Ask the way you talk
People type in Bangla script, English, or Banglish — sometimes all three in one chat. Ansary AI answers in the same style you asked in, so nobody has to switch to "software English".
- Replies mirror the user's language and tone
- Follow-up questions keep the context ("and in September?")
- Tables and lists come back clean, ready to read
Answers from the real reports
The model never guesses a number. It decides which report to open, the portal runs its own report code, and the assistant reads the result back — shipped value, share, change between quarters.
- Quarter-on-quarter comparisons in one question
- Top buyers, new orders, closing balance, KPI scores
- Numbers match what the page itself shows
From a question to a grounded answer
Every reply goes through the same short loop — and the loop is bounded, so a confused model can't spin forever.
You ask
Text, voice or an attached file
Model plans
Picks a tool, or answers directly
Tool runs
Portal's own report code, with your permissions
Model reads
Up to six rounds if it needs more data
Plain answer
In your language, with a Listen button
The 18 tools. Gold = sensitive, enforced by role.
Talk to it. Hand it a file.
Dictate a question with the microphone and have the answer read aloud with Listen. Attach a picture, PDF, Excel or Word file and ask about it — all inside the same chat.
- Voice input and spoken replies
- Images, PDF, Excel and Word attachments (5 MB)
- "New chat" clears the conversation in one tap
Built to keep working
If one AI provider is busy or out of quota, the chat quietly continues on the other. The page lets an administrator update provider keys without touching the server.
- Gemini first, Groq as automatic fallback
- Model names auto-discovered, so a retired model doesn't break chat
- Management-only access; sensitive tools check the user's role
Small stack, careful rules
Tool calling
Eighteen declared functions with typed arguments. The model asks; the portal executes and returns data only.
System prompt
Mirrors the user's language, states what it can't do, and is told never to invent a figure it didn't get from a tool.
Permissions
Checks live inside each tool, so a clever prompt can't reach data the signed-in user couldn't open themselves.