LanguageENES
WelcomeAI
Blog

Notes on retrieval, support and shipping

How we build WelcomeAI, and what we learn from the teams running knowledge bases on it. No growth hacks.

All33Product5Retrieval7Engineering3Guides6AI Search3Comparisons3Customer Support6
FeaturedRetrieval

How to fix bad retrieval scores by changing chunk size

When your assistant cites the wrong page or misses details, the problem is usually chunk size, not the model. Here's how to read scores and pick a better one.

Mauro CarreraSep 2, 2026 · 5 min read
RetrievalWhat an embedding is, and why two sentences with no shared words land togetherAn embedding turns a sentence into a position. Everything an assistant does next is geometry: what sits near the question gets read, and what sits far away never arrives.
Hernan MoyaOct 19, 2026 · 6 min
RetrievalWhat a vector database does, and why we still run PostgresEvery RAG tutorial opens by telling you to pick a vector database. Here is what one actually does, what Postgres already does, and the three conditions that make a separate system worth running.
Hernan MoyaOct 16, 2026 · 6 min
Customer SupportPlugin or script tag? Two ways to put AI chat on WordPressThe question sounds like a choice between two installation methods. It is not. One of them carries your content in and the other puts the chat on the page, and most sites want both.
Hernan MoyaOct 15, 2026 · 6 min
Customer SupportThe questions car buyers ask a dealership at 10 pmCar buying happens in the evening, on three tabs at once. The listing answers two questions out of five, and the showroom that answers the other three before morning gets the visit.
Hernan MoyaOct 14, 2026 · 6 min
Customer SupportAllergens, access and pets: hospitality questions with a liability edgeThree questions decide a booking and carry a consequence if the answer is wrong. An assistant can take all three, on one condition: somebody has to have written the policy down first.
Hernan MoyaOct 12, 2026 · 7 min
GuidesHow to write documentation an AI can actually answer fromRetrieval does not read your page. It reads a piece of it, chosen by a machine, with no memory of the heading six paragraphs up. Writing for that changes a few habits, not your whole style guide.
Hernan MoyaOct 9, 2026 · 7 min
RetrievalSemantic search vs keyword searchOne finds documents that share your words. The other finds documents that share your meaning. Both fail in ways the other does not, and knowing which is which decides what your search gets wrong.
Hernan MoyaOct 8, 2026 · 6 min
RetrievalStop your assistant mixing up product versionsTwo manuals that differ in one step are nearly identical to a retrieval system. It will find both, rank them the same, and blend them into an instruction that is right for neither device.
Hernan MoyaOct 7, 2026 · 6 min
GuidesHow long does it actually take to ship an AI chatbot?Installing takes five minutes and nobody believes it, because the thing that takes time is not installing. It is the state your content is in, and the people who have to approve it.
Hernan MoyaOct 6, 2026 · 6 min
Customer SupportWhat an AI assistant can and cannot see in your Shopify storeYour catalogue, policies and shipping rates reach the assistant. Your customers and orders never do. Here is the exact boundary, and the one thing that goes stale if nobody presses a button.
Hernan MoyaOct 5, 2026 · 6 min
Customer SupportWhy shoppers abandon carts over unanswered questionsThe cart was not abandoned at checkout. It was lost earlier, on the product page, at the moment a question arrived and nothing on the page answered it.
Hernan MoyaOct 2, 2026 · 7 min
Customer SupportWhat a clinic can and cannot automateMost of what a clinic is asked after hours is written down somewhere. The rest is clinical, and an assistant that knows the difference is the only kind worth putting on a practice website.
Hernan MoyaOct 1, 2026 · 7 min
ComparisonsAI chatbot pricing models compared: per seat, per message, per tokenFour ways vendors charge for an AI assistant, and what each one does to your bill in month three. The shape matters more than the sticker price.
Hernan MoyaSep 30, 2026 · 6 min
ComparisonsRAG chatbot vs rule-based chatbotA flow builder answers what you drew. Retrieval answers what you wrote. The difference shows up in the questions nobody predicted, and in what it costs to keep either one correct.
