A Technical Evaluation of the RuVector-Powered Intelligent Email Platform
Email remains the dominant medium for professional and personal digital communication, yet the volume of messages received by typical users has grown to a scale that overwhelms conventional management strategies. This paper presents a critical evaluation of Emailibrium, a proposed vector-native email intelligence platform that integrates approximate nearest neighbor (ANN) search, graph neural networks, self-adaptive learning, and multi-modal content extraction to transform email management from keyword-driven search and manual rule-based filtering into a semantic-first, continuously learning system. The platform's architecture centers on RuVector, a Rust-native vector database employing Hierarchical Navigable Small World (HNSW) indexing, SIMD-optimized distance computation, scalar quantization, and a three-tier adaptive learning model (SONA). We analyze the technical feasibility of the platform's claims --- including sub-10-minute inbox zero for 10,000+ message inboxes, sub-50ms hybrid search latency, and greater than 95% categorization accuracy --- against established benchmarks in information retrieval, graph-based learning, and human-computer interaction research. We identify both novel contributions and significant risks, and propose an evaluation framework grounded in standard IR metrics, controlled user studies, and ablation analysis.
The Radicati Group estimates that the average office worker receives 121 emails per day as of 2023, a figure projected to continue growing (Radicati Group, 2023). This volume creates what Whittaker and Sidner (1996) termed the "email overload" problem: users accumulate large backlogs, struggle to locate previously received information, and spend disproportionate time on triage activities relative to substantive work. Despite decades of research, commercial email clients continue to rely primarily on full-text keyword search, manual folder hierarchies, and simple heuristic filters --- tools that scale poorly with mailbox size.
Several trends motivate a re-examination of email management architectures. First, advances in dense retrieval models (Karpukhin et al., 2020) and efficient ANN indexing (Malkov & Yashunin, 2020) have made semantic search over millions of documents feasible at single-digit millisecond latencies. Second, sentence embedding models such as Sentence-BERT (Reimers & Gurevych, 2019) provide high-quality fixed-dimensional representations that capture semantic similarity far beyond lexical overlap. Third, graph neural networks (Hamilton et al., 2017) offer principled methods for discovering latent structure --- such as topic clusters and communication patterns --- in relational data. Finally, growing privacy concerns (Zuboff, 2019) have driven interest in local-first architectures that avoid transmitting personal data to cloud services.
This evaluation addresses the following questions:
-
RQ1: Can a hybrid search architecture combining full-text search (FTS) with HNSW-indexed dense vectors, fused via Reciprocal Rank Fusion (RRF), deliver retrieval quality competitive with state-of-the-art email search while maintaining sub-50ms latency?
-
RQ2: Is graph neural network-based clustering (specifically GraphSAGE over HNSW neighbor graphs) a viable approach for automatic email topic discovery and subscription detection?
-
RQ3: Can a three-tier adaptive learning model operating at the interaction, session, and long-term timescales materially improve search relevance and categorization accuracy over time without requiring explicit user training?
-
RQ4: Is a privacy-preserving, local-first architecture with all embedding generation and vector operations performed on consumer hardware technically feasible at the proposed scale (100K+ emails, sub-second latency)?
This paper does not report empirical results from a deployed system; rather, it provides a rigorous technical evaluation of the Emailibrium v1.4 implementation plan (2026-03-23). We situate the proposed architecture within the existing literature, formalize the key algorithms, analyze computational complexity, identify risks and limitations, and propose an evaluation methodology suitable for validating the platform's claims.
The Hierarchical Navigable Small World (HNSW) algorithm (Malkov & Yashunin, 2020) is a graph-based index for approximate nearest neighbor search that achieves logarithmic query complexity O(log n) with high recall. HNSW constructs a multi-layer navigable small-world graph where upper layers provide coarse navigation and the bottom layer contains all data points. The algorithm has become the de facto standard in production vector databases including Pinecone, Weaviate, and Qdrant.
Key parameters controlling the recall-speed tradeoff are M (the number of bi-directional links per node, typically 16-64), ef_construction (the beam width during index building), and ef_search (the beam width at query time). Emailibrium proposes M=16, ef_construction=200, ef_search=100 --- conservative values that favor recall over speed, appropriate for the expected corpus sizes.
