📖User Guide

This page gives TrusstGPT users an overview of how to use TrusstGPT

Once TrusstGPT has been deployed into your account, the first step to take is to add users to enable them to log into TrusstGPT.

Setting up Single-Sign-On (optional)

If single sign on is required/preferred, please create a support ticket via the Trusst Customer Support Portal and ensure to provide details regarding your Identity Provider (e.g. Okta/Auth0/Frontegg/Active Directory etc). One of our team will be in touch to provide support with integration of your IDP and TrusstGPT.

User Creation (excl. SSO)

Note: TrusstGPT User management and role based access control, coming soon (ETA: 2024 Q2) To enable a user in TrusstGPT, you will need to configure a user directly within an Amazon Cognito User Pool. To do this, you will need to have the appropriate IAM permissions, e.g. AWS technical team.

1. From AWS Management Console, search and click "Cognito":

2. Select the user pool:
3. Select create user and enter details as below and click "Create User". Make sure to select "Send an email invitation" and "Mark email address as verified".
4. The new user will receive an email from "[e-mail]" with their login details

User Sign-in

1. Once created, new users will receive an email from "[e-mail]" with their login details. This includes "Username" and a "Temporary Password", along with the link to access your organisations instance of TrusstGPT. To sign in, click the link in the email.

2. Sign in with the login details from the email in step 1:
3. Enter a new password, family and given name(s) and click "Confirm":
4. Install an authenticator app on your mobile device, e.g. Google Authenticator.

5. From your authentication app on your mobile device, click add (+) and select "Scan a QR code" option or equivalent.

6. Scan the QR code presented by TrusstGPT which will add a new "code" to your authenticator app:
7. Enter the generated code from your Authenticator app into the "Code" section of TrusstGPT above and click continue.

8. Congratulations, you're officially logged into TrusstGPT!

Before you dive in, we suggest having a quick read of our Suggested Iteration Cycle to familiarise yourself with some of the lessons we've learnt along the way whilst helping other customers maximise the value they achieve with TrusstGPT.

Importing interactions

If you're the first user to sign into TrusstGPT, there will be no contacts in the platform.

1. Before you create the import job, make sure you have the "S3 URI" to the location where the interaction data that you want to import has been stored. e.g. "s3://your-bucket-name/path/to/folder/"

2. Once you have this, click "Contact Import Jobs" in the lefthand side navigation bar, and then "Create Import Job":
3. Paste the "S3 URI" from step 1 into the first field. If you have already made updates to the prompts, select the "Prompt Revision" you created, otherwise leave it blank and it will use the Default Prompts. In the "Job Name" field, you can give the Import Job a relevant name, otherwise you can use the name generated using the name of the S3 bucket. Then click "Create":
4. TrusstGPT will begin processing the interactions from S3. You can monitor the progress of the Import Job in the status bar: Note: The time it takes to complete will depend on the format of the files imported, audio files take more time than transcriptions, as TrusstGPT needs to transcribe the audio before analysis

1. You can filter the list of interactions by it's category or sub-category by clicking the + next to the category you want to filter by: This will then filter the list of interactions to ones that match that category, you can remove the category by clicking the on the category in the "Active Filters":
2. You can view the results of a contact as they complete by clicking on the contact in the list lower down on this page:

Prompt Playground

From within any contact, you can use the Prompt Playground which is available at the bottom of the contact page. Use the Prompt Playground to test prompt adjustments or add questions to a prompt. This enables you to extract answers relevant to the questions you want answered (click the "^" to pull up the Prompt Playground tray):

Within the prompt playground, questions should be phrased as if you were asking a question to someone who has read the transcript of the conversation, but has perfect memory recall.

Depending on the type of question and the type of response you're looking for, you can also adjust the "Prompt Type" by selecting the appropriate type from the drop down list:

Understanding the Prompt Types:

Note: If TrusstGPT is unable to provide an answer to the question relative to a contact, the question will be excluded from the results. Chat: This provides generally a more verbose response to your question, and is akin to what a human would respond if you were "chatting" to them about the interaction. Binary: This will respond with a True/False response. e.g. "Did the agent greet the customer by their first and last name?"

Rating: This should be used when you are requesting TrusstGPT to provide a rating between x and y. e.g. "On a scale of 1 to 10, rate the level of customer effort experienced during the conversation".

Single Label: This should be used when you are looking to label the conversation, e.g. "What was the root cause for the customer calling?"

Synthesis: This should be used if you are looking to provide a synthesis e.g. "Summarise what the agent could have done to improve the customer's experience."

Viewing categorisation of interactions (Analytics)

1. Browse to the "Analytics" page from the navigation bar on the left of screen, and select the Import Job you created:
2. The vertical bar graph provides a heat map of categories about what your customers interact with you about: Note: To maintain clarity and focus, only the most prevalent categories are shown.

   The vertical stacked bar graph displays the top 20 categories as a result of TrusstGPT's topic clustering process, which organises conversation topics by relevance and frequency.

3. See tooltips [?] for descriptions of each graph for a description of their contents.

4. We welcome feedback regarding additional or alternative Key Performance Indicators you would like to visualise on this page. Alternatively, you can also use your existing Business Intelligence tools to visualise any/all of the data outputs from TrusstGPT by connecting directly to the Amazon Redshift database containing the TrusstGPT outputs in your AWS account.

