name: PMOVES Model Registry
description: |
Query, discover, and enrich the PMOVES.AI model catalog. Manages all AI model
metadata (LLM, embedding, TTS, vision), HuggingFace enrichment, TensorZero
TOML config export, and GPU deployment tracking across the fleet.
keywords: models, registry, catalog, embedding, llm, tensorzero, huggingface, gpu, deployment, discovery
version: 1.0.0
category: Infrastructure/AI
tier: 1
agent_class: Standard
agent_id: pmoves_model_registry
PMOVES Model Registry
Agent Class: Standard (Pmoves-)
Category: Infrastructure/AI
Version: 1.0.0
Tier: 1 (Core Infrastructure)
Status: Active — Supabase-backed model catalog + HuggingFace enrichment
Port: 8110
Capabilities
| Command |
What It Does |
list-models |
List all active models with optional type/provider filter |
get-model |
Get detailed metadata for a single model by ID |
enrich-hf |
Fetch dimensions, tags, CUDA support from HuggingFace API |
enrich-hf-bulk |
Batch-enrich all models that have hf_id in metadata |
export-tensorzero |
Generate TensorZero TOML config from catalog |
list-deployments |
Show active GPU model deployments across fleet |
register-deployment |
Register/update a model deployment (GPU Orchestrator) |
service-models |
List models mapped to a specific service |
Trigger Phrases (Pinokio 7 Interpreter)
| Phrase |
Action |
Endpoint |
"list available models" |
Show full model catalog |
GET /api/models |
"show embedding models" |
Filter catalog by type |
GET /api/models?model_type=embedding |
"show LLM models" |
Filter catalog by type |
GET /api/models?model_type=llm |
"get model details for [id]" |
Single model lookup |
GET /api/models/{id} |
"enrich model from huggingface" |
Fetch HF metadata + dimensions |
POST /api/models/{id}/enrich-hf |
"enrich all embedding models" |
Batch HF enrichment |
POST /api/models/enrich-hf-bulk |
"export tensorzero config" |
Download TensorZero TOML |
GET /api/tensorzero/config |
"show GPU deployments" |
List active model deployments |
GET /api/deployments |
"which models are on 5090" |
Filter deployments by node |
GET /api/deployments?node_id=5090 |
"what models does hi-rag use" |
Service-specific model lookup |
GET /api/services/hi-rag/models |
API Reference
Model Catalog
# List all models
curl http://localhost:8110/api/models
# Filter by type (embedding, llm, tts, vision, audio)
curl http://localhost:8110/api/models?model_type=embedding
# Filter by provider (ollama, openai, anthropic, venice)
curl http://localhost:8110/api/models?provider=ollama
# Get single model
curl http://localhost:8110/api/models/{model_id}
# Models for a service
curl http://localhost:8110/api/services/hi-rag/models
HuggingFace Enrichment
# Enrich a single model (requires metadata.hf_id set)
curl -X POST http://localhost:8110/api/models/{model_id}/enrich-hf
# Batch-enrich all embedding models
curl -X POST http://localhost:8110/api/models/enrich-hf-bulk?model_type=embedding
TensorZero Config Export
# Generate TOML config from catalog
curl http://localhost:8110/api/tensorzero/config -o tensorzero.toml
GPU Deployments
# List active deployments
curl http://localhost:8110/api/deployments
# Filter by node
curl http://localhost:8110/api/deployments?node_id=5090
# Filter by status
curl http://localhost:8110/api/deployments?status=loaded
Health Check
curl http://localhost:8110/healthz
# → {"status": "healthy", "timestamp": "...", "services": {"supabase": "...", "nats": "..."}}
Integration Points
- Supabase —
pmoves_core.models, pmoves_core.model_service_mapping, pmoves_core.v_active_deployments
- NATS — Publishes
model.registry.updated.v1 on catalog mutations (model enriched, deployment registered)
- GPU Orchestrator — Calls
POST /api/deployments when models are loaded/unloaded on GPU nodes
- TensorZero Gateway — Consumes exported TOML config for model provider routing
- HuggingFace API — Fetches model cards, config.json for embedding dimensions, tags, CUDA support
Prerequisites
- Supabase running with
pmoves_core schema seeded
- NATS message bus at port 4222 (optional — catalog changes still work without NATS)
- HuggingFace API accessible (no auth required for public models)
By