AI Context from Telegram Messages

This reference is about extracting useful Telegram context for AI features.

It is not a general persistence recipe.

Goal

When you pass Telegram data into an LLM, do not dump the raw Message object.

Instead, extract only the fields that change interpretation.

Good AI context is:

• compact
• explicit
• structured
• stable across message types

The most useful signals

1. Visible user text

Usually the first field to extract:

• message.text
• message.caption
• normalized HTML or markdown text if your bot preserves formatting

Use this as the main body of the AI input.

2. Quote or reply context

For modern Telegram clients, visible quoted text can arrive through message.quote.text.

This matters when the user is replying to something and the quote changes meaning.

Useful fields:

• quote text
• whether the message is a reply
• reply target message id when available
• thread or topic id when relevant

3. Poll data

Polls are useful AI context and should not be treated as empty messages.

Useful fields:

• poll question
• option texts
• poll type
• whether multiple answers are allowed
• whether the poll is closed
• explanation text when present

For AI formatting, keep only semantic fields, for example:

Poll question: Which feature should we ship first?
Options: Notifications; Calendar sync; CSV export
Poll type: regular
Multiple answers allowed: false
Closed: false

4. Sticker signals

Stickers often have little text, but they still carry meaning.

Useful fields:

• sticker emoji
• sticker set name if relevant to product logic
• whether it is animated or video

For AI, pass sticker data only if it changes interpretation. In many bots, emoji alone is enough.

5. Attachment summary

Attachments are often important even when you do not pass the binary itself.

Useful fields by type:

• photo: has_photo, caption, optional image OCR or vision output if your stack supports it
• video: has_video, caption, optional transcription or vision summary
• voice/audio: has_voice or has_audio, optional transcription
• document: file name, mime type, caption, optional extracted text
• animation/video note: media type plus caption or summary

For AI, prefer a summary field over raw Telegram metadata:

Attachment: voice message
Transcription: I need help connecting the bot to my channel.

6. Forward origin

Forward origin can change how AI should interpret a message.

Useful fields:

• forward origin type
• forwarded user id or chat id when available
• forwarded display name when available

Use scripts/forward_origin_context.py as a minimal typed example.

7. Entities and links

Entities are often more useful than raw text parsing.

Useful signals:

• URLs
• text links
• mentions
• hashtags
• bot commands
• custom emoji ids when your bot resolves them to text-like meanings

For AI, it is usually enough to pass a short entity summary:

Entities: url=https://example.com, mention=@support, hashtag=#launch

8. Sender and chat context

Sometimes meaning depends on who sent the message.

Useful fields:

• whether sender is from_user or sender_chat
• sender display name
• whether the message is a channel comment
• chat type: private, group, supergroup, channel
• whether the chat is a forum topic

Do not pass internal ids to the model unless they affect reasoning.

What to pass for common AI tasks

Moderation

Useful context:

• main text or caption
• quote text
• link and entity summary
• forward origin
• attachment type and transcription if any

Reply intent classification

Useful context:

• main text
• quote text
• reply flag
• topic or thread flag

Assistant response generation

Useful context:

• main text
• quote text
• poll summary if the message is a poll
• attachment summary or transcription
• sender or chat role only if it changes tone or permissions

Support bots

Useful context:

• main text
• transcription for voice notes
• document extracted text when available
• quote text showing what the user is responding to

Recommended formatting for AI

Two good universal formats are:

Option A: compact labeled text

Message text: How do I connect this?
Quote text: Click Bot Settings and then Add to Channel.
Chat type: supergroup
Is reply: true
Attachment: none
Entities: none
Forward origin: none

This is simple and works well for prompts.

Option B: small structured object

{
  "message_text": "How do I connect this?",
  "quote_text": "Click Bot Settings and then Add to Channel.",
  "chat_type": "supergroup",
  "is_reply": true,
  "attachment": null,
  "entities": []
}

This is better when a downstream service already expects structured inputs.

What not to pass

Avoid sending large low-signal blobs such as:

• raw serialized Message objects
• every Telegram field by default
• binary metadata that does not affect interpretation
• internal ids that the model cannot use meaningfully

Minimal principle

Only pass the Telegram context that changes model interpretation:

• visible message text
• quote text
• poll semantics
• attachment summary or transcription
• forward origin metadata
• reply or thread relation when it changes meaning