Skip to main content
One of the key strengths of AI Agents is their ability to improve over time. Cognigy.AI offers the Intent Trainer feature that allows you to refine AI Agents based on existing dialogs. This tool improves the AI Agent’s understanding by letting you review and add user inputs to Intents.

​
Key Features

  • Analysis of Collected User Inputs. Review and analyze Intent Trainer records to identify areas for improvement.
  • Adding Inputs to Intents. Choose which Intent Trainer records will enhance specific Intents and add them to the Intent Sentences list.
  • Instant Training. After refining Intent Trainer records, quickly update the NLU model on the Intent Trainer page. This action eliminates the need to rebuild the model on the Flow page.
  • Automatic Scoring. The Train capability automatically scores Intents, and you can visually track progress with color-coded icons and scoring data in the list.

​
Limitations

  • By default, the Intent Trainer records Time-to-Live (TTL) is set to 10 days (14400 minutes). If you have on-premises or dedicated SaaS installations, you can change this value using the following variable: TRAINERRECORD_TTL_IN_MINUTES. For example, to set the TTL to 30 days, configure the variable as follows: TRAINERRECORD_TTL_IN_MINUTES="43200".
  • The maximum file size for uploading Intent Trainer records is 150 MB.

​
Working with Intent Trainer

In Tweak > Intent Trainer, you can:
  1. Find Intent Trainer records by using filters
  2. Refine Intent Trainer records: add to Intents, skip, or ignore
  3. Train the NLU model
  4. (Optional) Import and Export Intent Trainer records

​
Filter Intent Trainer Records

Review the collected input records from users and search for them using filters.
The Filter Preset allows you to filter records based on predefined categories for more efficient navigation and analysis.
Filter Preset OptionDescription
Not UnderstoodView all records that the system didn’t understand.
Show AllView all available records without any filters applied.
Found IntentsView records with Intents identified by the system.
Found Lexicon SlotsView records where Lexicon Slots were detected.
Poor Intent ScoreView records with Intents that have a low confidence score.
Fair Intent ScoreView records with Intents that have a medium confidence score.
When you select Custom in the Filter Preset, you can apply any filter. If you modify a filter while a preset option is selected, the Filter Preset automatically changes to Custom.
FilterDescriptions
SnapshotSelect a specific Snapshot to view Intents and related records. By default, No Snapshot is selected.
LocaleFilter records by language or locale when multiple locales are used. By default, Any Locale is selected.
IntentFilter by specific Intents to refine the results. By default, Any Intent is selected.
Review StatusFilter records by their review status to track progress:
  • Not Reviewed — a record was not reviewed.
  • Reviewed — a record was added to the Intent and reviewed.
  • Ignored — a record, along with all future inputs containing the same text, is ignored.
  • Skipped — a record is skipped for now and can be revisited later, with the option to review it again.
Found IntentFilter records based on whether an Intent was identified or not. By default, No Intent Found is selected. You can select one of the following options:
  • Found Intent — shows records where an Intent was successfully matched.
  • No Intent Found — shows records where no Intent was identified.
Found a SlotView all Slots found based on the selected filter option. By default, Not Selected is chosen. You can select one of the following options:
  • Found Intents — shows records where a Slot was identified.
  • No Intent Found — shows records where no Slot was identified.
Intent ScoreFilter records by confidence score, shown as color-coded icons. By default, Not Selected is chosen. You can select one of the following options:
  • Poor (0–0.3) — low confidence score, meaning the Intent was not accurately identified.
  • Fair (0.3–0.7) — medium confidence score, showing a moderately accurate identification.
  • Good (0.7–1.0) — high confidence score, indicating a very accurate Intent identification.
Input TypesFilter records by user input type. You can select one of the following input types:
  • Positive Answer — user responded affirmatively. For example, yes, sure.
  • Negative Answer — user responded negatively. For example, no, never.
  • Greeting — user initiated a greeting. For example, hi, hello.
  • Goodbye Message — user expressed farewell. For example, goodbye, see you later.
  • Statement — user made a declarative statement. For example, I like this product.
  • Command — user issued a command. For example, turn on the light.
  • Why Question — user asked a why question. For example, Why is it included in the package?.
  • How Question — user asked a how question. For example, How does this work?.
  • Yes or No Question — user asked a yes/no question. For example, Do you have this item in stock?.
SlotFilter records by selected Slot type. You can select one of the following options:
  • None — no Slots identified.
  • Lexicon Slots — identified lexicon-based Slots, which are user-defined categories.
  • System Slots — Slots that are predefined in the system, such as dates or numbers.

​
Manage Intent Trainer Records

Based on the analysis, decide which user inputs will improve a particular Intent, add those inputs to the corresponding Intent.
  • GUI
  • API
  1. In the Intent Trainer, select a record from the list and click Add to Intent.
  2. In the Edit Record window, select an Intent to which you want to add a record.
    Change the text of the Intent in the Text field.
    Additionally, you can create a new Keyphrase or Synonym.
    Save changes and apply them. The Intent record will be added as a sentence to the selected Intent.
  • GUI
  • API
  1. In the Intent Trainer, select a record from the list and click Skip.
  2. Apply changes. This action moves the input to Skipped records, but it will reappear in Not reviewed if the same input is entered again.
  • GUI
  • API
  1. In the Intent Trainer, select an Intent record from the list and click Ignore.
  2. Apply changes. This action moves the user input to the Ignored records, and if a user enters the same input, it will also be ignored.

​
Train the NLU Model

After adding a record to the Intent, click Train in the top-right corner of the Intent Records page.
You don’t need to run Build Model in the Flow — the Intent Trainer has already scored the Intent, shown in the scoring data and color-coded icons.

​
Import and Export Intent Trainer Records

You can import and export records between production and development environments.
The Intent Trainer temporarily stores records. To ensure data safety, export the records for local storage.
  • GUI
  • API
  1. Go to Tweak > Intent Trainer.
  2. In the upper-right corner, click vertical-ellipsis > Import Trainer Records.
  3. Select a file in the CTRAIN format from your computer and click Open.
Once the file is uploaded, you will receive a system success message.
  • GUI
  • API
  1. Go to Tweak > Intent Trainer.
  2. In the upper-right corner, click vertical-ellipsis > Export Trainer Records.
  3. Select a date range by clicking the date and selecting the desired date in the calendar.
  4. To include reviewed records in the file, activate Include reviewed.
  5. Click Confirm, then Download Trainer Records.
The file will be downloaded in the CTRAIN format.

​
More Information

⌘I