Kubai Kevin

37 articles tagged Vector Search

BM25 vs dense embeddings: hybrid search reality

How BM25 and dense retrieval fail differently, why score fusion is not free, and how to choose RRF, weighted fusion, or a cascade for your corpus.

14 min read · 3 November 2026

3 LLM context techniques: when each fails

Compare fine-tuning, retrieval-augmented generation and long context by failure mode, cost arithmetic and a decision checklist, with code for each.

15 min read · 1 November 2026

7 IDP shifts for AI feature delivery

Seven platform shifts that let teams version prompts, pin models, run offline evals, attribute token cost, and roll out AI changes gradually.

15 min read · 10 October 2026

World models won't replace your backend

World models add a high-rate prediction stream to your system. Treat it as a sensor: schema, retention, backpressure, and rollback, not a new stack.

15 min read · 24 September 2026

Golden paths become load-bearing walls

Why thick internal SDKs for AI features become migration hazards, and how to design a thin runtime that enforces model ID, prompt version and embedding dimension.

13 min read · 21 September 2026

Agent drift: the failure mode you didn't log

Agent drift is a distribution failure, not a model failure. How to detect it with categorical signatures and PSI, and contain it with state guards.

15 min read · 18 September 2026

Designing a Schema That Survives Embedding Model Upgrades

Why embedding columns break on model upgrades, how to version vectors, migrate legacy JSON payloads, and when to offload to a vector store.

11 min read · 13 September 2026

IDP for AI: 2026 stack evolution

How to wire embedding model changes into an internal developer platform so dimension drift fails loudly instead of silently shipping broken indexes.

14 min read · 2 September 2026

Mobile-first AI fails you didn’t plan…

Mobile-first AI features fail on payload size, latency and device RAM, not model quality. A practical guide to measuring and fixing the network and device bottlenecks.

13 min read · 26 August 2026

AI errors hide in plain latency

Why token counts and GPU utilization miss user-visible AI degradation, and how to instrument end-to-end p95/p99 latency and error rate instead.

10 min read · 21 August 2026

AI unit costs: the metric both teams fight over

Why per-call AI cost estimates break down, and how to instrument a shared cost-per-outcome metric that product and finance both accept.

14 min read · 18 August 2026

Buy vs build AI: 3 hard rules for 2026

A framework for deciding which AI components to build versus buy, with failure modes, measurement methods, and code for caching and retries.

15 min read · 11 August 2026

AI cost attribution without the spreadsheet hell

How to attribute AI inference costs to users, features and models: instrumentation, cardinality limits, sampling error, and a 30-minute first step.

15 min read · 7 August 2026

RAG rots if you don’t refresh indexes

RAG indexes rot when source documents change. Learn change detection, chunk versioning, and incremental refresh patterns, with code and a cost model.

15 min read · 4 August 2026

Prompt Injection Is Not Your Biggest LLM Data Risk

Prompt injection dominates LLM security checklists, but poisoned training data and label drift often cause more damage. A pipeline-stage model for triage.

17 min read · 3 August 2026

When a vector DB misconfiguration breaks RAG silently

Vector DB misconfigurations cause RAG outages that look like app bugs. Learn to detect index errors, add circuit breakers, and instrument query latency.

11 min read · 29 July 2026

Postgres vector search: what breaks first

pgvector works until load grows. Here is what fails first, how to measure it, and when a dedicated ANN service earns its operational cost.

13 min read · 24 July 2026

Purpose-built beats general AI tools

When a general AI platform beats a purpose-built stack, how to choose, and how to measure latency, cost and fallback behaviour yourself.

16 min read · 19 July 2026

Agent context: short-term vs long-term memory

Compare Redis-style short-term agent context with Postgres event logs: latency, durability, auditability, cost, and how to measure each.

15 min read · 9 July 2026

Observability for LLM Features: Metrics That Drive Decisions

Log one structured event per LLM interaction instead of tracing every token, then alert only on user-facing outcomes that justify keeping a feature.

14 min read · 8 July 2026

LLMOps 2026: evaluation stack that replaced RAG dashboards

Why RAG dashboards miss real failures, and how to build an evaluation stack around task resolution, dataset rotation, prompt versioning, and drift checks.

13 min read · 5 July 2026

AI in slow networks: why models flop where data costs rise

How to cut payload size, round-trips and cost for AI features on 2G/3G networks, with a worked ASR example, failure modes and a measurement checklist.

14 min read · 5 July 2026

LLM drift: the silent quality drop metrics miss

Latency and error rate stay flat while LLM output quality quietly degrades. Here is how to log prompts, measure embedding drift, and catch it early.

11 min read · 4 July 2026

Debug AI agents: failures aren't just bugs

AI agent postmortems need different questions than service incidents. Compare structured failure narratives with LLM-assisted causal graphs, and when each wins.

11 min read · 1 July 2026

9 Vibe coding traps that break production code

Nine patterns that make AI-assisted prototypes fail under load, with the instrumentation to catch each one before it reaches production.

11 min read · 1 July 2026

RAG on legacy systems: vector DB vs. full-text SQL

Compare vector databases and in-database vector search for adding RAG to legacy systems, with latency, cost, and operational trade-offs.

14 min read · 30 June 2026

Choosing an AI Pair Programming Setup: A Practical Guide

How to evaluate AI pair programming tools by latency, signal-to-noise, cost, and privacy, with configs, failure modes, and a decision checklist.

15 min read · 29 June 2026

AI observability: the metrics that matter

AI observability needs drift, confidence, and attribution metrics. Covers instrumentation, cardinality limits, failure modes, and when not to build it.

13 min read · 28 June 2026

AI apps break when you treat them like regular software

AI apps fail differently from CRUD services: state accumulates, caches churn, and model outputs drift. Here is how to design, measure, and bound those failure modes.

15 min read · 27 June 2026

RAG is not enough: hybrid search + reranking in 2026

Dense retrieval alone ranks poorly. A two-stage pipeline of BM25 plus a cross-encoder reranker fixes it. Architecture, code, and how to measure it.

12 min read · 27 June 2026

AI observability isn’t monitoring — here’s why

Traditional monitoring proves a service is up; AI observability proves its outputs are still correct. Layers, code, drift math, and failure modes.

14 min read · 25 June 2026

LLM evals: skip the hype, track these 4 metrics

A practical guide to production LLM evaluation: four metric classes, code for deterministic validators, and how to detect prompt drift and scorer decay.

11 min read · 22 June 2026

LLM metrics that stop wasting time

A practical guide to LLM evaluation in production: grounding checks, drift detection, cost and latency metrics, and how to measure each one.

13 min read · 14 June 2026

Which AI wrappers made money in 2026

Most AI wrapper products fail on latency, cost, and churn rather than model quality. Here is a rubric, failure analysis, and code to build one that lasts.

10 min read · 12 June 2026

Build a search system in 2026: semantic, keyword

How to build a hybrid search service with PostgreSQL, pgvector, and Redis: schema, keyword and semantic queries, reranking, caching, and how to measure recall.

13 min read · 10 June 2026

RAG pipelines: 4 silent prod killers

A practical guide to the silent failure modes that break RAG pipelines at scale: tokenizer drift, GPU fragmentation, index skew, and cache stampedes.

13 min read · 8 June 2026

RAG in prod: what tutorials won’t tell you

Why nightly embedding refreshes break RAG latency, and how a hot/cold retrieval split with partial vector updates fixes it.

14 min read · 8 June 2026