SIMD (Single Instruction, Multiple Data) acceleration of distance computations has been extensively studied. Johnson et al. (2019) demonstrated 4-8x speedups for L2 and inner product computations using AVX-512 instructions in the FAISS library. The claim of 16 million operations per second for cosine similarity on modern hardware is consistent with published benchmarks on AVX2-capable processors.
Reimers and Gurevych (2019) introduced Sentence-BERT (SBERT), which fine-tunes BERT-family models using siamese and triplet network structures to produce semantically meaningful fixed-size sentence embeddings. The all-MiniLM-L6-v2 model selected by Emailibrium (384 dimensions, 22M parameters) represents a favorable point on the quality-efficiency Pareto frontier: Wang et al. (2020) showed it achieves approximately 95% of the retrieval quality of larger models at a fraction of the computational cost.
Dense passage retrieval (DPR) (Karpukhin et al., 2020) demonstrated that learned dense representations can outperform BM25 on open-domain question answering when sufficient training data is available. However, Thakur et al. (2021) showed in the BEIR benchmark that zero-shot dense retrieval often underperforms BM25 on out-of-domain tasks, motivating hybrid approaches that combine both.
Reciprocal Rank Fusion (RRF) (Cormack et al., 2009) is a rank aggregation method that combines multiple ranked lists without requiring score normalization:
where k is a constant (typically 60), r_i(d) is the rank of document d in the i-th result list, and the sum is over all contributing rankers. RRF has been shown to be robust, parameter-insensitive, and competitive with learned fusion methods (Cormack et al., 2009). Its use for combining BM25/FTS results with dense vector results has become standard practice, employed by Elasticsearch, Vespa, and other production systems.
Hamilton et al. (2017) introduced GraphSAGE, an inductive framework for learning node embeddings through neighborhood aggregation. Unlike transductive methods such as DeepWalk (Perozzi et al., 2014) or Node2Vec (Grover & Leskovec, 2016), GraphSAGE generalizes to unseen nodes, making it suitable for continuously growing email corpora.
Email network analysis has a substantial history. Klimt and Yang (2004) demonstrated topic modeling on the Enron corpus, while Dredze et al. (2008) explored intelligent email classification. The application of GNNs to email data is relatively novel; most prior work used shallow graph features or traditional community detection algorithms (Blondel et al., 2008). Emailibrium's proposal to construct a graph from HNSW neighbor relationships and apply GraphSAGE for topic clustering represents a non-trivial architectural contribution.
Online learning systems that adapt to user behavior have been studied extensively in the context of recommender systems (Rendle, 2010) and information retrieval (Joachims, 2002). Emailibrium's SONA (Self-Organizing Neural Adaptation) model proposes three learning tiers:
- Tier 1 (Instant): Per-interaction weight adjustments, analogous to bandit-style feedback (Li et al., 2010).
- Tier 2 (Session): Accumulated preferences within a session, similar to session-based recommendation (Hidasi et al., 2016).
- Tier 3 (Long-term): Persistent learned models updated across sessions, analogous to collaborative filtering updates.
This multi-timescale approach mirrors findings from cognitive science on the distinction between working memory, episodic memory, and semantic memory (Tulving, 1972), though the technical implementation details require further specification.
The movement toward local-first AI processing (Kleppmann et al., 2019) has accelerated with efficient inference runtimes such as llama.cpp (Gerganov, 2023) and ONNX Runtime. Running embedding models locally eliminates the need to transmit email contents to external servers, addressing privacy concerns that are particularly acute for email data containing PII, financial information, and privileged communications.
Quantization techniques --- including scalar quantization (Jacob et al., 2018), product quantization (Jegou et al., 2011), and binary quantization --- provide memory compression with controllable quality degradation. The proposed 4x compression via scalar quantization (fp32 to int8) is well-supported by the literature, with typical recall loss below 2% for cosine similarity tasks (Guo et al., 2020).
The "inbox zero" methodology, popularized by Merlin Mann (2006), advocates processing all incoming email to empty the inbox through a decision framework of delete, delegate, respond, defer, or archive. Whittaker et al. (2011) found that heavy email users who maintain clean inboxes report higher satisfaction but spend more time on email management. The research question of whether automated clustering and batch actions can reduce the time cost of inbox maintenance while preserving the benefits of a clean inbox remains open.
Emailibrium employs a layered architecture with four principal tiers:
Tier 1 --- Presentation Layer: A React TypeScript single-page application communicating with the backend via REST and Server-Sent Events (SSE). The migration from Tauri 2.0 desktop to a pure web SPA represents a deliberate trade of native desktop integration for universal accessibility and simplified deployment.
