Provider Nodes

Provider nodes route AI operations through fabric-providers, enabling text generation, image creation, audio transcription, and embedding computation.

How It Works

1. Node specifies an operation (e.g., audio.transcribe) and ProviderRuntimeConfig (modality, model, tier)
2. ProviderRuntimeExecutor builds a ProviderRequest from the config and shared context
3. Request is routed through ProviderRouter (tier-aware, cost-based, health-checked)
4. Usage metadata is captured in the NodeExecutionResult

Routing

The ProviderRouter selects a provider based on:

1. Tier preference — basic (local/free) or premium (remote/paid)
2. Cost — cheapest provider first within a tier
3. Health — skip unhealthy providers
4. Capability — provider must support the requested modality and operation

If a provider fails, the router falls back to the next eligible provider.

Streaming

Set "stream": true in the node config to use stream_execute. Chunks are collected and concatenated into the final output.

Streaming protocols by provider:

• OpenAI — SSE (Server-Sent Events)
• Ollama — NDJSON (Newline-Delimited JSON)

Example

NodeDefinition::provider("transcribe", "audio.transcribe", "audio")
    .input("audio_url", "$context.audio.url")
    .output("transcript", "$context.transcript", MergePolicy::Replace)
    .requires("provider.audio.transcribe")

Cost Tracking

Provider executions capture usage metadata in the result:

Field	Description
input_tokens	Number of input tokens consumed
output_tokens	Number of output tokens generated
estimated_cost_usd	Estimated cost in USD

This data flows into Fabric’s cost tracking system, enabling per-request attribution by org, principal, API key, job, and workflow.

Available Providers

Provider	Tier	Modalities	Streaming
OpenAI	Premium	Text, Image (DALL-E), Embedding	Yes (SSE)
Anthropic	Premium	Text	No
Ollama	Basic	Text, Embedding	Yes (NDJSON)
Whisper	Basic	Audio	No
ComfyUI	Basic	Image (Stable Diffusion)	No (async poll)
ONNX	Basic	Configurable	No
Candle	Basic	Text, Embedding	No
Echo	Basic	All (test stub)	No

Cost Estimation

Before executing a provider node, estimate the cost:

curl -X POST http://localhost:3001/v1/providers/estimate \
  -H 'content-type: application/json' \
  -d '{
    "modality": "text",
    "model": "gpt-4",
    "input": {"prompt": "Hello world"}
  }'

Local models (Ollama, Whisper) always return $0.00.