If the tool (backend) itself was at fault (for example, a timeout when accessing the system) — did you pass traceback errors to the models?
Machine Learning / AI
Was there a check for record duplication in the system after a failure following a successful write (idempotency)?
What dataset was used to measure model selection metrics?
How was the entire multi-agent system traced — at which stages do errors, latency, etc. occur?
Have other embedding models been tested?
What metrics were used to measure retrievability (which dataset and tool)?
Why was Human-in-the-Loop not used in the multi-agent system, only in ReAct?
What was the chunk size and did you do overlap?
What is BERT and what tasks was it trained on?
Describe the ReAct agent cycle: what were the steps and how did the cycle end?
What types of annotations for NER tasks do you know? What have you worked with?
What is VLM (Vision Language Model)?
In terms of speed, which is faster — a regular agent with tools or a multi-agent system?
Which part (BM25 or dense) has unnormalized values?
Was the toolcalling native or through prompt (parsing)?
Used one model for all agents or different ones?
And if the model gave an invalid argument, what did you do?
How do models report that an argument is invalid?
What is the difference between fine-tuning and few-shot prompting? When and what to choose?
Was there parallelism in the execution of agents or was everything sequential?