The Technical Context
I dug into the Stanford paper itself after seeing another claim that "RAG works up to 10,000 vectors but becomes garbage after a million." I immediately hit the brakes: the paper makes no such empirical claim.
The paper everyone cites is Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools. It examines legal RAG tools like Westlaw and LexisNexis, comparing them against GPT-4 on legal tasks to see how often they invent facts, cite wrong sources, or distort conclusions.
The findings are sobering but useful: retrieval reduces hallucinations compared to a base model, but it doesn't eliminate them. Depending on the tool, error rates are still significant, ranging from 17% to 33%.
However, I found no mention of a "10k vector limit" or "1M noise threshold" in the paper. There are no charts on index size, no precision@k benchmarks on large collections, and no analysis of retrieval degradation as the corpus grows. As an engineer, I wouldn't cite this as proven fact.
But the idea isn't baseless. In production, I've repeatedly seen poorly designed vector search pull in semi-relevant junk due to weak embeddings, incorrect chunking, flawed metadata filters, or an overly generous top-k. People then blame the model when the real issue is the AI architecture and retrieval layer.
In short, Stanford didn't show a 10k vector limit. It revealed something more practical: even an expensive legal RAG system doesn't give you a free pass to skip fact-checking or assume the reliability problem is solved.
What This Means for Business and Automation
For teams building AI automation on documents, the takeaway is simple: the "just add RAG" approach to ensure accuracy doesn't work. This is especially true in legal, compliance, policy, and internal knowledge bases.
The winners are those who engineer retrieval as a complete system: hybrid search, strict filtering, reranking, context limiting, and source tracing. The losers are those who just dump millions of chunks into a vector database and hope for magic.
At Nahornyi AI Lab, we specialize in solving these bottlenecks. We determine where RAG is appropriate, where a knowledge graph is better, or where a different AI integration layer is cheaper and more reliable than endlessly pumping vectors. If your document search is noisy or generates untrustworthy answers, let's review your architecture and build a robust AI solution without fragile assumptions.