Full analysis vs. search (RAG) - why fragments cannot explain a system

RAG (retrieval-augmented generation) finds fragments similar to the question and hands them to the AI. It works for document search, but it has a structural limit when it comes to understanding a system. A system's meaning is not in the fragments. It is in the connections between them.

"We tried RAG and it wasn't great." This is the thing we hear most often in enterprise meetings. The tool is not bad. The kind of problem is different.

The number of connections grows far faster than the number of fragments

If code contains n symbols (functions, classes, tables), the potential connections between them grow close to n². Ten thousand symbols means ten thousand pieces but tens of millions of connections. Search picks up a few pieces; it cannot retrieve the connections. Yet the answers to day-to-day questions like "what breaks if I change this?" and "how is this amount calculated?" all live in the connections.

Break a link in the chain and the answer is wrong

A settlement amount starts in payment module A, passes through fee calculation B, and ends in settlement batch job C. Ask "how is settlement calculated?" and search returns A and C, the ones containing the word "settlement." B has no such word, so it drops out. Seeing only A and C, the AI invents a plausible B. This is what enterprise hallucination actually is. Not because the model is stupid, but because the middle of the chain is missing.

The same question gets different answers

Search picks up different fragments depending on how the question is phrased. Ask today and get this answer, phrase it differently tomorrow and get that one. To use it as a standard for your work, the same state has to produce the same answer. An answer you cannot reproduce will not survive an audit, a review, or automation.

How Platty does it. Lookup, not search

Platty decomposes the entire codebase into structure before any question arrives. A function call is a real connection, not "similar text," so static analysis can link all of them. The A→B→C chain is in the graph in its entirety from the start. When a question comes in, Platty does not search; it follows the graph. Nothing in the middle drops out, and the same code yields the same answer.

To put it plainly: RAG is the right tool for "find me similar documents," while understanding a system is a problem of "knowing every connection." Different problems need different tools.

Next : How do you trust the AI's answer - evidence, gaps, confirmed and inferred