# On-Device Commercial Intent Retrieval Under Size, Latency, and Privacy Constraints: A 3 MiB Retrieval System with Typed Egress Boundaries

arXiv: https://arxiv.org/abs/2610.00170 · HTML: https://mlsanity.com/paper/2610.00170/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 25 of its 49 references that are in the map.
- First posted on arXiv 16 Sep 2026.

## Builds on (most cited first, 25 of 25)

- [Learning Transferable Visual Models From Natural Language Supervision](https://mlsanity.com/paper/2103.00020/) (2021)
- [Distilling the Knowledge in a Neural Network](https://mlsanity.com/paper/1503.02531/) (2015)
- [Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks](https://mlsanity.com/paper/1908.10084/) (2019)
- [Representation Learning with Contrastive Predictive Coding](https://mlsanity.com/paper/1807.03748/) (2018)
- [Communication-Efficient Learning of Deep Networks from Decentralized Data](https://mlsanity.com/paper/1602.05629/) (2016)
- [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://mlsanity.com/paper/1910.01108/) (2019)
- [Dense Passage Retrieval for Open-Domain Question Answering](https://mlsanity.com/paper/2004.04906/) (2020)
- [Distributed Representations of Words and Phrases and their Compositionality](https://mlsanity.com/paper/1310.4546/) (2013)
- [Advances and Open Problems in Federated Learning](https://mlsanity.com/paper/1912.04977/) (2019)
- [Deep Learning with Differential Privacy](https://mlsanity.com/paper/1607.00133/) (2016)
- [Billion-scale similarity search with GPUs](https://mlsanity.com/paper/1702.08734/) (2017)
- [GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers](https://mlsanity.com/paper/2210.17323/) (2022)
- [Enriching Word Vectors with Subword Information](https://mlsanity.com/paper/1607.04606/) (2016)
- [ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT](https://mlsanity.com/paper/2004.12832/) (2020)
- [LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale](https://mlsanity.com/paper/2208.07339/) (2022)
- [Federated Learning with Personalization Layers](https://mlsanity.com/paper/1912.00818/) (2019)
- [MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices](https://mlsanity.com/paper/2004.02984/) (2020)
- [ABO: Dataset and Benchmarks for Real-World 3D Object Understanding](https://mlsanity.com/paper/2110.06199/) (2021)
- [Concentrated Differential Privacy: Simplifications, Extensions, and Lower Bounds](https://mlsanity.com/paper/1605.02065/) (2016)
- [Matryoshka Representation Learning](https://mlsanity.com/paper/2205.13147/) (2022)
- [RAPPOR: Randomized Aggregatable Privacy-Preserving Ordinal Response](https://mlsanity.com/paper/1407.6981/) (2014)
- [KLUE: Korean Language Understanding Evaluation](https://mlsanity.com/paper/2105.09680/) (2021)
- [Federated Learning of Gboard Language Models with Differential Privacy](https://mlsanity.com/paper/2305.18465/) (2023)
- [Shopping Queries Dataset: A Large-Scale ESCI Benchmark for Improving Product Search](https://mlsanity.com/paper/2206.06588/) (2022)
- [Timing-Aware Repurchase Prediction for Web-Scale E-Commerce: Survival Models for Multi-Surface Grocery Recommendation](https://mlsanity.com/paper/2608.28393/) (2026)

## Abstract (arXiv)

We study commercial intent inference that runs entirely on the user's device, under three constraints frozen before the work began: the downloaded payload under 3 MiB, Tier-0 inference under 20 ms at p95, and no raw text, content embedding, or stable identifier leaving the device. Under them we build a retrieval path over a 6,020-leaf commercial taxonomy: a static embedding table distilled from a Korean sentence transformer, quantized to 4 bits, no inference runtime. Our main result is where that constraint costs accuracy. On real Korean commerce text labelled by others (22,900 AI-Hub shopping reviews), mid-category top-5 on real product names is 75.0% against an 18.4% permutation baseline, but splits on one observable: a query containing some leaf name as a substring scores 83.5%, one containing none 45.2%. A generic 196.6x larger teacher seemed to localize the gap (+20.1 pp without an anchor, +0.1 with). That null was two effects cancelling: the same teacher fine-tuned on the student's own contrastive pairs reaches 0.8586 and beats the pure-encoder student by +10.6 pp with an anchor and +20.9 pp without. The cost is not uniform, but it is not free anywhere; where the anchor is absent, task adaptation buys the teacher nothing, so what the constrained encoder lacks there is capacity. The expensive regime is detectable on-device from the ranker's own score margin: declining the least confident fifth lifts the rest to 0.8296. A second axis we first reported, a manufacturer model code, does not survive source-category fixed effects (-4.0 pp, p=0.51); the anchor does (+13.0 pp). Payload is 2,942,652 bytes, all three library links measured. Tier-0 p95 is 4.431 and 3.670 ms on two iPhones (A14, A16) and 5.080 ms on a budget Android tablet (Snapdragon 695), all slower than three server CPUs on the same code. Taxonomy supervision is mostly synthetic Korean utterances.