The AI Assistant
,
end to end.
A tour through the new AskDona AI Assistant — what changed, why, and what it can do now.
Hi there!
I'm Dona, an AI assistant. I generate answers from documents in the knowledge base.
Here's what we'll cover.
A new look isn't enough — How it moves, how it feels, and how it works underneath.
AskDona is our enterprise RAG platform. The AI Assistant is its embeddable version — a widget customers drop onto their own sites for internal and public users.
What is an AI Assistant — and why AskDona?
A quick grounding before the redesign tour starts.
An AI assistant that lives on a website and answers visitors’ questions in natural language — instantly, around the clock.
Cuts repeat support tickets, makes long documents discoverable, and gives users help at 3am — without growing the team.
Answers from customer documents, not the public web. Bilingual JP/EN, themeable, citation-rich.
<script src="…/askdona.js"></script><script>AskDona.init({chatflowId: 'your-chatflow',displayMode: 'modal', // or 'fullscreen'primaryColor: '#…', // your brand color});</script>
one script tagDrop it on any page — no build step, no framework required.modal / fullscreenShow it as a floating launcher or take over the whole page.your brandOne brand color plus JP / EN — it adapts to match your site.your knowledgePoint it at your own documents; every answer comes from your content.A single tab, three ways to ask.
Users move between Chat, Search, and Deep Research without ever leaving the widget — shown as a floating modal or expanded to fullscreen. One consistent experience, whichever you choose.
Simple. Smart. Expert.
One toggle, three response styles. The chat mode flows down to the backend so the model, retrieval depth, and reasoning budget all change accordingly.
Every client’s widget feels like theirs.
Pass any brand color and the widget makes it its own — it automatically picks readable text and matching shades across every surface.
It notices. It asks. It doesn’t push.
If a user is in EN mode but types in Japanese (or vice versa), a small banner offers to switch — once. Ignore it and it goes away.
Motion that earns its keep.
Every animation does one of three jobs — otherwise we cut it.
A small ding when you’re looking away.
If the user has switched to another tab while waiting on an answer, the AI Assistant pings them when the response finishes streaming — so they don’t miss it.
Only when you’ve switched away to another tab while waiting. If you’re looking right at the chat, it stays silent.
Quiet, short, and never repeats. Designed to gently let you know your answer is ready — a nudge, not an interruption.
Watch the answer take shape.
As your answer is being written, a live hand-drawn animation shows it coming together in real time.
One launcher. Eight host palettes.
Same widget code on any host site — the launcher takes on each brand's color.
A guided tour that knows when to leave you alone.
First-time users get a short, spotlighted walkthrough of every feature the admin has enabled — and nothing they haven't.
A trash icon that asks before deleting.
We replaced the old refresh icon — users were clicking it expecting a reload and losing their entire conversation without warning.
Feedback you can take back — until you commit it.
Thumbs up / down on any assistant reply. The selection is editable until the user clicks Submit, so a mis-click isn't punished.
Copy the format that matches where it's going.
Two-option copy on every assistant reply — raw markdown for docs and Notion, plain text for Slack and email.
Sources, on demand — or always on.
Old design had citations everywhere — both inline and in a long list. Feedback was mixed: some loved the auditability, others found it noisy.
When a user says “Hi” — we suggest what to ask.
Some users open the chat and type a greeting, not a real question. Instead of forcing a useless answer, the AI Assistant follows up with three suggested questions tuned to the chatflow’s content.
A short heuristic flags greeting-only or filler messages — hi, hello, thanks, cool — before they hit retrieval. Real questions skip this path entirely.
The AI Assistant offers the same three example questions configured for the chatflow’s landing screen — recognizable to the user, and known to work against the corpus.
Suggested questions render as chips. Tapping one fires it as a real user message — no extra confirm step. Users can also ignore the chips and keep typing.
A full RAG pipeline for “hi” would burn tokens and make the AI Assistant look bad. Suggested questions stop the request before it goes deep.
Open filters from one button.
A single Filter pill in the chat surface opens a focused popup — everything filterable lives inside.
The popup at work.
Filter relationship, search-as-you-type, and a live preview of what’s selected.
A two-button toggle at the top — AND means match all conditions, OR means match any. Within a single key, values are always OR’d (e.g. Department: Sales OR Marketing). The AND/OR toggle only governs the relationship between keys.
Type Batch — matching values surface in their parent key, with the matched text highlighted.
Picking a key promotes it into the right pane. Selected values stay highlighted; the summary line at the bottom names exactly what got picked.
Header gains a (N conditions) badge as soon as anything is active — instant scanability before Apply.
If the filter combination matches no documents, the popup tells you immediately and suggests relaxing your filters — so you never wait on a search that can’t return anything.
What would you like to investigate?
Deep Research is a separate mode for the hard questions — the ones that need clarifying first, then a thorough investigation across many documents. The AI Assistant asks a few follow-up questions, then investigates and writes a sourced report.
A form that doesn’t look like a form.
For complex questions, Deep Research asks a few follow-ups first — as a turn-by-turn conversation that asks the next question only after your previous answer lands. It feels like talking, so it’s easy to finish.
I’ll ask you 5 follow-up questions. Each answer helps us run a more accurate investigation, but every question is optional — press “Skip this question” to move on, or “Skip questionnaire” to skip the entire follow-up.
Ask a hard question. Walk away. Come back to a report.
Deep Research doesn't pretend to be instant. It asks for clarification, works through your sources, shows its progress, and produces a sourced report with a one-click PDF.
Close the tab. Come back. Done.
A multi-minute research run doesn't depend on you keeping the tab open. The work keeps going on its own — open the chat on a different device and the result is waiting.
A report worth keeping.
Every research run produces an artifact users can share, archive, or paste into a doc. The report includes the original Q&A history so readers can audit how the question was scoped.
· APAC region grew fastest, +34% QoQ[3]
Q: Subsidiaries included? → Parent only
Q: Quantitative comparison? → Yes
What are you looking for?
Search is a second tab in the AI Assistant — same content, same filters, but it gives you the document instead of a generated answer. It surfaces the matching passages directly, with your search terms highlighted.
Search — find the doc, not the answer.
A second tab that swaps the generated answer for a direct document lookup. Same content, same filters — just results with highlighted matches and links into the source.
When Search is on, it sits right next to Ask AI as a separate tab — switch between a generated answer and a direct document lookup in one click.
Search gives you the source material directly instead of a written answer. Results show the matching passages with your search terms highlighted.
The categories you pick in the chat carry over to Search. Filter your content and your results narrow to match — both linked pages and documents surface here.
Ask AI — you want an answer in your own words.
Search — you know there’s a doc that says it, you just need the page.
Thank you.
Questions, demos, or wild ideas — let's talk.