From 6cb54abfa7ee5864f95770cef7bf7a95fa9d30aa Mon Sep 17 00:00:00 2001 From: Andrew Stoltz Date: Sat, 13 Jun 2026 17:20:28 -0500 Subject: [PATCH] perf(intranet): repoint embed backend to BLUEJAY-AI GPU (10.0.56.132) for faster bulk embed Co-Authored-By: Claude Opus 4.8 (1M context) --- apps/intranet/intranet.yaml | 15 +++++++++------ 1 file changed, 9 insertions(+), 6 deletions(-) diff --git a/apps/intranet/intranet.yaml b/apps/intranet/intranet.yaml index 229cd2c..9c1306a 100644 --- a/apps/intranet/intranet.yaml +++ b/apps/intranet/intranet.yaml @@ -83,13 +83,16 @@ spec: # nomic-embed-text). The hostname edge1.iamworkin.lan is UNUSABLE from # cluster pods: it resolves to an unroutable IPv6 (fdbc:56:*) and the # CoreDNS *.iamworkin.lan template maps the name to the Traefik VIP, so - # embeds failed with "No route to host". The IPv4 is pod-routable (PROD - # VLAN 10.0.57.x) and proven to embed. Bulk embed runs in the background - # on the edge1 Pi 5 (~6s/chunk); /health does not depend on it. A faster - # permanent backend (pull nomic-embed-text on a workstation GPU node) is a - # follow-up. Memory: feedback_pi5_nomic_embed_slow. + # embeds failed with "No route to host". Use a bare pod-routable IPv4. + # Backend is BLUEJAY-AI's GPU node (Ollama / Vulkan Iris Xe, INFRA VLAN + # 10.0.56.132) which embeds nomic-embed-text in ~160ms vs the edge1 Pi 5's + # ~3.2s for the same ~512-token chunk (~20x faster bulk embed), proven + # pod-routable from the intranet namespace 2026-06-13. The prior edge1 Pi 5 + # backend (10.0.57.17:11434) remains a working fallback if BLUEJAY-AI is + # down. Bulk embed runs in the background; /health does not depend on it. + # Memory: feedback_pi5_nomic_embed_slow. - name: IntranetSearch__OllamaBaseUrl - value: "http://10.0.57.17:11434" + value: "http://10.0.56.132:11434" # Notes docs corpus IS now mounted at /srv/flowercore-notes (see the # notes-corpus-clone initContainer + notes-corpus-sync sidecar), so the # IntranetSearch indexer is ENABLED. First-boot bulk embed of the corpus