Hernan MoyaSep 29, 2026 · 7 min
ComparisonsAI chatbot vs live chat: when each one winsThe question is rarely which tool to buy. It is which questions belong to an assistant, which belong to a person, and what happens at the moment one hands over to the other.
Hernan MoyaSep 28, 2026 · 7 min
AI SearchShould you let AI crawlers index your site?Three different things get called AI crawlers, and they take different decisions. Here is what each one does, what blocking them really buys, and the policy we run on our own two hosts.
Hernan MoyaSep 25, 2026 · 6 min
AI SearchHow to get your documentation cited by ChatGPTBeing cited takes two things: a page an assistant is allowed to fetch, and a passage it can lift without the rest of the page. The second one is where most documentation fails.
Hernan MoyaSep 24, 2026 · 7 min
AI SearchWhat is llms.txt, and should your site have one?A Markdown map of your site, written for assistants instead of crawlers. What goes in it, what actually reads it today, and when your site is better off without one.
Hernan MoyaSep 24, 2026 · 8 min
ProductWhatsApp is next: your assistant, under your business's nameYour assistant is coming to WhatsApp. It will answer from your own number, under your business's name, and every conversation will start with the customer's phone number already on the lead.
Hernan MoyaSep 22, 2026 · 6 min
ProductIntroducing WelcomeAI for Instagram: your assistant, in your DMsOn 15 October your assistant moves into Instagram Direct Messages: answers from your own content, a voice written for a phone, and conversations that end as leads.
Hernan MoyaSep 21, 2026 · 7 min
RetrievalFine-tuning vs RAG: which one do you actually need?Fine-tuning changes how a model writes. Retrieval changes what it can read. For company knowledge that difference settles the question almost every time.
Hernan MoyaSep 18, 2026 · 6 min
GuidesHow to add an AI chatbot to your websiteInstalling takes one line of code. Four decisions before it (surface, pages, sources and domains) decide whether anyone uses it.
Hernan MoyaSep 17, 2026 · 7 min
EngineeringWhy we serve citations instead of confidence scoresA percentage next to an answer looks rigorous and measures the wrong thing. A citation shows the evidence, and anyone can check it.
Hernan MoyaSep 16, 2026 · 5 min
ProductHow to measure whether your AI chatbot is actually workingMost chatbot dashboards are built to make a bad week look good. Four numbers tell you whether the assistant is doing its job, and none of them is on the default view.
Hernan MoyaSep 15, 2026 · 5 min
ProductHow to stop your AI chatbot from hallucinatingA hallucination is not the model lying. It is the model filling a gap you left, and the fix is upstream of the prompt.
Hernan MoyaSep 14, 2026 · 5 min
GuidesHow to train an AI chatbot on your own contentEvery AI chat product says you can train it on your own content. The word is wrong, and believing it changes what you expect it to cost.
Hernan MoyaSep 11, 2026 · 5 min
RetrievalWhat is RAG? Retrieval-augmented generation explainedA language model does not know your refund policy. RAG is how it reads your documents at the moment of the question, and answers from what it read.
Hernan MoyaSep 10, 2026 · 7 min
EngineeringThe cost of an AI answer, broken downThe expensive part of an AI answer is not the answer. It is the several thousand tokens the model reads before writing three hundred.
Hernan MoyaSep 9, 2026 · 9 min
GuidesHow to build an AI knowledge base for your businessMost knowledge bases get worse as they grow. Start from the twenty questions people actually ask, index in the right order, and test it before your customers do.
Hernan MoyaSep 9, 2026 · 7 min
GuidesEight questions to ask before you buy an AI chat toolEvery AI chat product demos well. These are the eight questions that tell you what it will do in month three, plus a scorecard to run live in the demo.
Mauro CarreraSep 7, 2026 · 8 min
ProductWhy your AI assistant must cite every answerCitations are not just a trust signal. They let your team find and fix answers that are wrong before a customer acts on them.
Mauro CarreraAug 18, 2026 · 4 min
EngineeringWhat actually happens when you paste the script tagThe widget is 20 KB, renders in a shadow root, and never sees your page's CSS. Here is the whole path from tag to answer.
Mauro CarreraJul 29, 2026 · 4 min