Spaces:
Sleeping
Sleeping
| # Operations notes | |
| These are project-authored demonstration notes for Document RAG Lab. | |
| The Streamlit demo creates a separate in-memory Chroma collection for each browser session. Documents stay in the server process until it stops. The separate FastAPI service uses a persistent local Chroma directory. | |
| Local retrieval requires no language-model API key. Generated answers require NVIDIA_API_KEY and NVIDIA_MODEL. NVIDIA_BASE_URL defaults to the NVIDIA integrate API. The demo asks for permission before sending the question and selected excerpts to that provider. | |
| Never add API keys or uploaded private documents to Git. The example environment file contains variable names and empty values only. Provider error messages shown to the user omit request details. | |
| For slow requests, measure embedding initialization separately from warm search. Compare request times on the same machine with the same corpus. A small local search time does not describe full generated-answer latency. | |
| Uploads are limited to 2 MB per file, 60 PDF pages, and 200,000 extracted characters per document. Scanned PDFs need OCR outside this app. | |