Tier 2 --- API Gateway: An Axum 0.8 (Rust) web framework handling authentication (OAuth2/JWT), request routing, and response streaming. Axum's tower-based middleware architecture enables composable cross-cutting concerns such as rate limiting, logging, and authentication.
Tier 3 --- Intelligence Layer: The novel core of the system, comprising the RuVector engine (HNSW indexing, SONA learning, GNN clustering, quantization), the embedding pipeline, the category classifier, and the pattern detector. This layer transforms raw email data into semantic representations and actionable intelligence.
Tier 4 --- Data Layer: A heterogeneous persistence layer combining SQLite/PostgreSQL for structured relational data and full-text search (FTS5), RuVector's REDB-backed storage for vector data, and Moka/Redis for caching.
The ingestion pipeline processes each email e through six sequential stages:
- Parse: Extract structured metadata M(e) = {subject, sender, recipients, date, headers, labels}
- Embed: Generate dense vector v(e) = f(concat(M(e), body(e))) where f is the all-MiniLM-L6-v2 encoder
- Extract Assets: Process HTML body, inline images (OCR + CLIP), attachments (PDF/DOCX/XLSX text extraction), and URLs
- Classify: Assign category c(e) = argmaxc cos(_v(e), mu_c) where mu_c is the centroid of category c, with LLM fallback when max similarity falls below threshold tau
- Detect Patterns: Identify subscriptions via header analysis (List-Unsubscribe), sender frequency modeling, and template similarity clustering
- Apply Rules: Execute user-defined and system-generated rules against the enriched email representation
The pipeline supports two execution modes: fast mode (text embedding only, ~5ms per email) for interactive responsiveness during initial sync, and deep mode (full multi-asset extraction, ~20ms per email excluding heavy attachments) for background enrichment.
The choice of all-MiniLM-L6-v2 for text embeddings reflects an explicit optimization for inference latency over absolute retrieval quality. With 22 million parameters and 384-dimensional output, the model runs at approximately 5ms per inference on consumer hardware with ONNX Runtime acceleration. Larger models such as all-mpnet-base-v2 (110M parameters, 768 dimensions) would approximately double recall@10 improvement on semantic textual similarity benchmarks but at 4-5x the computational cost and storage overhead.
For image embeddings, the plan selects CLIP ViT-B-32 (Radford et al., 2021), which projects images into a 512-dimensional space shared with text. This enables cross-modal search queries such as finding emails containing images similar to a text description. The architectural decision to maintain separate vector collections for text, image-text (OCR), image-visual (CLIP), and attachment-text reflects a pragmatic recognition that heterogeneous embedding spaces require collection-level separation to avoid dimensionality mismatches.
The email text template concatenates subject, sender, recipients, date, labels, and truncated body (400 characters). This design choice prioritizes metadata-enriched representations over body-only embeddings, which is well-motivated: in email search, knowing who sent a message and when is often as informative as the message content itself.
For a corpus of N emails, the HNSW index provides:
- Construction: O(N log N) time, O(N * M) space, where M is the number of links per node
- Query: O(log N) time for a single nearest-neighbor search with beam width ef
- Insertion: O(log N) amortized for adding a single vector
- Memory: For 100,000 emails at 384 dimensions with fp32 precision: 100,000 384 4 bytes = ~147 MB for vectors alone, plus graph structure overhead
With scalar quantization (fp32 to int8), vector storage reduces to ~37 MB, validating the plan's claim of ~100 MB for 100K emails (including graph structure and metadata). The 4x compression ratio is consistent with published quantization results (Guo et al., 2020).
The hybrid search procedure combines lexical and semantic retrieval:
Let Q be a query string. The algorithm proceeds as:
- Compute query embedding: v_q = f(Q), cost O(d) where d = 384
- Vector search: retrieve top-k_v results R_v = HNSW.search(v_q, k_v, ef), cost O(log N)
- Full-text search: retrieve top-k_t results R_t = FTS5.search(Q, k_t), cost O(log N) with inverted index
- Rank fusion: for each document d in R_v union R_t:
where k = 60 and r_i(d) is the rank of d in result set i (or infinity if absent).
- Filter: apply structured predicates (date range, sender, labels)
- Re-rank: apply SONA learned weights
The claimed total latency of ~20ms is plausible given that embedding inference (~5ms), HNSW search (~2ms), FTS5 search (~10ms), and fusion/filtering (~3ms) can be partially parallelized (steps 2 and 3 execute concurrently).
