Set up an agent

01Name it

How you will find it on the agents page. Versions you save later group under this name.

02Answer model

One model for now, chosen because it runs on a laptop rather than because it is the largest. It downloads once, on your users' first visit, and is cached after.

Model	Qwen3-0.6B, 4-bit GGUF
First-load download	~397 MB — about 30 seconds on broadband, then cached
Runs on	the browser — CPU, or WebGPU where available

03How it finds things

Both keyword and semantic is the usual answer — they fail in different places, and together they cover each other.

Keyword Exact terms, names, codes, error strings. Unbeatable when the user knows the word that appears in the document. Semantic Meaning rather than wording — finds the right passage when the question and the document share no vocabulary. Needs the embedder, which downloads alongside the model. Graph — optional Follows links between documents, so an answer can pull in what a passage refers to. Costs a build step over your corpus; leave it off until you have a reason.

04How much it reads, and how widely

The token budget caps how much retrieved text reaches the model per question — larger is more thorough and slower. Breadth is how readily the pipeline widens a thin result set.

05Prompt instructions — optional

Added to every question. Use it for tone, formatting, or domain vocabulary. It cannot remove the grounding rule: the agent answers only from retrieved passages and says so when they do not cover the question — that is what makes an answer trustworthy and it is not editable.