Conversational Generative AI
Users can upload documents in various formats and interact with the generative AI with the Hallucination detection feature.
Feature
Chat UI
User can try in the same way as a general chat service.
Retrieval Augmented Generation(RAG)
AI models can provide answers by referring to external documents that cannot be answered by itself. Besides, user can confirm the original text refered in AI response.
Hallucination Detection
The hallucination detection feature can detect Hallucinations (plausible errors not based on data) in sentences answered by the Conversational Generative AI.
A metric called the hallucination score is calculated based on whether the same meaning answer can be obtained stably. The closer the hallucination score is to 100, the more likely it is that the answers contain errors. Please use this feature to determine whether or not to confirm facts using reliable information sources.
However, the Hallucination Detection is also a technology under development and cannot completely detect hallucinations. Please always verify the AI's response by yourself.
Phishing URL Detection
The phishing URL detection feature determines whether the URL in the interaction text with the interactive AI is a phishing URL and presents it. If a phishing URLs are detected, exclamation marks (!) are displayed and the URL cannot be clicked. You can use interactive AI with highly confidence.
Keyword Mask
The Keyword Mask function detects sentences of keyword stuffing in your uploaded documents and subsequently masks them. The keyword stuffing is a listed keyword that is deliberately designed to gain an unfair ranking advantage in document searches, making responses of LLMs less trustworthy. This function detects keyword stuffing injected by adversaries. Besides, it detects unintentional keyword lists, such as the pages of Table of Contents and Indexes. These result in making RAG-integrated LLM applications more secured and accurate.
Available AI Models
- GPT-3.5 (powered by Azure OpenAI Service)
- GPT-4o (powered by Azure OpenAI Service)
- Claude Instant (powered by Amazon Bedrock)
- Fugaku-LLM (CPU version only available in the free trial)
- Since it uses the CPU to process information, it takes about tens of seconds to a minute to return a response.
- If you wish to use the GPU version, which offers faster processing, please contact us through the Fujitsu Kozuchi website.
Notes
Please be aware that generative AI models may generate inaccurate or inappropriate information.
When using these models, please verify the output results at your own responsibility, and make necessary corrections or adjustments as needed.
OSS list
License information for OSS, etc. used by Conversational Generative AI.