CubelaPetarandClaude Opus 4.8
75296ce58c
- rename host/group/folder everywhere to match the server's actual hostname and physical label - share-analysis.ps1: sanitize the output filename prefix — '-Paths "D:"' produced 'D:-file-types.csv' and Export-Csv failed with 'path format not supported', so no CSVs were written - share-analysis.ps1: new -FolderDepth so folders can be aggregated at D:\Abteilungen\<Share> level, which matches the share layout - list-shares.ps1: -WithSize walks local paths instead of UNC when run on the server itself (UNC was orders of magnitude slower and looked stuck) and prints progress every 50k files Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2.3 KiB
2.3 KiB
TODO
phy-srv-gpu01
Per projektplan (§ references below).
Server is delivered; base setup + NVIDIA driver are done (see SETUP.md).
Blocking the knowledge base (customer conversation)
- Scope the corpus (§2.6) — share analysis found 1.26M extractable documents, ~12× the planning threshold. Decide with the customer which shares/folders actually form the knowledge base. Everything in Phase 5 depends on this.
- Evaluate the two CSVs still sitting on the file server:
D-toplevel-folders.csv(basis for the include/exclude list) andD-file-types.csv(resolve the 1.58M unidentifiedotherfiles) - Run
scripts/list-shares.ps1on Z-FILESERVER — map shares → local paths, so the analysis ofD:can be tied to actual shares - Data protection / works council: are chat logs stored? (§2.2) — before rollout
- AD details still needed: bind service account + base DN (group
llm_usersand the read-only SMB account are agreed) - German eval set (§2.5): 20–30 Q&A with the customer — acceptance criterion
Server / Ansible
- Role
cifs_mounts— read-only mounts, credentials fromgroup_vars/secrets.yml - Role
llm_stack— docker compose: vLLM (fixed--gpu-memory-utilization), embedding server, Open WebUI + pgvector, reverse proxy, oikb timer. Tag compose tasks withcomposesojust composeworks. - Smoke-test a vLLM container on the server (official images may not support SM120)
— pin the working image, then build
llm_stack. Start model: Qwen3-32B FP8 - OCR pipeline for ~166k scanned PDFs (§2.1) — own work block, competes with vLLM for the GPU
- TLS: currently plain HTTP on
chat.phytron.local; retrofit an internal CA certificate (AD passwords travel in clear text until then) - Hostname mismatch: server reports
phy-srv-gpu01, repo/inventory/label usephy-srv-gpu01— align - Finish base hardening in
SETUP.md(ssh, updates) viajust run phy_srv_gpu01
Done
- Share analysis script →
scripts/share-analysis.ps1, run on 2026-07-14 (§2.1) - Storage decision (§2.3) — 876 GB usable, no extra disks needed
- NVIDIA driver via role
nvidia_gpu(595 open kernel modules)