Author: Dan Petrovic

  • Top 10 Most Recent Papers by MUVERA Authors

    Top 10 Most Recent Papers by MUVERA Authors

    —

    by

    Dan Petrovic
    in , ,

    MUVERA Authors: 1. Laxman Dhulipala (Google Research & UMD) Top 10 Recent Papers (2023-2025) Research Focus Areas 2. Majid Hadian (Google DeepMind) Top 10 Recent Papers (2023-2025) Research Focus Areas 3. Jason Lee (Google Research & UC Berkeley) Top 10 Recent Papers (2023-2025) Research Focus Areas 4. Rajesh Jayaram (Google Research) Top 10 Recent Papers…

  • Training Gemma‑3‑1B Embedding Model with LoRA

    Training Gemma‑3‑1B Embedding Model with LoRA

    —

    by

    Dan Petrovic
    in , ,

    In our previous post, Training a Query Fan-Out Model, we demonstrated how to generate millions of high-quality query reformulations without human labelling, by navigating the embedding space between a seed query and its target document and then decoding each intermediate vector back into text using a trained query decoder. That decoder’s success critically depends on…

  • Training a Query Fan-Out Model

    Training a Query Fan-Out Model

    —

    by

    Dan Petrovic
    in , ,

    Google discovered how to generate millions of high-quality query reformulations without human input by literally traversing the mathematical space between queries and their target documents. Here’s How it Works This generated 863,307 training examples for a query suggestion model (qsT5) that outperforms all existing baselines. Query Decoder + Latent Space Traversal Step 1: Build a…

  • Cosine Similarity or Dot Product?

    Cosine Similarity or Dot Product?

    —

    by

    Dan Petrovic
    in ,

    Google’s embedder uses dot product between normalized vectors which is computationally more efficient but mathematically equivalent to cosine similarity. How Googler’s work and think internally typically aligns with their open source code (Gemini -> Gemma) and Chrome is no exception. It’s why I look there for answers and clarity on Google’s machine learning approaches. After…

  • Universal Query Classifier

    Universal Query Classifier

    —

    by

    Dan Petrovic
    in , ,

    Generalist, Open‑Set Classification for Any Label Taxonomy We’ve developed a search query classifier that takes any list of labels you hand it at inference time and tells you which ones match each search query. No retraining, ever. Just swap in new labels as they appear. Old workflow Pain New workflow Build + label data + retrain…

  • Another failed attempt to kill SEO

    Another failed attempt to kill SEO

    —

    by

    Dan Petrovic
    in

    If Marie Haynes, Barry Schwartz or Cindy Krum had written an article declaring SEO dead and proposing we rebrand our industry you’d seriously consider it. Wouldn’t you? What about Zach Cohen and Seema Amble? I don’t know either. Looked them up just now. Two VC people with insignificant footprint or long-term interest in SEO, Machine…

  • Vector Embedding Optimization

    Vector Embedding Optimization

    —

    by

    Dan Petrovic
    in

    Embedding Methods Evaluation: Results, Key Findings, and a Surprising Insight On June 6, 2025, we ran a comprehensive evaluation comparing four different embedding methods—regular, binary, mrl, and mrl_binary—on a dataset of paired sentences. The goal was to measure each method’s speed, storage footprint, similarity quality, and accuracy against a ground-truth of sentence pairs. Below, we…

  • Dissecting Gemini’s Tokenizer and Token Scores

    Dissecting Gemini’s Tokenizer and Token Scores

    —

    by

    Dan Petrovic
    in ,

    As a technical SEO, you might be diving into machine learning (ML) to understand how tools like Google’s Gemini process text. One foundational concept is subword tokenization—breaking words into smaller pieces called “tokens.” While tokens themselves are context-agnostic (they don’t consider surrounding words), they do carry an inherent bias: each token’s likelihood reflects how prominent…

  • There’s a small army of on-device models coming to Chrome

    There’s a small army of on-device models coming to Chrome

    —

    by

    Dan Petrovic
    in ,

    1. ULM128M 2. LLMIT1B 3. GEMINI2_NANOV2 4. GEMINI2_NANOV2_EE2Q 5. GEMINI_XS 6. GEMINI_XS_DRAFTER_6LAYER_CAUSAL_USM_700M_RESIDUAL 7. GEMINI_XS_LUSM_700M_RESIDUAL_BOTTOM15 8. GEMINI2_NANOV2_EE12Q 9. GEMINI2_NANOV2_EE2_LUSM_700M 10. GEMINI2_NANOV2_CAUSAL_700M 11. GEMINI2_NANOV2_EE20_CAUSAL_LUSM_700M 12. GEMINI_XL_DRAFTER_24LAYER 13. GEMINI_XS_FA1 14. GEMMA2_8B 15. GEMMA2_7B 16. GEMMA2_2B 17. GEMMA3_1B 18. GEMMA3_4B 19. GEMMA3_12B 20. GEMMA3_27B 21. STABLELM_4E1T_3B_PHI_2_TF_LITE

  • Multi-Step Research Agent

    Multi-Step Research Agent

    —

    by

    Dan Petrovic
    in

    This post is the output from the implementation of Google’s query fan-out in an agentic framework inspired by Google’s Gemini Agent repo. Query: What services does DEJAN AI offer? The following is raw copy/paste from the agent’s output: Research Progress Initial Search Strategy: The search strategy aims to identify the range of services offered by…