Bot Data Strategy: Real-time Skills vs Ingested Knowledge

Your bot can answer questions in two completely different ways. Picking the right one for each kind of question is the single biggest decision you'll make when setting up your bot. This guide will help you choose.

Real-time API Skills vs Ingested Knowledge — when to use each

Real-time API skills (sometimes just called "skills") let the bot reach into a connected system — Shopify, HubSpot, your CRM — and pull a specific record on demand. The bot calls an API, gets back a single answer, and reads it to the customer.

Ingested knowledge is different. You upload a catalog, a price book, a product manual, or a CSV, and Velaro processes the whole document into a searchable index that the bot can read all at once. The bot can then compare, rank, and summarize across every item — not just look up one.

The simple rule: look up one thing → use a skill. Compare or rank many things → ingest the catalog.

If you have 10,000+ products and need comparison or ranking — ingest your catalog

Real-time skills are great at fetching a single record, but they cannot compare or rank items. The bot would have to call the skill thousands of times to look at every product, which is slow, expensive, and unreliable.

If your customers ask things like:

• "What are your top 5 most durable kitchen knives?"
• "Which laptops under $1,500 have the best battery life?"
• "Compare your three best-selling running shoes."
• "Which of your products work for sensitive skin?"

...you need ingested knowledge. Upload your full catalog (CSV, JSON, or PDF) on the bot's Knowledge tab. The bot can then read across every product at once and produce a ranked, comparative answer.

If you just need lookups (price check, order status, stock) — real-time skills are better

Skills shine when the answer is exactly one record and that record changes minute-to-minute. Stock counts, order status, balance due, recent orders — these change constantly, so reading from your live system is the right call.

Common skill-shaped questions:

• "Do you have the iPhone 15 in stock?"
• "What's the status of my order #12345?"
• "How much do I owe on invoice 7821?"
• "What's the price of SKU ABC-100?"

For these, leave the catalog alone — connect the integration (Shopify, NetSuite, etc.) and let the bot use the skill.

Concrete examples — which to use

You can do both — and most successful bots do

Most production bots use a mix: ingested knowledge for catalog comparisons, policies, and FAQs; real-time skills for stock, orders, prices, and account data. You don't have to pick one.

If you're unsure which to use for a particular question, check your bot's Unanswered Questions dashboard. If you see a lot of "needs ingestion" entries, that's a signal the bot is being asked comparison-style questions and needs catalog knowledge added.

How big does my catalog need to be to ingest it?

There's no hard minimum, but the rule of thumb is: if you have more than 100 items and customers ever ask aggregate questions ("top," "best," "compare," "which"), ingest it. Below 100 items, skills usually cover everything. Above 100, comparison questions become very hard for skills to answer well.

Where to upload

Go to your bot's Knowledge tab. You can upload CSV, JSON, PDF, or paste text directly. Velaro processes the file in the background and the bot will start using it within a few minutes. There's no extra coding required.

How your bot finds the right answer in your knowledge

When a customer asks a question, your bot does not simply guess. It runs a three-stage retrieval process across your knowledge before it writes a single word:

1. Knowledge Overrides — pinned, authoritative facts you have set. These always take priority. If you have told the bot "our holiday hours are X," that wins over anything else.
2. Vector knowledge search — a semantic search of your ingested content. This is not just keyword matching: it understands intent, so "do you have anything for sore feet" can match a product described as "plantar fasciitis support."
3. Help Center articles — your published help articles, matched to the question.

Results from all three are combined and given to the AI as grounded context. The bot then writes its answer from that context — not from the open internet, and not from guesswork.

Why this is more accurate than a generic chatbot

Generic AI chatbots answer from whatever the language model happened to learn during training. That produces confident-sounding answers that can be out of date or simply wrong — the well-known "hallucination" problem.

Velaro grounds every answer in your content. The bot retrieves your real policies, your real catalog, and your real help articles, then answers from those. The search is hybrid — it combines exact keyword matching (so a specific model number or SKU is found) with semantic similarity (so intent-based questions are understood). Most chatbots do only one or the other; doing both is what catches the questions a single method would miss.

See exactly where every answer came from

Open Test Your Bot and ask any question. Each reply shows a retrieval trace: which of the three stages fired, which documents were retrieved, and which knowledge index served them. If the bot answered from your content, you will see the sources. If nothing matched, the trace says so plainly — which tells you the topic needs content added rather than leaving you guessing.

This transparency means you are never wondering whether the bot is grounded in your knowledge or improvising. You can see it, on every answer, before a customer ever does.