasonal biases in training data. What the hell?
Retrieval is no longer just about finding the right records. It’s about understanding the real intent behind questions, extracting signal from noise, finding clear answers in fuzzy semantic space.
Agent: Thinking Tools or IF-ELSE in AI Clothing?
If retrieval is RAG’s brain, then intelligent agents are its hands and feet. In the past, we built passively responsive systems; now, we’re creating ecosystems that actively think and act.
The Agent paradigm has become the new hotness. Why? Because just returning information isn’t enough anymore. Users don’t want data, they want answers; not options, they want decisions; not possibilities, they want certainty.
We orchestrate complex Agent Workflows, share tool ecosystems with MCP, create specialized thinking agents, search agents, computation agents, planning agents. These intelligent agents also need standard protocols to interact and collaborate. Projects like IBM’s beeAI are trying to solve this problem.
But honestly, most so-called “intelligent” agents are still if-else monsters under preset paths. True agent intelligence would think independently, but we haven’t really reached that step yet. We’ve just written more conditional statements in more complex ways.
Except this time, the conditional statements are dynamically generated by language models. This is what we call progress.
Grounding Truth: The Connection Point with the Real World
Ultimately, all these fancy technologies must return to a simple question: is it correct?
Grounding truth is RAG’s value anchor. If the systems we build can’t provide accurate, verifiable information consistent with the real world, then it’s no different from a randomly babbling drunk.
Human In The Loop isn’t an option, it’s a necessity. AI hallucination isn’t a technical problem, it’s a philosophical one: when there’s no human oversight, who defines what’s real?
We need to continuously accumulate grounded factual foundations, need human experts to participate continuously, need to establish feedback loops to correct errors and biases. This isn’t just about improving accuracy, but about giving these systems real value and meaning.
New Layers of Complexity
Let me be straight: RAG is ten times more complex than traditional databases.
When dealing with databases before, we at least knew what the data looked like. Now? PDFs, images, videos, audio, 3D models, even IoT sensor data - any garbage gets thrown in. What’s more fucked up is that all this heterogeneous data has to maintain semantic consistency.
When doing query optimization before, we cared about index selection and execution plans. Now? We have to consider vector similarity, semantic relevance, multimodal transformations, and worry about whether the large model will suddenly go crazy and return a bunch of nonsense.
When ensuring data consistency before, we had ACID. Now? Knowledge evolves in real-time, concepts dynamically reconstruct, even the word “consistency” itself becomes ambiguous.
My engineer brain is experiencing unprecedented challenges, and I’m actually fucking enjoying this process.
Why Am I Still in This Pit?
Why am I still enduring this complexity? Because it’s too damn interesting.
We’re not building a simple chatbot. We’re redefining how knowledge is acquired, processed, and applied:
- We’re building information systems that understand human intent
- We’re designing learning ecosystems that can self-adapt
- We’re creating intelligent agents that can reason and act
- We’re trying to connect abstract language understanding with the concrete real world
This is way more exciting than building a database, an API, or an application. This is being part of creating the next generation of computing infrastructure.
Layers are Illusion, Infrastructure is Eternal
After this year of RAG development, I finally get it: there’s no such thing as “higher-level” or “lower-level” technology. The so-called hierarchical distinction is just a placebo we give ourselves.
No matter how high you go, you’ll eventually find yourself dealing with infrastructure problems. It’s just that this time the infrastructure is more complex, more abstract, more closely related to human cognition.
When you really want to build something valuable, you ultimately have to return to infrastructure. Because that’s the foundation of everything.
So when someone asks me: “Why did you switch from databases to AI applications?”
I’ll say: “I never switched at all. I just changed the way I solve the same problems. These problems now have a better name: the trinity of retrieval, intelligent agents, and grounding truth.”
Now, excuse me, I need to deal with some damn vector quantization issues, educate a few confused intelligent agents, and make sure my text splitting strategy doesn’t tear facts to pieces. This is what real RAG work looks like.