Adapting Insights in Prompt Revisions

Jump on over to "Prompt Revisions" from the navigation bar. This is where you will be able to create, read, update, and delete the sets of questions used (prompt revisions) to ask TrusstGPT to provide the insights you're looking to get.

To get started with adapting TrusstGPT's responses to answer your specific questions, you can either:

1. Click "Create Prompt", this allows you to modify and expand on the default list of questions. Note: There is a soft limit of 40 questions per Prompt Revision. This is to ensure timely responses (the more questions, the longer responses will take to compute).
2. If you already have Prompt Revision created, you can click on one of the prompt revisions from the list. As long as it is not the "Live" (Published) Prompt Revisions, you will be able to then click "Edit", where you can modify and expand on the list of questions, same as above

   By clicking "Publish" you will make this Prompt Revision "Live" which means that it will be used for all future interactions processed by TrusstGPT.

3. When creating or editing a Prompt Revision, you will note that the prompt is broken out into sections, e.g. "Synopsis", "Soft Skills", "Compliance". Synopsis: contains questions that will adapt the outputs of the Summary details produced by TrusstGPT. Here you can adapt the outputs of TrusstGPT by adding or modifying questions: Soft Skills/Compliance: Treat these questions as if you're asking TrusstGPT a question about what happened during an interaction. TrusstGPT currently responds with either "True", or "False". If the "Positive" toggle is turned on, a "True" response to the question will result in the agent achieving the weight set for that question. Conversely, if the "Positive" toggle is turned off, a "False" response to the question will result in the agent achieving the weight set for that question, otherwise a score of 0 is achieved. Conditional Quality Assurance: If you want to ask conditional questions, you can toggle "CQA" on. This will create a follow up question, e.g.: Note: The second question will only be evaluated if the result of the condition is "True", otherwise it will be marked as "N/A" and will not impact the score of the contact. When the condition is "True", only the score/weight of the second question is calculated for the interactions score. "Weight" is flexible, such that you can use whichever scale of weighting you prefer. Regardless of what scale you chose (1-10, or 1-50, or 1-100), the score is calculated based on Achieved Score/Total Weight possible, expressed as percentage (%). For critical criteria or questions, you can toggle "Auto fail" on. This will result in the interaction receiving a score of 0, whilst flagging the particular question which resulted in the failure. You can re-arrange questions by dragging and dropping the question (this is activated by hovering over the question beneath the "Title" of the question. By clicking "Publish" you will make this Prompt Revision "Live" which means that it will be used for all future interactions processed by TrusstGPT.

   If you want to regenerate insights of interactions already processed by TrusstGPT, you can do this by clicking "Regenerate Insights" from the Contact Import page. Note: Automated Quality Evaluations will always use the Published Prompt Revision. TrusstGPT will then reprocess the interactions using the revised prompt you created. You can monitor the process in the same way you would a standard Import Job.

Retry/Regenerate/Re-Import

To accelerate the iteration cycle, we have added the ability to re-perform actions on interactions at an Import Job or Contact level.

To re-perform an action at an Import Job level, you can select the drop down in the top right corner of an Import Job:

• "Retry Import" action will re-import all contacts within the contact import Job

• "ReImport Failed" action will only re-import contacts with a "Failed" status within the contact import Job

• "Regenerate Insight" action will request for regenerating the inference/insight on all calls for a contact import Job

• "Retry Failed Insights" action allows you to request for inference/insight on calls with failed inference

• "Regenerate Transcription" action will request for regenerating the transcription on all audio calls for a contact import Job

• "ReTranscribe Failed" action allows you request for transcription only for audio calls with failed transcription

To re-perform an action at an individual contact level, from within a contact, select the action you want to perform from the drop down:

• "ReImport Contact" action will re-import the selected contacts.

• "Regenerate Insight" will regenerate insights for selected contacts.

• "Regenerate Transcription" will re-generate the transcription for selected contacts. Enabled only on Audio calls (.MP3/.WAV/.OGG)

Handling Ambiguous or Multi-part Prompts in TrusstGPT

When using TrusstGPT, it's important to understand how the system handles questions, especially when they are ambiguous or contain multiple parts. Here's a straightforward guide to help you navigate these situations.

Single Question Response: When you ask a clear, single question, TrusstGPT will provide a straightforward response of either True or False.

Example: Input: "Did the agent introduce themselves articulately?" Output: True

Ambiguous or Multi-part Question Response: If your question is ambiguous or contains multiple parts, TrusstGPT may consider the first part of the question and respond to that only. Any additional parts will be ignored.

Example: Input: "Did the agent introduce themselves articulately and maintain a positive tone while offering assistance?" Output: True (response to the first part of the question only)

Best Practice

To ensure you get clear and accurate responses, break down your questions into separate, distinct prompts. This helps TrusstGPT understand and respond appropriately.

Example: Instead of asking: "Did the agent introduce themselves articulately and maintain a positive tone while offering assistance?" Ask:

1. "Did the agent introduce themselves articulately?"

2. "Did the agent maintain a positive tone while offering assistance?"

Summary

By breaking down your questions and following these guidelines, you can ensure TrusstGPT provides accurate and useful responses. This approach helps avoid ambiguity and ensures that each part of your query is addressed properly.