Machine Learning / AI

How to determine if an embedding model performs well? What metrics are used?

Tell me about cross-encoder and bi-encoder, what is the difference?

Is there a way to verify that generated text does not contain factual errors relative to the source?

How were data extracted from PDF documents for the RAG system?

Tell about quantization of models: what it is, why it is needed, and what types exist.

How was the chunking strategy chosen for the RAG system?

How to test a service implemented with FastAPI?

Tell about the performance metrics of an ML service: latency, RPS, how was it organized in your project?

Tell about vector databases, which ones exist, how do they differ, and why did you choose pgvector?

Tell me about LLM generation parameters: temperature, top-k, top-p, and others.

Tell me about the most complex task or difficulty in one of your projects.

How are Recall and Precision related, and is it possible to improve one without decreasing the other? Provide examples of tasks.

How to evaluate the quality of summarization?

How were the data organized for dialogue summarization and how was the training conducted?

Which quantization method (int8 or float) will degrade the model quality less?

How would you approach the task of summarizing a multi-person conference dialogue (for example, 10 developers in Slack)?

What model and approach would you choose for processing conference audio (speech-to-text, diarization)?

Tell me about LoRA: what is it, and why is it needed?

How to work with private data during local system deployment?

Why did you choose hybrid search BM25 + embeddings instead of only cosine similarity?