A KAIST School of Computing team said its CONDA index kept search accuracy stable while documents were continuously added and removed, addressing a failure mode that hits Korean banks and retailers retraining retrieval-augmented generation systems on weekly policy PDFs. The university announced the results Oct. 5 ahead of commercial integration into startup GraphAI’s AkasicDB product.

Problem CONDA targets

Vector databases underpin corporate chatbots that must cite HR manuals, loan guides, and product specs. When corpora churn, static indexes misroute queries even though the correct paragraph still exists somewhere in storage. Professor Min-Soo Kim’s group built a connectivity-aware dynamic index that preserves search paths as graphs evolve.

In benchmark runs with ongoing inserts and deletes, CONDA improved recall up to 24.5 percent versus prior state-of-the-art methods and raised update throughput as much as 1.9 times, according to the peer-reviewed paper presented at VLDB 2026.

Why Korean adopters care

Domestic conglomerates deploying generative AI under the AI Basic Act must document data sources and maintain audit trails. If retrieval layers drift silently, compliance teams cannot trust citations shown to customers. CONDA’s pitch is operational: fewer full reindex jobs during business hours.

GraphAI, founded by Kim, plans to ship the index in AkasicDB’s fourth-quarter release, giving enterprises a locally supported option instead of patching open-source vector stores manually.

Technical mechanism in plain terms

Traditional approximate nearest-neighbor indexes treat deletions as lazy tombstones, which warps cluster boundaries. CONDA tracks how nodes connect and rewires neighborhoods when edges break, reducing the chance that a query wanders into an obsolete partition. The approach trades a modest memory overhead for predictable latency during heavy churn.

First author Darae Lee, a recent master’s graduate, tested workloads modeled on ticketing systems and news archives; industry reviewers asked for finance-specific traces in the commercial build.

Limits and competition

CONDA does not solve hallucination inside the generator model; it only raises the odds the right chunk is retrieved. Hybrid search with keyword filters and metadata guards remains necessary for regulated answers.

Global cloud vendors market managed vector services with auto-scaling; Korea’s selling point is on-prem latency and data residency for won-denominated institutions that hesitate to export embeddings.

Funding and follow-on work

Kim’s lab is exploring GPU-accelerated rebuild paths for billion-vector corpora common in telecom log analysis. Government AI safety working groups have asked whether retrieval drift should trigger mandatory regression tests; CONDA could supply metrics if vendors expose connectivity health scores.

For now, the headline is empirical: dynamic corpora no longer force a choice between fresh data and reliable search. Korean RAG pilots that stalled on nightly reindex windows have a local paper—and soon a product—to cite in procurement memos.