Guide of : Eurêka AI
- Faye Muchery
Analytics
What is Feedier’s Eureka AI ?
The vast majority of Feedier's platform value relies on Artificial Intelligence. To ensure an extremely secure environment (ISO 27001 certified), compliant with GDPR, the AI Act, and leveraging all the data within Feedier, our teams have created an application layer called Eureka AI.
Eureka AI is composed of 4 components:
User access control (based on role, team, and view)
A vector database fed by feedback
A query module (with over 20 filters for fast and precise searches)
A LLM already trained and requiring no "fine-tuning" (Mistral)
Today, Feedier uses Artificial Intelligence (AI) for analysis on 4 types of tasks :
Assistance in analysis: This involves automating tasks related to the synthesis of verbatim responses, feedback analysis, report generation, naming reports, annotations, questions, etc.
Sentiment analysis: This refers to the ability to determine sentiment in a given verbatim response.
Entity detection: This means identifying “entities” such as personal data, product names, brand names, etc.
Topic detection: This refers to the ability to tag verbatim responses with topics related to the project context (bugs, technical issues, improvement ideas, etc.).
Feedier uses LLM (Large Language Model) technology for all AI-related tasks. We primarily work with Mistral (based in France), and the models are hosted on Microsoft Azure within the European Union (Sweden).
No customer data is used by Feedier (or Mistral) for training the models.
Our commitments
We do not use our clients' data for training the AI models used. The models are pre-trained models from Mistral.
No client data is stored outside of the Feedier infrastructure when using AI. We use pre-trained models for inference, and anonymization tasks are performed in advance.
Once content is generated by AI on the Feedier Platform, a notice is available on the Platform to clearly indicate that the content is AI-generated.
We are committed to respecting the laws and recommendations of the French government and the European Union regarding the use of AI, including the IA Act.
For any questions, feel free to contact our compliance team at: support@feedier.com.
General Configuration
Eurêka AI on Feedier uses the company’s context to ensure that our models deliver results aligned with the operational environment.
Access your organization’s page (Admin role required)
Add a complete description in the “Description” field
It is possible to disable AI modules based on roles from the “Roles & Access” page
[Content Generation] Task 1: Analysis Assistance
The Feedier Platform offers the automation of multiple time-consuming tasks using “generative” AI (LLM), leveraging the company’s context (provided during the configuration step) and the user’s intent.
The available tasks are :
Generating action plans
Generating annotations on reports
Generating questions
Generating translation
Auto matching imported data
[Semantic Analysis] Task 2: Sentiment Analysis
Each analyzed verbatim is assigned a sentiment : Positive, Neutral ou Negative.
Sentiment | Explication | Average sentiment score |
|---|---|---|
| The verbatim is linked to a positive emotion from the customer: Joy, Trust, Serenity, Admiration. | 97.5% |
| The verbatim is not directly linked to an emotion, or the emotion is not strongly enough expressed to be categorized. | 50% |
| Le verbatim is linked to a negative emotion from the customer : Anger, Contempt, Sadness, Disgust. | 2.5% |
In the text analysis module, you have an individual sentiment score for each topic and can view a breakdown of the sentiment score.
Let’s take an example: the topic "service" received 2000 responses, which are distributed as follows:
As explained earlier, each response has a sentiment score, so to calculate the average, we proceed as follows:
[9% * 97.5%] + [21% * 50%] + [70% * 2.5%] = 21%, which gives an overall negative sentiment score.
Now, let's look at other examples.
As a user, you have three different topics that are distributed as follows:
| Positives reponses | Neutral reponses | Negative reponses | Sentiment score |
|---|---|---|---|---|
Theme 1: Improvements | 30% | 40% | 30% | 49% → Neutral |
Theme 2: Facilities | 71% | 18% | 11% | 78% → Positive |
Theme 3: Management | 20% | 0% | 80% | 21.8% → Negative |
[Semantic Analysis] Task 3: Entity Detection
With Feedier, the "entities" identified in the verbatims are added as attributes.
Attribute label | Explication |
|---|---|
| The value corresponds to a personal name. It may include the first name, last name, or both. |
| The value corresponds to the name of a publicly available company.. |
| The value corresponds to the company's products, which may not necessarily be public but are detected as products mentioned by the customer. |
| The value corresponds to a physical location. |
[Semantic Analysis] Task 4: Topic Detection
Topics allow you to group verbatims based on common subjects. Once topics are created, you can:
Track topic trends
Organize topics into subtopics
Filter topics using Feedier's various filters
| Best practices for creating topics and ensuring relevance of results : |
|---|---|
It is strongly recommended not to use keywords in addition to AI for topic recognition. | |
It is strongly recommended to use topics that encapsulate one idea at a time. | |
It is strongly recommended to add specific instructions for topics related to your industry/operational context. |