💬 TGMix

TGMix is a powerful tool that processes your Telegram chat export into an AI-friendly dataset. Perfect for feeding the full context of long and complex conversations to Large Language Models (LLMs) like Claude, Gemini, GPT-4o, and more. Inspired by Repomix.

🛠️ Beta Version Note

Please note that TGMix is currently in a beta phase. This means that while the core features are functional, you may encounter occasional bugs or unexpected behavior with certain edge cases.

Your feedback is invaluable during this stage. If you find any issues, please report them on GitHub Issues.

Core features

• Save Costs & Time: Processing a chat history that's up to 3x smaller in token count directly translates to lower API costs for paid models (like GPT-4o and Claude 3 Opus) and significantly faster response times.
• Fit More Data into Context: The token reduction is crucial for fitting large chat histories into a model's limited context window — something that is frequently impossible with raw Telegram exports.
• Higher Quality Analysis: By stitching fragmented messages, you provide the LLM with a more natural and complete context. This prevents misinterpretations and leads to more accurate and insightful summaries, analyses, or role-playing sessions.
• Efficient Output Format: TGMix uses the TOON format instead of JSON. This provides a 30-60% token reduction(as documentation says) over standard JSON, directly cutting API costs and speeding up LLM processing. Benchmarks also show that models retrieve data more accurately from TOON.
• Flexible Anonymization: Automatically mask sensitive data like phone numbers, emails, author names, and custom patterns to protect privacy before sharing chat data.
• Data for RAG & Fine-Tuning: The clean, structured output is a perfect dataset for advanced applications. Use it to build a knowledge base for Retrieval-Augmented Generation (RAG) or to fine-tune a custom model on a specific person's conversational style.

Roadmap

The development of TGMix is planned in stages. Here is what's available now and what to expect in the future.

Current Version

• Significant Token Reduction: By simplifying the structure and removing redundant metadata from the original Telegram export, TGMix reduces the final token count by up to 3 times.
• Message Stitching: Automatically combines messages sent by the same user in quick succession into a single, coherent entry.
• Media Marking: Uses MarkMyMedia-LLM to automatically add filenames to media like videos and voice messages, improving context for AI analysis.
• AI-Ready Toon Output: Produces a single, clean tgmix_output.toon.txt file in the TOON (Token-Oriented Object Notation) format, which is more token-efficient than JSON and specifically designed for LLMs.
• Advanced Anonymization: A flexible system for masking sensitive data.

Planned for Future Releases

• Advanced Media Processing: Optional conversion of voice/video messages into text via transcription.
• Official Package Manager Support: Easy installation via AUR.

Requirements

• Python 3.10+

Installation

Via PyPI

pip install -U tgmix

From GitHub (For development)

Install tgmix directly from this repository:

pip install -U git+https://github.com/damnkrat/tgmix.git

How to Use

Step 1: Export Your Telegram Chat

1. Open Telegram Desktop.
2. Go to the chat you want to export.
3. Click the three dots (⋮) in the top-right corner and select Export chat history.
4. Crucially, in the export settings:
   • Set the format to "Machine-readable JSON".
   • Choose a date range and media settings as desired.
5. Let the export process complete. You will get a folder containing a result.json file and media subfolders.

Step 2: Run TGMix

1. Navigate to your exported chat directory in your terminal.

   cd path/to/your/telegram_export
   

2. (Optional) Create a local configuration file.

   tgmix --init
   

   This will create a tgmix_config.json file. You can edit it if your export has non-standard file names.

3. Run the processor.

   tgmix
   

   To enable anonymization, add the -a flag:

   tgmix -a
   

Step 3: Use the Output

Once finished, you will find:

• tgmix_output.toon.txt: The final, processed file in Toon format, ready for your LLM.
• tgmix_media/: A new folder containing all processed and copied media files.

Configuration

You can control TGMix's behavior by editing the tgmix_config.json file. Create one in your project directory by running tgmix --init.