The plan proposes constructing an email graph G = (V, E) where vertices represent emails and edges connect HNSW neighbors (i.e., emails with high vector similarity). GraphSAGE (Hamilton et al., 2017) is then applied to learn node embeddings that capture both content similarity and structural patterns in the communication graph.
The GraphSAGE aggregation for node v at layer l is:
where AGG is a permutation-invariant aggregation function (mean, LSTM, or max-pool), W is a learnable weight matrix, and sigma is a non-linearity.
This approach has the advantage of operating on the already-constructed HNSW graph, avoiding the need for a separate graph construction step. However, HNSW neighbor graphs are optimized for navigability rather than semantic coherence, and the relationship between HNSW connectivity and meaningful email clusters requires empirical validation.
The SONA model implements learning at three temporal scales:
-
Tier 1 (Instant feedback): When a user clicks a search result, opens an email, or provides explicit relevance feedback, the system performs a lightweight weight update. This is analogous to online gradient descent with a high learning rate, affecting only the current query context.
-
Tier 2 (Session accumulation): Over a user session, the system aggregates interaction signals to build a session-level preference model. This captures short-term intent (e.g., "I am currently looking for emails about project X").
-
Tier 3 (Long-term model): Periodically, session-level signals are consolidated into a persistent user model that influences future search ranking and categorization. This corresponds to the slowly-updated user profile in collaborative filtering systems.
The multi-timescale design is principled, but the plan lacks formal specification of the update rules, convergence guarantees, and safeguards against catastrophic forgetting or feedback loops where early misclassifications reinforce themselves.
The plan proposes adaptive quantization based on corpus size:
| Corpus Size | Quantization | Precision | Memory per 384D Vector |
|---|---|---|---|
| < 10K emails | None (fp32) | Full | 1,536 bytes |
| 10K--50K | Scalar (int8) | ~99% recall | 384 bytes |
| 50K--200K | Product (PQ) | ~97% recall | ~96 bytes |
| > 200K | Binary | ~90% recall | 48 bytes |
The recall degradation estimates are consistent with published benchmarks (Jegou et al., 2011; Guo et al., 2020). The automatic scaling strategy is pragmatic, though transitions between quantization levels require index reconstruction, which should be performed as a background operation to avoid service disruption.
The hybrid search system should be evaluated using standard information retrieval metrics:
- Recall@k: The fraction of relevant documents appearing in the top-k results. For email search, k values of 5, 10, and 20 are most relevant to user experience.
- NDCG@k (Normalized Discounted Cumulative Gain): Measures ranking quality, penalizing relevant documents that appear lower in the results.
- MRR (Mean Reciprocal Rank): The average of the reciprocal of the rank of the first relevant result, particularly appropriate for navigational queries.
- Latency at p50, p95, p99: End-to-end query latency percentiles.
An ablation study should compare: (a) FTS5 alone, (b) vector search alone, (c) hybrid without SONA re-ranking, and (d) full hybrid with SONA. This would isolate the contribution of each component.
Categorization accuracy should be measured as:
- Macro-averaged F1: Averaged across all categories, ensuring minority categories are not ignored.
- Precision-recall curves per category: To identify categories where vector centroid classification underperforms.
- LLM fallback rate: The fraction of emails requiring LLM classification due to low centroid confidence, which directly impacts latency and cost.
The claimed >95% accuracy requires a labeled evaluation dataset. The Enron corpus (Klimt & Yang, 2004) provides one benchmark, though its distribution differs significantly from modern email patterns. A purpose-built evaluation set from consenting users would be more representative.
GNN-based clustering should be evaluated using:
- Silhouette coefficient: Measuring intra-cluster cohesion versus inter-cluster separation.
- Adjusted Rand Index (ARI): Comparing discovered clusters against human-labeled topic assignments.
- Subscription detection precision and recall: For the specific task of identifying subscription/newsletter emails.
The claimed >98% subscription detection recall is achievable given that most subscription emails contain explicit List-Unsubscribe headers (RFC 2369); header-based detection alone should achieve high recall, with vector clustering providing incremental improvement for subscriptions lacking standard headers.
The headline claim --- that users can achieve inbox zero for 10,000+ emails within 10 minutes --- requires a controlled user study with the following design:
- Participants: N >= 30, recruited from a population of heavy email users (>100 emails/day).
