Machine Learning / AI
What can be done to prevent the model from answering if there is no information in the context for answering the question?
How to technically pass only a part of the state in LangGraph, not the entire state?
How to prevent a situation where a mock system, based on a modified agent request, inserts irrelevant historical data or cannot find data for the request?
Что такое Wav2Vec 2.0 и self-supervised pretraining?
What embedding model to choose for RAG?
What is over-smoothing in deep GNNs and how to combat it?
Do all of this are done by different agents or different tools of the same agent?
Tell me how the RAG pipeline is generally arranged: what stages does it include?
What is model selection vs hyperparameter optimization?
How to measure that the embedding model and retrieval are working properly?
What is speculative decoding and how does it work?
What is a HOLD-OUT test of long-term effects?
What features are considered when evaluating the quality of browse agents (WebArena)?
Do you use state in LangGraph and pass the entire state in the context?
How to process a table in a document when preparing data for RAG?
What can be done when the agent's context gradually fills up during long work?
On which project did you work on deploying services and developing FastAPI services: did you implement it yourself or only support and observe?