The problem
Retrieval-augmented generation depends on finding relevant documents quickly. Comparing queries with many document representations creates substantial memory traffic and computation.
What we did
We introduced confidence-guided dimension adaptivity for nearest-neighbor retrieval and paired it with an in-DRAM processing architecture. The approach adapts the dimensions used for retrieval and parallelizes Hamming similarity matching between queries and documents within memory.
What we found
The evaluation summarized in the resume reports fewer memory accesses and approximately 10× speedup compared with a GPU baseline. CASPER combines algorithmic adaptivity with memory-side execution to accelerate information retrieval.
Paper
Ashiqur Rasul, Anurup Saha, Sai Pranav Komaragiri, Abhijit Chatterjee. “CASPER: Confidence-Guided Adaptive Nearest Neighbor Search with In-DRAM Processing for Efficient Information Retrieval.” DATE 2027, under review, as listed in the latest resume.