- Protocol: Participants connect a real email account, perform initial ingestion, and then use the Emailibrium interface to triage their inbox. Time-to-inbox-zero is measured.
- Controls: Comparison against (a) the same participants using their existing email client, and (b) a version of Emailibrium without vector intelligence (keyword search only).
- Metrics: Time to zero unread, number of user actions required, user satisfaction (SUS scale), and a 1-week follow-up to measure inbox rebound rate.
This study design would provide evidence for or against the 10-minute claim while controlling for confounds such as prior inbox organization and email volume.
Performance claims should be validated through:
- Microbenchmarks: Single-embedding latency, batch-embedding throughput, HNSW search latency at various index sizes (1K, 10K, 100K, 500K vectors).
- End-to-end benchmarks: Full ingestion pipeline throughput (emails/second) for text-only and multi-asset modes.
- Memory profiling: Resident set size (RSS) measured at steady state for various corpus sizes, compared against the claimed 700MB for 100K emails.
- Hardware matrix: Performance on representative consumer hardware (Apple Silicon M-series, Intel/AMD with AVX2, systems without AVX).
Language-homogeneous architecture: The decision to implement the entire backend --- including the vector database, embedding inference, and GNN clustering --- in Rust eliminates FFI overhead, serialization costs, and the operational complexity of managing multiple runtime environments (Python, Go, Java). This is a meaningful architectural contribution, as most comparable systems require polyglot deployments.
Multi-asset semantic search: Extending vector search beyond email body text to include OCR'd images, attachment text, and CLIP visual embeddings provides a richer search space than any commercial email client currently offers. The ability to search for "the chart Dave sent about Q3 revenue" across all content modalities is a compelling capability.
Hybrid search with adaptive re-ranking: The combination of FTS5, HNSW, RRF, and SONA learning addresses the known weakness of pure dense retrieval on exact-match queries (Thakur et al., 2021) while adding personalization. The architecture correctly positions keyword search as a complement to, rather than a replacement for, semantic search.
Privacy-first design: Local embedding generation and vector storage address legitimate privacy concerns that make cloud-based email intelligence solutions unacceptable for many users and organizations. The absence of any requirement to transmit email content to external services is a significant differentiator.
RuVector maturity: RuVector is a relatively new project within the Rust ecosystem. The plan acknowledges API instability risk but does not address the absence of long-term production track record. Critical concerns include: (a) REDB durability guarantees under crash scenarios, (b) correctness of the HNSW implementation under concurrent read-write workloads, and (c) the maturity of the SONA and GNN components. A comparison against established alternatives (Qdrant, Milvus) for reliability and correctness would strengthen the architecture.
SONA specification gap: The three-tier learning model is described at a high level but lacks formal specification of update rules, learning rates, convergence properties, and safeguards against feedback loops. Without these details, the claimed +12% recall improvement over time cannot be evaluated. Degenerate cases --- such as a user who consistently clicks the first result regardless of relevance --- could corrupt the learned model.
Embedding model limitations: The all-MiniLM-L6-v2 model, while efficient, has known weaknesses with domain-specific vocabulary, non-English text, and very short queries. Email communication frequently involves jargon, abbreviations, and code-switching. The plan does not address fine-tuning or domain adaptation strategies.
Multi-asset extraction reliability: OCR quality varies dramatically with image quality, font, and language. PDF text extraction is unreliable for scanned documents without OCR. The pipeline's overall utility depends heavily on the quality of content extraction, which is not uniform across input types.
Scalability ceiling: While the architecture is designed for local execution, the combination of embedding generation, vector indexing, GNN clustering, and SONA learning creates significant CPU and memory pressure on consumer hardware. The claimed performance targets assume modern hardware with SIMD support; degradation on older or resource-constrained systems is not addressed.
| System | Semantic Search | Local Processing | Adaptive Learning | Multi-Modal |
|---|---|---|---|---|
| Gmail (Google) | Yes (cloud) | No | Yes (cloud ML) | Partial |
| Outlook (Microsoft) | Yes (cloud) | No | Yes (cloud ML) | Partial |
| Apple Mail | No | Yes | No | No |
| Thunderbird | No | Yes | No | No |
| Superhuman | Limited | No | Limited | No |
| Emailibrium | Yes (local) | Yes | Yes (SONA) | Yes |
Emailibrium occupies a unique position as the only system proposing full semantic search with local-only processing. However, it must be noted that Google and Microsoft have vastly larger training datasets and compute budgets for their ML models, and their cloud-based approach enables capabilities (such as cross-user learning and real-time model updates) that a local-first system cannot replicate.
