GitHub - emil-celestix/celestix-ifr: Meet IFR: a bio-inspired engine solving RAG's biggest flaws. It achieves true O(1) scaling latency stays <5ms even as data grows 1000x. With +14% accuracy over RAG-rerank, it uncovers multi-hop links standard search misses. Patent-pending tech for enterprise-scale discovery. All benchmark JSONs and test results are now live on GitHub.
Briefly

GitHub - emil-celestix/celestix-ifr: Meet IFR: a bio-inspired engine solving RAG's biggest flaws. It achieves true O(1) scaling latency stays <5ms even as data grows 1000x. With +14% accuracy over RAG-rerank, it uncovers multi-hop links standard search misses. Patent-pending tech for enterprise-scale discovery. All benchmark JSONs and test results are now live on GitHub.
"IFR successfully finds multi-hop targets that are entirely invisible to RAG, achieving a 15% Hit@20 on multi-hop evaluation, while traditional methods scored 0%."
"Dynamic query mutation is strictly required for success; setting the mutation rate to zero resulted in a 0% multi-hop hit rate and near-zero MRR."
"At high data scales, Beam Search with a novelty exploration bonus outperformed greedy traversal, demonstrating a significant performance difference in retrieval effectiveness."
Induced-Fit Retrieval (IFR) addresses multi-hop reasoning by treating query retrieval as dynamic graph traversal. Unlike traditional methods, IFR adapts the query vector at each hop based on node embeddings. This approach allows for the discovery of semantically distant concepts. Testing showed IFR outperformed traditional methods in retrieval but faced challenges in end-to-end generation quality due to query mutation drift. A conditional pass verdict indicates the need for improvements in ranking and drift-damping for future versions.
Read at GitHub
Unable to calculate read time
[
|
]