Research Dashboard

    Deep dives into the two research projects driving my current work — hierarchical ANN quantization and knowledge-graph-augmented retrieval.

    R1 · Vector Retrieval

    Reusable Codebooks: Hierarchical TurboQuant for Two-Tier ANN Retrieval

    Two-tier ANN pipelines train two independent Product Quantization codebooks — one for the cheap prefilter, one for the expensive reranker — duplicating storage and training cost. Can a single nested codebook serve both tiers?

    660–1300×
    Faster index build
    +2.2
    R@1 vs PQ + rerank (GloVe)
    200 / 1536 / 3072
    Embedding dims
    3
    Nesting strategies

    Recall@1 at 4 bits, by method & dataset

    Top-down matches native TurboQuant exactly; bottom-up loses up to 0.116 (paper Tables 6–7).

    Index construction time (seconds, log scale)

    Two k-means trainings (PQ-2 + PQ-4) vs one Lloyd-Max codebook in under 1 s (paper Table 4).

    Recall@1 across all methods and datasets.
    Recall@1 across all methods and datasets.
    Bottom-Up · Top-Down · Middle-Anchor nesting.
    Bottom-Up · Top-Down · Middle-Anchor nesting.
    Middle-anchor Lloyd-Max nested codebooks.
    Middle-anchor Lloyd-Max nested codebooks.

    Key Insight

    Top-down decimation matches unconstrained TurboQuant within ±0.005 R@1 on all three datasets and beats PQ + reranking by +2.2 R@1 on GloVe-200 — using one shared codebook instead of two, with 660–1300× faster index construction. Bottom-up degrades by up to 12 points.

    Limitations

    Nesting forces an MSE redistribution: no nested family reaches zero penalty at two bit-widths at once, so low-bit tiers (b=2) pay a measurable cost. Data-oblivious codebooks lose ground to PQ on heavily anisotropic, high-dimensional corpora. Manuscript, May 2026 (advisor: Prof. Ainesh Bakshi).

    R2 · Knowledge-Graph RAG

    Erica – LLM Knowledge Graph Reasoning

    repo

    Pure vector-based retrieval produces hallucinated or incoherent answers lacking structured reasoning paths.

    5,200+
    KG Nodes
    13.4k
    Relations
    −38pp
    Hallucination ↓
    Neo4j + Vectors
    Backend

    Knowledge graph growth

    LLM-driven semantic extraction over five pipeline stages.

    Vector-only vs Graph-augmented RAG

    Explicit reasoning paths dominate on faithfulness & hallucination.

    Erica pipeline

    UnstructuredSourcesLLM SemanticExtractionConcepts +PrerequisitesNeo4j KG+ EmbeddingsGraph-augmentedAnswer

    Key Insight

    Graph-augmented retrieval provides explicit reasoning paths that vector similarity alone cannot capture, measurably improving consistency over pure vector retrieval.

    Limitations

    Knowledge graph construction from unstructured sources requires careful schema design. Long-form outputs need asynchronous execution pipelines.