Local vector embeddings provide strong privacy guarantees against external data collection, but introduce new considerations:
- Embedding invertibility: While dense embeddings are not directly human-readable, recent work (Morris et al., 2023) has demonstrated that text can be partially recovered from sentence embeddings. The vector store should therefore be treated as sensitive data requiring encryption at rest.
- Metadata leakage: Even without embeddings, the SQLite database contains full email metadata (sender, subject, dates) that constitutes sensitive information.
- Device security: Local-first shifts the threat model from cloud breaches to device compromise. Disk encryption and secure key management are prerequisites.
The plan addresses some of these concerns (Web Crypto API for browser storage, OWASP-compliant secrets management) but does not discuss embedding invertibility risks.
Several research directions emerge from this evaluation:
-
Federated learning for cross-user intelligence: Exploring privacy-preserving federated learning techniques (McMahan et al., 2017) to enable subscription detection and categorization models to benefit from aggregate patterns without sharing individual email data.
-
Retrieval-augmented generation (RAG) for email composition: Leveraging the vector store as a retrieval backend for LLM-assisted email drafting, enabling responses that reference prior correspondence.
-
Temporal graph networks: Extending the static GNN approach with temporal graph networks (Rossi et al., 2020) to model the evolution of communication patterns over time.
-
Multilingual and cross-lingual embeddings: Evaluating multilingual sentence embedding models (e.g., multilingual-e5-large) to support non-English and multilingual email corpora.
-
Formal evaluation of SONA: Designing controlled experiments with synthetic feedback signals to characterize SONA's convergence properties, stability, and robustness to adversarial or noisy feedback.
-
Differential privacy for vector stores: Investigating mechanisms to add calibrated noise to stored embeddings to provide formal privacy guarantees against embedding inversion attacks.
-
Energy and sustainability analysis: Quantifying the computational and energy costs of local embedding generation versus cloud-based alternatives, particularly relevant for mobile and battery-constrained devices.
Emailibrium's v1.4 plan presents an ambitious and technically grounded architecture for reimagining email management through vector-native intelligence. The core technical decisions --- HNSW indexing, hybrid search with RRF, GraphSAGE clustering, multi-modal content extraction, and local-first processing --- are each well-supported by the existing literature. The integration of these components into a coherent, single-language (Rust) system with a modern React TypeScript frontend represents a genuine engineering contribution.
The primary risks lie in the maturity of the RuVector ecosystem, the underspecification of the SONA adaptive learning model, and the challenge of delivering reliable multi-asset content extraction across the diversity of real-world email. The "10-minute inbox zero" claim, while plausible for well-structured inboxes with clear subscription and topic patterns, requires empirical validation through controlled user studies.
If the implementation achieves the stated performance targets --- particularly sub-50ms hybrid search, >95% categorization accuracy, and the ability to process 10K emails in under 3 minutes --- the platform would represent a significant advance in privacy-preserving email intelligence. The evaluation framework proposed in this paper provides a roadmap for rigorous validation of these claims.
Blondel, V. D., Guillaume, J.-L., Lambiotte, R., & Lefebvre, E. (2008). Fast unfolding of communities in large networks. Journal of Statistical Mechanics: Theory and Experiment, 2008(10), P10008.
Cormack, G. V., Clarke, C. L. A., & Buettcher, S. (2009). Reciprocal rank fusion outperforms condorcet and individual rank learning methods. Proceedings of the 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval, 758--759.
Dredze, M., Lau, T., & Kushmerick, N. (2008). Automatically classifying emails into activities. Proceedings of the 13th International Conference on Intelligent User Interfaces, 70--79.
Gerganov, G. (2023). llama.cpp: Port of Facebook's LLaMA model in C/C++. GitHub repository. https://github.com/ggerganov/llama.cpp
Grover, A., & Leskovec, J. (2016). Node2vec: Scalable feature learning for networks. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 855--864.
Guo, R., Sun, P., Lindgren, E., Geng, Q., Simcha, D., Chern, F., & Kumar, S. (2020). Accelerating large-scale inference with anisotropic vector quantization. Proceedings of the 37th International Conference on Machine Learning, 3887--3896.
