RAG-вопрос
A video-analytics platform wants an assistant that answers support requests using documentation and prior tickets. How would you design the RAG system?
Ответить самому
Сначала сформулируйте ответ как на собеседовании, затем откройте разбор и оцените себя.
Короткий ответ
Ingest docs/tickets with permissions, chunk and index them, retrieve with hybrid search and reranking, generate grounded answers with citations, and evaluate on real support cases plus hallucination/coverage metrics.
Полный разбор
Start with sources and permissions: product docs, runbooks, ticket history, known incidents and customer-specific configuration. Clean and chunk documents so each chunk has metadata such as product area, version, customer, language and access level. Support data often contains sensitive customer information, so access control must be enforced before generation.
Retrieval should usually be hybrid: lexical search for exact error codes and product names, vector search for semantic paraphrases, and reranking to improve precision. The answer prompt should require grounded responses, cite retrieved evidence and admit when documentation is insufficient.
Evaluate with a held-out set of real tickets: answer correctness, retrieval recall, citation support, escalation rate, latency and hallucination rate. Production should keep human handoff, feedback capture and monitoring for new product versions or stale documentation.
Типичные ошибки
- Use vector search only and lose exact error-code matching.
- Ignore document permissions and customer-specific data.
- Evaluate only with generic LLM quality impressions.