Оценка RAG-модели сессионных вложений

★ 7.2 · testing

session-rag-eval is a Claude Code skill that runs and debugs Chatbox session attachment RAG model evaluation using synthetic and real long-file fixtures. It verifies two sides of model behavior: that the model calls `query_session_attachment` when an answer depends on an uploaded large file, and that it skips retrieval for clearly unrelated requests. Fixtures include synthetic long documents with hidden needle facts, Wikipedia-derived real documents, direct and implicit positive cases, and unrelated negative cases. Full product validation runs through the Chatbox Electron harness, exercising the renderer, local API, file upload, indexing, tool registration, SQLite session RAG DB, and persisted messages; a `--dry-run` mode enables fast model-behavior iteration without Electron. The skill is aimed at QA engineers testing RAG pipeline correctness and regression in chat applications that handle large file attachments.

• #rag-evaluation
• #session-attachment
• #model-testing
• #quality-assurance