Concepts

Assistant

The top-level object you create in the app. One assistant has a name, brand colour, sources, scenarios, and produces one embed snippet. Most teams need one assistant per site; you can have several if you want different personalities or knowledge bases on different sites.

Source

A file or web page added to an assistant's Sources tab. Accepted file types include PDF, Word (.docx / .doc), Markdown, plain text, CSV, and Excel. URLs are fetched and crawled. Each source is split into chunks and embedded.

Chunk

A short passage of text extracted from a source. The embedder turns each chunk into a vector so the AI fallback can retrieve the most relevant passages for any visitor question (this is the “R” in RAG — retrieval-augmented generation).

Vector

The numeric representation of a chunk, used for similarity search. The collection of vectors for an assistant is published as vectors.json and cached in the visitor's IndexedDB the first time they open a page with the widget.

Scenario

A curated question/answer pair you author by hand in the Scenarios tab. Scenarios are matched first — before the AI runs — using a fast lexical search. They're how you guarantee an exact answer to your most important questions.

Variant

An alternative phrasing of a scenario's question. Each scenario can have several variants so the matcher catches different wordings of the same intent. The app can auto-generate variants for you.

Mode and engine

The widget picks its runtime on two orthogonal axes, automatically, based on a device-capability probe:

• Mode — desktop or mobile. Controls layout (floating panel vs fullscreen) and whether the device can carry a model.
• Engine — llm or scenarios. The llm engine runs Qwen3-0.6B in-browser (WebGPU when available, WASM otherwise) with scenarios as a fast short-circuit. The scenarios engine matches curated Q&A only — used on mobile and as a safety fallback when the LLM can't start.

Override the probe with data-mode-override or data-engine-override for QA. See Browser support for the full device matrix.

Build pipeline

The job that turns a source into a usable knowledge base. Stages:

1. Queued — waiting for a worker.
2. Parsing — extracting text from the file or URL.
3. Chunking — splitting text into retrievable passages.
4. Embedding — turning chunks into vectors.
5. Publishing — writing vectors.json to the CDN.
6. Done — ready for visitors.

Publish

Two things publish independently: the knowledge base (after every successful build) and the scenarios (when you click Publish on the Scenarios tab). Publishing emits new JSON files to the CDN that the embed downloads on its next load.

Config.json / scenarios.json / vectors.json

The three files the embed fetches per assistant. config.json carries branding and behaviour, scenarios.json holds the curated Q&A, and vectors.json holds the embedded chunks. Their URLs are derived from your data-assistant-id.