Hamilton, W. L., Ying, R., & Leskovec, J. (2017). Inductive representation learning on large graphs. Advances in Neural Information Processing Systems, 30.
Hidasi, B., Quadrana, M., Karatzoglou, A., & Tikk, D. (2016). Parallel recurrent neural network architectures for feature-rich session-based recommendations. Proceedings of the 10th ACM Conference on Recommender Systems, 241--248.
Jacob, B., Kligys, S., Chen, B., Zhu, M., Tang, M., Howard, A., Adam, H., & Kalenichenko, D. (2018). Quantization and training of neural networks for efficient integer-arithmetic-only inference. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2704--2713.
Jegou, H., Douze, M., & Schmid, C. (2011). Product quantization for nearest neighbor search. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(1), 117--128.
Joachims, T. (2002). Optimizing search engines using clickthrough data. Proceedings of the 8th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 133--142.
Johnson, J., Douze, M., & Jegou, H. (2019). Billion-scale similarity search with GPUs. IEEE Transactions on Big Data, 7(3), 535--547.
Karpukhin, V., Oguz, B., Min, S., Lewis, P., Wu, L., Edunov, S., Chen, D., & Yih, W.-t. (2020). Dense passage retrieval for open-domain question answering. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, 6769--6781.
Kleppmann, M., Wiggins, A., van Hardenberg, P., & McGranaghan, M. (2019). Local-first software: You own your data, in spite of the cloud. Proceedings of the 2019 ACM SIGPLAN International Symposium on New Ideas, New Paradigms, and Reflections on Programming and Software, 154--178.
Klimt, B., & Yang, Y. (2004). The Enron corpus: A new dataset for email classification research. European Conference on Machine Learning, 217--226.
Li, L., Chu, W., Langford, J., & Schapire, R. E. (2010). A contextual-bandit approach to personalized news article recommendation. Proceedings of the 19th International Conference on World Wide Web, 661--670.
Malkov, Y. A., & Yashunin, D. A. (2020). Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(4), 824--836.
Mann, M. (2006). Inbox Zero. 43 Folders. https://www.43folders.com/izero
McMahan, B., Moore, E., Ramage, D., Hampson, S., & Arcas, B. A. (2017). Communication-efficient learning of deep networks from decentralized data. Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, 1273--1282.
Morris, J. X., Kuleshov, V., Shmatikov, V., & Rush, A. M. (2023). Text embeddings reveal (almost) as much as text. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing.
Perozzi, B., Al-Rfou, R., & Skiena, S. (2014). DeepWalk: Online learning of social representations. Proceedings of the 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 701--710.
Radicati Group. (2023). Email Statistics Report, 2023-2027. The Radicati Group, Inc.
Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., & Sutskever, I. (2021). Learning transferable visual models from natural language supervision. Proceedings of the 38th International Conference on Machine Learning, 8748--8763.
Reimers, N., & Gurevych, I. (2019). Sentence-BERT: Sentence embeddings using siamese BERT-networks. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing, 3982--3992.
Rendle, S. (2010). Factorization machines. Proceedings of the 2010 IEEE International Conference on Data Mining, 995--1000.
Rossi, E., Chamberlain, B., Frasca, F., Eynard, D., Monti, F., & Bronstein, M. (2020). Temporal graph networks for deep learning on dynamic graphs. ICML 2020 Workshop on Graph Representation Learning.
Thakur, N., Reimers, N., Ruckle, A., Srivastava, A., & Gurevych, I. (2021). BEIR: A heterogeneous benchmark for zero-shot evaluation of information retrieval models. Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks.
Tulving, E. (1972). Episodic and semantic memory. In E. Tulving & W. Donaldson (Eds.), Organization of Memory (pp. 381--403). Academic Press.
Wang, W., Wei, F., Dong, L., Bao, H., Yang, N., & Zhou, M. (2020). MiniLM: Deep self-attention distillation for task-agnostic compression of pre-trained transformers. Advances in Neural Information Processing Systems, 33.
Whittaker, S., & Sidner, C. (1996). Email overload: Exploring personal information management of email. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, 276--283.
Whittaker, S., Bellotti, V., & Gwizdka, J. (2011). Email in personal information management. Communications of the ACM, 49(1), 68--73.
Zuboff, S. (2019). The Age of Surveillance Capitalism: The Fight for a Human Future at the New Frontier of Power. PublicAffairs.