https://relu.chat/blog/ Technical articles about on-device AI, browser-based chatbots, NLP, and privacy-first machine learning. en Thu, 06 Aug 2026 13:16:22 GMT https://relu.chat/assets/logo.png https://relu.chat https://relu.chat/blog/retrieval-metrics-precision-recall-mrr/ https://relu.chat/blog/retrieval-metrics-precision-recall-mrr/ Thu, 06 Aug 2026 17:30:00 GMT How to evaluate a knowledge-base retriever with precision, recall, MRR, and NDCG: definitions, worked examples, and a tuning loop for dense/sparse ensembles. ReLU.chat retrieval ndcg mrr evaluation https://relu.chat/blog/chunking-strategies-knowledge-base/ https://relu.chat/blog/chunking-strategies-knowledge-base/ Thu, 06 Aug 2026 16:45:00 GMT How chunk size, boundaries, and overlap decide what a knowledge base retriever can find — fixed, overlapping, and structure-aware chunking compared. ReLU.chat chunking retrieval knowledge-base embeddings bm25 https://relu.chat/blog/query-expansion-without-llm/ https://relu.chat/blog/query-expansion-without-llm/ Thu, 06 Aug 2026 15:30:00 GMT Improve retrieval without an LLM: index-time alias weighting, query-side synonym expansion, edit-distance spelling correction, and boosting before re-ranking. ReLU.chat query-expansion bm25 spelling-correction retrieval https://relu.chat/blog/knowledge-base-fragments-fields-metadata/ https://relu.chat/blog/knowledge-base-fragments-fields-metadata/ Thu, 06 Aug 2026 14:00:00 GMT Knowledge base entries built from categorized fragments, weighted fields, and confidence metadata — and why structure decides retrieval quality. ReLU.chat knowledge-base retrieval metadata bm25 rag https://relu.chat/blog/indexeddb-client-storage/ https://relu.chat/blog/indexeddb-client-storage/ Thu, 06 Aug 2026 13:00:00 GMT Store embeddings as binary and session state as records with IndexedDB: schema design, vector search, transactions, and local-first privacy. ReLU.chat indexeddb storage embeddings local-first privacy https://relu.chat/blog/float32-vs-float64-precision/ https://relu.chat/blog/float32-vs-float64-precision/ Thu, 06 Aug 2026 12:30:00 GMT When to use float32 versus float64 in browser ML: tensors stay float32, accumulators like EMA summaries deserve float64, and int8 quantization shrinks weights. ReLU.chat float32 float64 precision quantization https://relu.chat/blog/browser-vector-math-optimization/ https://relu.chat/blog/browser-vector-math-optimization/ Thu, 06 Aug 2026 11:15:00 GMT What cosine similarity actually costs in the browser, and how precomputed norms, typed arrays, bounded top-k, and memoization cut that cost. ReLU.chat performance javascript vectors embeddings retrieval https://relu.chat/blog/offline-first-chat-architecture/ https://relu.chat/blog/offline-first-chat-architecture/ Thu, 06 Aug 2026 10:00:00 GMT Designing browser chat that works offline: persisted work queues, retries with backoff and idempotency, and sync-free design with one local copy of state. ReLU.chat offline-first service-worker indexeddb chat-architecture https://relu.chat/blog/model-size-vs-quality/ https://relu.chat/blog/model-size-vs-quality/ Thu, 06 Aug 2026 09:30:00 GMT Smaller can be better: distillation, quantization, and domain fit explain when a 22 MB model beats a much larger one on edge devices. ReLU.chat model-size distillation quantization edge-ai on-device https://relu.chat/blog/conversation-state-machines/ https://relu.chat/blog/conversation-state-machines/ Thu, 06 Aug 2026 08:45:00 GMT How modes, intents, and follow-up routing fit together in a conversation state machine, with JavaScript examples and a look at ReLU.chat's policy. ReLU.chat conversation state-machine intents routing chatbots https://relu.chat/blog/debugging-browser-ml/ https://relu.chat/blog/debugging-browser-ml/ Thu, 06 Aug 2026 07:30:00 GMT Practical techniques for debugging browser-based ML: profiling inference stages, structured on-device logging, and graceful fallback paths. ReLU.chat debugging browser-ml profiling fallbacks https://relu.chat/blog/bias-in-retrieval-chatbots/ https://relu.chat/blog/bias-in-retrieval-chatbots/ Thu, 06 Aug 2026 06:15:00 GMT Where bias enters retrieval chatbots — corpus, embeddings, term weighting, ranking — and a practical audit procedure for each stage. ReLU.chat bias retrieval evaluation nlp chatbots https://relu.chat/blog/privacy-checklist-browser-ai/ https://relu.chat/blog/privacy-checklist-browser-ai/ Wed, 05 Aug 2026 11:30:00 GMT What actually leaves your device when AI runs in the browser? A six-step audit: network, offline test, storage, and source. ReLU.chat privacy security checklist browser on-device https://relu.chat/blog/named-entity-recognition-no-llm/ https://relu.chat/blog/named-entity-recognition-no-llm/ Tue, 04 Aug 2026 12:00:00 GMT Gazetteers, regex, Levenshtein fuzzy matching, and knowledge bases deliver solid entity extraction in the browser — no LLM required. ReLU.chat ner nlp regex fuzzy-matching on-device https://relu.chat/blog/bm25-math-worked-example/ https://relu.chat/blog/bm25-math-worked-example/ Tue, 04 Aug 2026 11:00:00 GMT Step-by-step BM25 ranking: smoothed IDF, term-frequency saturation (k1), length normalization (b), and field weighting. ReLU.chat bm25 retrieval search sparse ranking https://relu.chat/blog/webassembly-browser-ml/ https://relu.chat/blog/webassembly-browser-ml/ Tue, 04 Aug 2026 11:00:00 GMT What WebAssembly is, why compiled numeric kernels beat JavaScript for ML, and how runtimes like ONNX Runtime run models entirely in the browser. ReLU.chat webassembly wasm onnx-runtime inference browser https://relu.chat/blog/latency-budgets-conversational-ui/ https://relu.chat/blog/latency-budgets-conversational-ui/ Mon, 03 Aug 2026 11:00:00 GMT First-token latency under 400 ms keeps a conversation feeling instant. Budgets, streaming, and progressive loading for conversational UI. ReLU.chat latency performance conversational-ui streaming ux https://relu.chat/blog/katex-vs-mathjax-math-rendering/ https://relu.chat/blog/katex-vs-mathjax-math-rendering/ Mon, 03 Aug 2026 10:30:00 GMT KaTeX vs MathJax for in-browser math: speed, bundle size, TeX coverage, accessibility, and streaming chat output. ReLU.chat katex mathjax math rendering latex https://relu.chat/blog/int8-quantization-explained/ https://relu.chat/blog/int8-quantization-explained/ Mon, 03 Aug 2026 09:30:00 GMT How INT8 quantization maps FP32 weights and activations to 256 buckets, why accuracy usually survives, and how calibration keeps it that way. ReLU.chat quantization int8 onnx optimization browser https://relu.chat/blog/cosine-similarity-vector-search/ https://relu.chat/blog/cosine-similarity-vector-search/ Sun, 02 Aug 2026 10:00:00 GMT The math of cosine similarity and vector search: dot products, normalization, brute-force top-k in the browser, and hybrid dense-sparse ranking. ReLU.chat vector-search cosine-similarity embeddings retrieval browser