AI features
Conversational replies, voice transcription, music recognition, and the two picture lookups
Five things in Cookiebot call something outside Telegram. They are grouped here because they share a property nothing else in the bot has: they cost money per use, and they can be unavailable while the rest of the bot is fine.
Talking to the bot
Mention the bot by name, or reply to one of its messages, and it answers in character. No command, no prefix.
- Seven consecutive triggers is the cap. The first six get an answer; the seventh gets nothing. Every ordinary message the same person sends puts one back. It stops one member turning the bot into a private chat toy without the bot ever having to say no.
- A message that is only the bot's name gets
?— and no model is asked, so an empty prod costs nothing. - Other handlers win. If you reply to the
/newrulesprompt with something that also mentions the bot, that is a rules reply, not a conversation. - With Fun Functions off, it goes quiet rather than announcing that fun is disabled — a chat reply refusing to chat is worse than silence.
Replying to the bot with a sticker gets a sticker back instead, from the pool it has collected in safe-for-work groups.
Voice notes
Two different things happen to a voice note, and they are independent.
A voice note that replies to the bot is transcribed and answered like any other message. The transcript itself is never posted — you get the reply, not a readback. Notes longer than 300 seconds are refused rather than truncated.
/transcribe (also /transcrever, /transcribir), sent as a reply to a
voice note, posts the transcript in the chat. Sent without replying to
anything, it says so. This one follows Utility Functions.
Music recognition
Every voice note in the group can be fingerprinted against Shazam, and the bot answers when it recognises the song. Nothing is posted when it does not.
This is off unless the operator turns it on (CB_MUSIC_DETECTION_ENABLED),
because it calls an unofficial API on the busiest path in the bot. It also
follows Utility Functions. Neither switch produces a refusal message: a
voice note never asked the bot for anything.
Finding where a picture came from
/searchsource/buscarfonte/buscarfuenteReply to an image with it, and the bot looks for the original.
A confident match comes back with the title, the author and a link. Nothing found is reported as "this looks original" — the bot does not guess.
Sent without a picture to work on — no reply, or a reply to plain text — it answers with the instructions instead. There is a daily allowance, and the bot says when it is spent rather than failing quietly. It follows Utility Functions.
Any unknown command is an image search
/french fries is not a command, so the bot searches Google Images for
"french fries" and posts what it finds. This is v1's catch-all, kept exactly:
- Safe search follows the group. A group marked safe-for-work gets filtered results; one marked otherwise does not.
- Two daily limits apply — one per person, one across the deployment. The request that crosses the line is the one refused, and the bot says so.
- Real commands are never hijacked.
/isaliveanswers as itself. - Nothing happens for a command carrying a link, a command addressed to another bot, or a word on the blocklist. No search, no message.
/anythingon its own (also/qualquercoisa,/cualquiercosa) explains the shape rather than searching for nothing.
It follows Utility Functions, and with that switch off it is silent.
What this costs, and who chooses
For a group: nothing to configure beyond the two family switches.
For an operator: the model behind every AI task is configuration, not code — provider and model are chosen per task, tokens and cost are recorded per call, and a provider that starts failing is cut off rather than retried into the ground. A refusal from a model comes back as a refusal, not a crash. The architecture page has the details.