O
Homelab AI Runtime

Oppa AI Gateway

Local-first routing for embeddings, fast chat, long-form reasoning, and bridges into Codex and Claude Code. Live roles stay stable while training outputs land in candidate storage until you explicitly promote them.

API Docs Widget JS Health JSON
System Status
Gateway Healthy
Public URLhttps://ai.bourdeauhomelab.com
Widget/widget/oppa-ai-widget.js
Inbox bridgeAI Inbox on home.bourdeauhomelab.com
MCP Bridge Connected
StatusConnected to staged MCP server
EmbeddingsFallback JSON embeddings
Use caseMemory, tasks, snapshots, shared context
Model Roles
Embedding Agent Online
DescriptionSemantic search, retrieval, and memory lookup.
Port8781
URLhttp://127.0.0.1:8781
Fast Agent Online
DescriptionQuick answers, triage, and routing decisions.
Port8782
URLhttp://127.0.0.1:8782
Large Agent Online
DescriptionLonger reasoning, richer answers, and fallback synthesis.
Port8783
URLhttp://127.0.0.1:8783
Embed In Oppa
Widget Snippet Ready
<script src="https://ai.bourdeauhomelab.com/widget/oppa-ai-widget.js" data-gateway="https://ai.bourdeauhomelab.com"></script>
Obsidian Vault Index
Index Status Indexed
Vault path/Volumes/Storage Drive/Homelab_Apps_storage/obsidian-sync/vault
Files indexed621
Total chunks6901
Last indexed2026-05-03T01:19:51.117650+00:00
Scan runningNo
New and changed notes are indexed automatically via file watcher. Nightly full scan runs at 2 AM. Deleted notes are never auto-purged — their embeddings persist until you explicitly remove them.
Last Scan Results
Indexed (new)
Updated
Skipped (no change)622
Errors
Total scanned622
OCR Document Indexer
Upload Document docs
Supported: PDF (digital + scanned), PNG, JPG, JPEG, TIFF, BMP, WEBP. Max 50 MB. Each file is chunked and embedded into the vector store. Embeddings are never auto-deleted — use the Delete button to remove a document's chunks. Requires: pypdf, Pillow, pytesseract, and tesseract binary (brew install tesseract).
Index Summary
Documents uploaded
Total chunks
Uploaded Documents
Filename Document ID Chunks Uploaded
No documents uploaded yet.
Training Workspace
Manual DatasetDatasets
Build From Gateway SamplesMCP / Inbox
This pulls from `ai_training_samples` so you can stage supervised examples without touching your live model roles.
Datasets
No datasets created yet.
Training Jobs
Queue LoRA JobCandidates Only
Jobs train into candidate storage first. Nothing replaces the daily live model until you promote a candidate.
Queue Summary0 jobs
Datasets0
Candidates0
Live overrides0
On the base M4 Mac mini, keep this focused on adapters, prompt/routing improvements, and small eval loops instead of full model retraining.
No training jobs queued yet.
Candidate Models
No trained candidates yet.
Live Promotion & Rollback
Embedding Live Slot Base Only
No promoted override. Using the daily base model.
Fast Live Slot Base Only
No promoted override. Using the daily base model.
Large Live Slot Base Only
No promoted override. Using the daily base model.