A guide to using Eureka AI
- Carlos Carasco
- Federico Delucchi
What is Eureka AI?
Eureka AI transforms data management by combining real-time feedback and business records from all sources while providing excellent security with user access control, name anonymization, and ISO 27001 certification. It uses Feedier's query language with 20+ filters for fast and accurate data searches. Built on cutting-edge LLM technology, Eureka AI delivers precise text insights and information retrieval.
How it works?
Eureka AI serves as your personal AI analyst within the platform. In this initial iteration, it will be integrated into two key modules:
Assignment Module: Previously known as the Copilot module, the Assignment module now leverages Eureka AI to generate insightful results for your assignments.
Feedier Report: Eureka AI also powers the annotations with AI-generated text in Feedier Reports, providing in-depth analysis to enhance your understanding and efficiency in interpreting your data.
Text Analysis : In the Feedier platform, some of the data processed is unstructured data (textual feedback). Today, Feedier uses artificial intelligence (AI) to analyse 4 types of tasks:
Sentiment analysis. In other words, the ability to determine a sentiment in a given verbatim.
Entity detection. The ability to identify "entities" such as personal data, product names, brand names, etc.
Theme detection. The willingness to mark the verbatim with themes related to the context of the project (bug, technical problem, idea for improvement, etc.).
Content generation. In practical terms, this means being able to automate tasks such as summarising verbatim, analysing feedback and generating reports.
Feedier uses LLM (Large Language Model) technology for all AI-related tasks. We work mainly with Mistral (based in France) and the models are hosted on Microsoft Azure in the European Union (Sweden).
There are 3 types of text analysis used on the platform:
1: Sentiment analysis
2: Entity detection
3: Theme (topic) detection
General configuration
Feedier’s Eureka AI uses the context of the business to enable our models to deliver results that make sense in the operational context.
Go to your organisation page(Admin role)
Add an organisation name
Add a description of the organisation
Type 1 : Sentiment analysis
Each verbatim analysed is assigned to a sentiment: Positive, Neutral ou Negative.
Sentiment | Explanation | Sentiment rate average |
|---|---|---|
| The verbatim is linked to a positive emotion on the part of the customer: Joy, Confidence, Serenity, Admiration. | 97.5% |
| The verbatim is not directly linked to an emotion or the emotion is not expressed strongly enough to be categorised. | 50% |
| The verbatim is linked to a negative emotion on the part of 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.
So let's dig into an example, What is the issue has 20 answers with the following breakdown:
As explained earlier, each answer has a sentiment rate so to calculate the average we do:
[20% * 97.5%] + [0% * 50%] + [ 80% * 2.5%] = 21.8% which leads to an overall negative sentiment score
Lets dive into further examples.
As a user you have 3 different topics which have the following breakdown:
| Positive Answers | Neutral Answers | Negative Answers | Sentiment Score |
|---|---|---|---|---|
Topic 1: Improvements | 30% | 40% | 30% | 49% → Neutral |
Topic 2: Facilities | 71% | 18% | 11% | 78% → Positive |
Topic 3: Management | 20% | 0% | 80% | 21.8% → Negative |
Type 2 : Entity detection
With Feedier, the "entities" extracted from the verbatims are added as attributes.
Attribute | Explanation |
|---|---|
| The value corresponds to a personal name. The value can contain the surname, first name or both. |
| The value corresponds to the name of a company in the public domain. |
| The value corresponds to company products that are not necessarily in the public domain, but are detected as being products mentioned by the customer. |
| The value corresponds to a physical location. |
Type 3 : Theme (topic) detection
Themes allow verbatims to be grouped according to common subjects. Once you have created a topic, you can:
Track the development of topics
Organise topics by sub-topics
Filter topics using the different Feedier filters
| Best practices if you are manually adding themes to ensure effective results: |
|---|---|
We strongly recommend that you do not use keywords in addition to the AI to recognise themes. | |
We strongly recommend using short themes that encapsulate one idea at a time. |
Choose topics that you’d like to track, if available, Feedier will group text responses that would fit under this topic type.
Our commitments
We do not use our customers' data to train the AI used (except in the case of a specific request in the "Feedier NLP" contract). The models are pre-trained.
No customer data is saved outside the Feedier infrastructure when using AI. We use pre-trained models in Inference, and anonymisation tasks are performed upstream.
Whenever content is generated by AI in the Feedier Platform, a notice is available on the Platform to clearly explain that the content is generated by AI.
We undertake to comply with the laws and recommendations of the French State and the European Union concerning the use of AI, in particular the AI Act.