Overview
This article provides answers to frequently asked questions about Ask JoVE AI, including its underlying technology, data handling practices, conversation retention, and methods used to improve response accuracy.
1. What foundation model powers Ask JoVE AI, and where is it hosted?
Ask JoVE uses multiple open-source large language models (LLMs).
These models are further refined using reinforcement training to improve their performance for JoVE use cases.
Hosting region: United States.
2. Are customer queries or AI-generated responses used to train third-party AI models?
No.
Customer prompts and AI-generated responses are not used to train vendor or third-party AI models.
3. How long are conversations retained?
Currently, Ask JoVE retains the two most recent conversations.
Future enhancements may allow conversation history to be retained for up to three years so users can review previous interactions.
4. Does Ask JoVE AI answer questions using only JoVE content?
Ask JoVE AI uses a Retrieval Augmented Generation (RAG) architecture.
Responses are generated using multiple sources of information:
Foundation knowledge
The underlying language models possess general language understanding acquired during pretraining on publicly available and licensed datasets.
JoVE knowledge
When responding to users, Ask JoVE primarily retrieves information from JoVE resources, including:
- Video transcripts
- Scientific protocols
- Experiment methodologies
- Internal knowledge resources
- Documents and structured datasets
Relevant information is retrieved first and then supplied to the model to generate grounded responses.
5. How does Ask JoVE AI reduce hallucinations?
Although no AI system can eliminate hallucinations, Ask JoVE uses several methods to improve factual accuracy.
These include:
- Retrieval-Augmented Generation (RAG) using verified JoVE content
- Grounding responses in retrieved source material
- Prompt instructions that discourage speculation
- Output validation and confidence checking
- Reinforcement learning and domain-specific model refinement
- Internal verification and multi-step reasoning techniques
6. Are audit logs available?
Yes.
Ask JoVE maintains audit information internally.
7. Is customer data used to train Ask JoVE AI?
No.
Customer prompts and responses are not used to train Ask JoVE AI or third-party AI models.
Summary
Ask JoVE AI is designed to respond grounded primarily in JoVE's scientific content while protecting customer data.
Key points include:
- Hosted in the United States.
- Customer prompts are not used to train third-party AI models.
- Uses Retrieval Augmented Generation (RAG) to improve response accuracy.
- Retains the two most recent conversations.
- Employs multiple techniques to reduce hallucinations.