Anonymization Settings

The core of the new functionality lies in the anonymization settings. Here is an example of a fully configured setup:

{
  "export_json_file": "result.json",
  "media_output_dir": "tgmix_media",
  "final_output_file": "tgmix_output.toon.txt",
  "anonymize": true,
  "default_phone_region": "RU",
  "mask_presets": {
    "phone": "[PHONE]",
    "email": "[EMAIL]",
    "authors": "[AUTHOR]"
  },
  "mask_literals": {
    "Project Capture": "[SECRET_PROJECT]"
  },
  "mask_regex": {
    "\\b\\d{16}\\b": "[CARD_NUMBER]"
  }
}

• "anonymize": true or false. The master switch for the anonymization feature.
• "default_phone_region": A two-letter country code (e.g., "RU", "US", "GB") to help detect local phone numbers that don't have a + prefix. This is a tradeoff for accuracy.
• "mask_presets": A dictionary of built-in, ready-to-use masking rules. The key is the preset name, and the value is the placeholder text that will replace the found data. Available presets are:
  • "phone": Finds and replaces phone numbers. It intelligently detects both international formats (e.g., +1-541-754-3010) and local formats based on the "default_phone_region" (e.g., 8 (999) 123-45-67).
  • "email": Finds and replaces email addresses using a robust, tested regular expression.
  • "authors": This is a special preset that anonymizes the authors themselves. It modifies the author_map in the final output, replacing the author's name with the provided template, and a unique number (e.g., [AUTHOR_1], [AUTHOR_2]).
• "mask_literals": A dictionary for replacing exact, case-insensitive phrases. The key is the phrase to find, and the value is its replacement. Perfect for redacting names, project titles, or other specific keywords.
• "mask_regex": A dictionary for replacing content based on regular expressions. The key is the regex pattern, and the value is its replacement. This gives you the power to mask any custom data format, like ID numbers, bank accounts, or tracking codes.

  Important: In JSON format, the backslash \ is an escape character. Therefore, you must escape your backslashes in regex patterns. For example, to match a digit (\d), you must write it as \\d in the tgmix_config.json file.

Command-Line Overrides and Examples

While tgmix_config.json is great for setting up your default anonymization rules, you can easily override or supplement them for a specific run using command-line flags. CLI flags always take precedence over the configuration file.

Command-Line Usage

You can control TGMix directly from the command line. Flags and options provided here will always take precedence over the settings in tgmix_config.json.

Options

• path (Positional Argument) The path to the directory containing the Telegram export. If omitted, TGMix will process the current directory.
• --init Creates a tgmix_config.json configuration file in the current directory from a built-in template.
• --version Displays the installed version of TGMix and exits.
• -a, --anonymize Enables the anonymization feature for the current run.
• --no-stats Disables the computation and printing of processing statistics at the end of the run.
• --no-mark-media Disables media marking. Files will be copied to the output directory without changes.
• --mask-preset <preset1> <preset2> ... Overrides the list of active presets from the config file. Only the presets you list here will be used.
• --mask-literal "phrase:replacement" Overrides the mask_literals dictionary from your config file. You can provide multiple rules.
• --mask-regex "pattern:replacement" Overrides the mask_regex dictionary from your config file. You can provide multiple rules.

Example 1: Enabling Anonymization with Default Rules

If your "anonymize" key is set to false in the config, you can easily enable it for a single run:

tgmix --anonymize

• What it does: This command enables the anonymization feature and uses all the rules exactly as they are defined in your tgmix_config.json file.

Example 2: Using Only Specific Presets

Imagine you only want to anonymize authors for a particular analysis, ignoring other presets from your config.

tgmix --anonymize --mask-preset authors

• What it does: This command overrides the mask_presets from your config. It will only anonymize author names in the author_map, ignoring the phone and email presets even if they are present in the JSON file.
• Note: Authors are not anonymized in message text. The reason is author name can be, for example, just "a". It will just break most messages in the export. To anonymize it in message text, you need to add it to mask_literals or mask_regex.

Example 3: Adding a One-Time Redaction Rule

You need to redact a new project name just once, without permanently editing your config file.

tgmix --anonymize --mask-literal "Project Capture:[SECRET_PROJECT]"

• What it does: This command uses all the default rules from your config (presets, regex) but overrides the mask_literals dictionary, adding a new, temporary rule to mask "Project Hydra".

Example 4: The Power User Command

This command demonstrates full control from the CLI, overriding all rule types for a highly specific task.

tgmix --anonymize \
      --mask-preset email phone \
      --mask-literal "damnkrat:[USER_A]" "secret password:[REDACTED]" \
      --mask-regex "\b\d{4}-\d{4}\b:[ID_CODE]"

• What it does:
  • --anonymize: Enables the feature.
  • --mask-preset email phone: Uses only email and phone presets.
  • --mask-literal ...: Overrides all literals from the config and uses only the two provided here.
  • --mask-regex ...: Overrides all regex rules from the config and uses only the one for a custom ID code.

License

This project is licensed under the GNU General Public License v3.0. See the LICENSE file for details.