AI Agents & RAG
Designing, building and running AI agents and retrieval-augmented generation systems, from tool calling and memory to vector databases, voice agents and Claude Code.
Categories in AI Agents & RAG
Learning series
Articles in AI Agents & RAG (page 1 of 7)
- RAG Knowledge Base Freshness: Staleness and Incremental Updates — A RAG system is only as current as its index. This guide covers detecting changes, updating incrementally, handling deletions, measuring staleness and telling users honestly when an answer may be out of date.
- Re-Embedding Migrations: How to Change Embedding Models Safely — Vectors from different embedding models cannot be compared, so switching models means re-embedding everything. Here is a migration plan with versioned storage, backfill, evaluation before cutover, a rollback path and a cost estimate.
- pgvector HNSW vs IVFFlat: Index Tuning for RAG in Postgres — An exact vector scan works until it does not. Then you must pick an approximate index and tune it. This guide compares HNSW and IVFFlat in pgvector, explains the knobs that trade recall for speed, and shows how to measure the result on your own data.
- Multimodal RAG: Scanned PDFs, Charts and Page Images — Plain text extraction fails on scans, charts and complex tables, so RAG answers quietly go wrong. This guide compares OCR, layout parsing and vision models, explains page-image retrieval, and shows how to test the whole pipeline.
- Multilingual RAG: Cross-Language Retrieval and Answers — Your documents are in one language and your users write in another, or both are mixed. This guide compares multilingual embeddings with translation, shows how to evaluate each language separately, and how to control the answer language.
- Query Rewriting, Multi-Query and HyDE: Better Retrieval for RAG — Users ask short, vague or context-dependent questions that embed poorly. Rewriting the query before retrieval can help a lot or make things worse. Here is when to use each technique, what it costs, and how to measure the difference.
- Verifiable RAG: Citations, Grounding and Knowing When to Say "I Don't Know" — A RAG system that answers confidently but cannot show its work is a liability. Citation, grounding and abstention are what make answers checkable. Here is how to build each, verify citations in code and measure whether it is working.
- Chunking Strategies for RAG: How to Split Documents So Retrieval Actually Works — When a RAG system answers badly, the cause is often the step tutorials mention in one line: how the documents were cut into chunks. This guide compares the main strategies, handles tables, code and PDFs, and shows how to test chunking with evidence.
- AI Agent vs Workflow: When Not to Use an Agent — Agents are impressive in demos and expensive in production. This guide defines the difference using primary sources, shows how reliability decays across steps, and gives a checklist for deciding when a plain workflow is the better answer.
- AI Won't Replace You. But People Who Know How to Automate With AI Might. — AI is not only about generating text. The bigger opportunity is AI plus data, tools, APIs, automation and human oversight. This flagship guide explains the difference between using AI and building AI automation, what agents and RAG really are, and how to start.
- When an AI Agent Has Too Many Tools: Fixing Tool-Selection Ambiguity — Give an agent three different ways to look up a customer -- by email, by ID, by name search -- and it will pick a different one each time for reasons that have nothing to do with which is actually best.
- When an AI Agent Calls a Tool That Doesn't Exist — A looping agent at least calls a real tool over and over. A more disorienting failure is an agent confidently invoking a function name that was never in its tool list at all.
- DeepSeek V4.1 Flash: A 552B MoE Model Built to Cut Agent Memory Costs — DeepSeek V4.1 Flash is a 552-billion-parameter Mixture-of-Experts model released September 10, 2026, built around one specific goal: cutting the memory cost of running long-lived AI agents by roughly 4x.
- Claude Fable 5.1 and Mythos 5.1: Why Anthropic Shipped Two Versions of the Same Model — Anthropic shipped Claude Fable 5.1 and Claude Mythos 5.1 on the same day as the same underlying model with different safeguard levels. Here is what actually changed, what the benchmark numbers mean, and who can access which version.
- Grok 4.6 and the Road to Grok 4.8: Inside xAI's New Release Cadence — Grok 4.6 is SpaceXAI's newest shipped flagship, and Grok 4.8 has already been named as the next model, skipping a public 4.7. Here is what actually shipped, what is still just a tease, and what the cadence itself signals.
- Gemini 3.8 Flash and the Cyber Variant: What Google Actually Shipped — Gemini 3.8 Flash is Google's third Flash release in six weeks, built on the 3.7 Flash base with a locked-down Cyber sibling. Here is what actually changed in capability, benchmarks, and pricing.
- Semantic Caching for AI Automation: Cutting Redundant LLM Calls Without Losing Freshness — Most AI automations re-ask the model questions it has already, functionally, answered before. Semantic caching AI automation pipelines catches the near-duplicates prompt caching misses — but only if you tune the similarity threshold instead of guessing at it.
- AI Outbound Voice Sales Automation: What an AI Cold-Calling Agent Can and Can't Do — An AI voice agent can dial a list all day without getting tired, but that is not the same as running a good cold-calling program. Here is what outbound voice AI is actually good at, where it needs a human handoff, and the compliance rules that don't go away just because a script is speaking instead of a person.
- WhatsApp Ecommerce Automation: Order Updates, Cart Recovery, and Support in One Thread — WhatsApp messages get opened at a rate email cannot match. The three automations, order updates, cart recovery, and support handoff, that make WhatsApp the primary post-purchase channel for an ecommerce store.
- ServiceNow AI Automation: Automating IT and HR Service Workflows With Now Assist — ServiceNow spent two decades as an IT ticketing system. In 2026 it repositions around AI agents that take action directly across ITSM, HRSD, and CSM -- and opens that system to outside AI agents through a new MCP server.
- GoHighLevel AI Automation: Turning Sub-Accounts Into an AI-Powered Growth Engine — GoHighLevel bundled five separate AI products into "AI Employee" in 2026. A practical look at what Voice AI, Conversation AI, and Reviews AI actually automate, what it costs, and where agencies still have to build the guardrails themselves.
- Sandboxing AI Agent Code Execution Safely — An agent that can write and run its own code is far more capable than one limited to a fixed set of tools -- and far more dangerous if that code runs with the same access as everything else on your system. Sandboxing is what makes the capability safe to actually use.
- Plan-and-Execute vs ReAct: Two Approaches to How an AI Agent Reasons — ReAct interleaves thinking and acting one step at a time, reassessing after every tool call. Plan-and-execute commits to a full plan up front, then works through it. Neither is universally better -- they trade off differently on cost, latency, and how well they handle a task that changes shape mid-way.
- Long-Running AI Agents: Designing for Tasks That Take Hours, Not Seconds — Most agent frameworks assume a request-response shape -- a user asks, the agent works for a few seconds or minutes, and returns an answer. A genuinely long-running agent task, spanning hours or requiring a wait for an external event, breaks that assumption at almost every level of the stack.