Spaces:
Running
Running
| # Migration: v1.1 -> v1.2 | |
| Replace the v1.1 source/docs with the v1.2 patch and push a normal new commit. No database migration is required because workspaces are ephemeral. | |
| ## Important behavioral changes | |
| - `gr.State` session ID -> `gr.BrowserState` with compatibility fallback. | |
| - `Workspace.ingest()` accepts an optional `progress_callback`. Existing callers remain valid. | |
| - query traces now include `workspace` metadata. | |
| - retrieval trace separates unique `selected_sources` from `retrieved_chunks`. | |
| - the graph has an explicit `abstain` terminal node. | |
| - UI demo mode can auto-index the bundled corpus when the workspace is empty. | |
| - source snippets are stored longer for diagnostics but rendered as clean word-boundary previews. | |
| ## Deployment | |
| The v1.1 dependency pins and Hugging Face filesystem permission fix are retained. The NIST PDF remains tracked through Git Xet/LFS; do not remove `.gitattributes`. | |
| After deploy, test: | |
| 1. fresh page + demo enabled -> ask `What is the collection about?` without manually indexing; demo should initialize and answer from documents. | |
| 2. refresh page -> corpus should remain available while the same Space process is alive. | |
| 3. reset session + disable demo -> a document-only query should return the explicit no-corpus abstention without retrieval/verification loops. | |
| 4. manual indexing -> button should show `Building corpus...` and progress stages. | |
| 5. source panel -> long snippets should end with `...` rather than a cut-off word. | |