Skip to main content

User Guide

How do I manually trigger a training pipeline?​

You can manually trigger a training pipeline in the internal Admin portal under the Cronjobs section. To launch a text training pipeline for a specific profile:

  1. Navigate to the Admin portal and open the Cronjobs page.
  2. Locate the row labeled text-training-<account>-<profile> corresponding to the account and profile you wish to train.
  3. Click the Run now button in that row to immediately start the training pipeline for that profile.

This action will initiate the model training and deployment process for the selected profile, updating response suggestions based on the latest available data.

How do I configure training parameters?​

Yes, you can configure training parameters for each profile in Deepdesk. Each profile has an active profile config, which controls how training is performed and what data is used:

  • Exclude as suggestion: JsonLogic rules for messages that should not be recommended by the model.
  • Ignore messages: JsonLogic rules for messages that should be ignored altogether (these are excluded during ingestion and never stored).

The profile config also references a text model config and a url model config, which provide more specific training parameters for the pipeline:

Text Model Config Parameters:

  • period: The window of historic data used for training. This is a string in the format <amount>-<unit>, where unit can be days, weeks, or months (e.g., 12-months, 1000-days).
  • include_full_messages: When one of the splitting flags is enabled, this flag determines whether the full message (not just the split sections) is also allowed to be recommended. Of the found frequently occurring responses, this controls if the entire message can be suggested in addition to the smaller split suggestions.
  • split_paragraphs: When enabled, response messages are split into smaller sections by paragraph, resulting in more fine-grained suggestions.
  • split_sentences: When enabled, response messages are split into smaller sections by sentence, resulting in even more granular suggestions.
  • agent_character_limit: The maximum number of characters from historic data to use (overrides the period if set).
  • apply_email_denoising (matching method only): Attempts to normalize frequently occurring response emails by identifying templates and removing noise from custom openings/closings.

These parameters allow you to tailor the training process to your specific needs, improving the relevance and quality of response and URL suggestions for your agents.

How do I disable training pipelines for a profile?​

Yes, you can disable training pipelines for a profile using two different switches:

  • The main "Is pipeline enabled" flag in the active Profile Config enables or disables scheduled URL and GPT2 pipeline jobs (covering model promotion and bi-weekly training).
  • There is a separate "pipeline enabled" switch in the Profile Content Types config. This controls the Text training pipelines specifically.

Disabling either switch will prevent the corresponding scheduled and manual training runs for that profile. No new models will be trained or deployed for the disabled pipeline type until the flag is re-enabled.

How can I filter out bot messages from the recommendations?​

You can filter out bot messages from response recommendations by configuring the exclude as suggestion or ignore messages rules in the profile config.

For example, you can use a JsonLogic rule like the following to exclude messages from bots or specific authors:

{
  "or": [
    { "in": [ "bot", { "var": "source" } ] },
    { "in": [ "Chat Bot", { "var": "author_name" } ] },
    { "!=": [ { "var": "rouLOB" }, "SSC" ] }
  ]
}

This rule will exclude any message where the source is bot, or the author name contains Chat Bot, or the rouLOB field is not SSC. Adjust these values as needed to match your bot identification logic.

JsonLogic

For details on how to use JsonLogic to define assignment and filtering rules, see JsonLogic Assignment & Filtering.