Your corpus is growing 10x a year. What breaks first?

As the corpus grows, the probability of finding the ‘gold’ chunk in the top-k results decreases unless k increases. Once the context window limit (e.g., 32k tokens) is reached, performance degrades. A reranker helps by keeping the generator’s context window usage efficient, while a better retriever improves the precision of the initial retrieval. Additionally, index memory requirements grow significantly, making quantization essential for scaling.