# Models, and which one to use

> What the model choice changes, what it does not, and how to keep the cost sensible.

Source: https://welcomeai.dev/docs/ai-models
Category: Start here
Updated: 2026-09-06

---
Two different models are involved in every answer, and they do unrelated jobs.

## The embedding model

It turns your content into numbers so that questions can find it. It runs once per chunk when a
document is indexed, and once per question. It never writes anything.

Changing it means re-indexing everything, because vectors from two different models are not
comparable. Pick one and leave it alone unless you have a reason.

## The answering model

It reads the passages retrieval found and writes the reply. This is the one you choose per knowledge
base, and the one your plan gates.

Bigger models write better: they follow instructions more exactly, handle a question that spans three
passages, and say "the content does not cover this" instead of inventing something. They also cost
more per answer and take longer to start.

**What the model does not change is what was retrieved.** If the right passage was never found, no
model can answer from it. Look at the retrieved context before you look at the model -- a bad answer is
far more often a retrieval problem, and moving up a model to fix one is expensive and does not work.

## The context window

Every model has a limit on how much it can hold in mind at once, called its **context window**.
Everything sent for one answer has to fit inside it: the passages retrieval found, the conversation so
far, your instructions, and the question.

You will rarely hit it. Top K caps how many passages go in, and a conversation is trimmed to the recent
turns before it is sent. It matters in two places: a very high Top K on a knowledge base of long
documents, and a chat that has been going for an hour. In both, the oldest part is dropped first, so an
assistant that "forgets" what was said at the start of a long conversation is not broken -- it ran out
of room, and a model with a larger window is the fix.

## A reasonable default

Use the cheapest model that answers your questions correctly, and check by asking the same real
questions in the Playground on each. For a well-organised knowledge base answering factual questions,
a small model is often indistinguishable. For questions that need several passages reconciled, or a
careful tone, the difference is obvious immediately.

## What it costs

Cost is per token, and both directions are counted: the question plus every retrieved passage going
in, and the answer coming back. Output is about five times the price of input.

That has one practical consequence: **Top K is a cost setting as well as a quality one.** Six passages
instead of three roughly doubles the input of every single answer. Billing and usage breaks the month
down by model so you can see where it went.

## Bringing your own

On the Bring Your Own plan you supply your own Anthropic and OpenAI keys and the provider bills you
directly at their prices. Everything else works the same. There is no fallback to our account: if your
key stops working the assistant says so rather than quietly running on ours.