Configuration Reference

This page documents every configuration option for the worker node, indexer, and dashboard.

Worker Configuration

The worker reads a YAML configuration file at startup:

dfpn-worker --config config.yaml

Two example files are provided in the repository root:

config.yaml -- Full GPU node with all modalities
config-cpu.yaml -- CPU-only node (reduced concurrency, no video)

Network

network: devnet          # devnet | testnet | mainnet
rpc_url: https://api.devnet.solana.com
indexer_url: https://indexer.devnet.dfpn.network

Key	Type	Default	Description
`network`	string	`devnet`	Solana cluster to connect to
`rpc_url`	string	--	Solana JSON-RPC endpoint
`indexer_url`	string	--	DFPN indexer base URL

Wallet

wallet_path: ~/.config/solana/id.json

Key	Type	Description
`wallet_path`	string	Path to the Solana keypair JSON file used to sign transactions

Worker Settings

worker:
  modalities:
    - image_authenticity
    - video_authenticity
    - face_manipulation
    - voice_cloning
    - generated_content
  min_fee: 1000000
  max_concurrent: 4
  task_timeout: 300
  poll_interval_ms: 5000

Key	Type	Default	Description
`worker.modalities`	list	--	Detection modalities this node can process. Must match the models configured below.
`worker.min_fee`	integer	`1000000`	Minimum fee in lamports to accept a task (1,000,000 = 0.001 SOL)
`worker.max_concurrent`	integer	`4`	Maximum tasks processed in parallel. Reduce for CPU or low-VRAM GPUs.
`worker.task_timeout`	integer	`300`	Seconds before a task is considered timed out
`worker.poll_interval_ms`	integer	`5000`	How often to poll for new tasks (milliseconds)

Models

Each entry in the models list configures one detection model.

models:
  - id: "face-forensics-sbi"
    path: ./models/face-forensics
    modalities:
      - face_manipulation
    gpu_required: true
    runtime: external        # external (Python subprocess) | http (microservice)
    on_chain_id: null        # Set after registering model on-chain

Key	Type	Description
`models[].id`	string	Unique identifier for this model
`models[].path`	string	Local filesystem path or HTTP URL for the model runtime
`models[].modalities`	list	Modalities this model handles
`models[].gpu_required`	bool	Whether the model needs a GPU
`models[].runtime`	string	`external` for Python subprocess, `http` for a sidecar microservice
`models[].on_chain_id`	string	On-chain model account public key (set after registration)

Inference

inference:
  device: cuda       # cuda | cpu
  batch_size: 1
  precision: fp16    # fp32 | fp16 | int8

Key	Type	Default	Description
`inference.device`	string	`cuda`	PyTorch device. Use `cpu` for CPU-only nodes.
`inference.batch_size`	integer	`1`	Inference batch size. Keep at 1 for single-item processing.
`inference.precision`	string	`fp16`	Floating-point precision. `fp16` saves VRAM on GPU; use `fp32` on CPU.

Storage

storage:
  temp_dir: /tmp/dfpn
  max_file_size_mb: 500
  cleanup_after_seconds: 3600

Key	Type	Default	Description
`storage.temp_dir`	string	`/tmp/dfpn`	Directory for downloaded media files
`storage.max_file_size_mb`	integer	`500`	Maximum file size to process (MB)
`storage.cleanup_after_seconds`	integer	`3600`	Remove temporary files after this many seconds

Monitoring

monitoring:
  metrics_port: 9090
  health_port: 8080
  log_level: info

Key	Type	Default	Description
`monitoring.metrics_port`	integer	`9090`	Prometheus metrics endpoint port
`monitoring.health_port`	integer	`8080`	Health check endpoint port
`monitoring.log_level`	string	`info`	Log verbosity: `trace`, `debug`, `info`, `warn`, `error`

CPU-only Configuration

The config-cpu.yaml file adjusts several defaults for nodes without a GPU:

Setting	GPU Value	CPU Value	Reason
`inference.device`	`cuda`	`cpu`	No GPU available
`inference.precision`	`fp16`	`fp32`	Half-precision has no benefit on CPU
`worker.max_concurrent`	`4`	`2`	CPU inference is slower; fewer parallel tasks avoid overload
`worker.task_timeout`	`300`	`600`	Longer timeout accounts for slower processing
`worker.poll_interval_ms`	`5000`	`10000`	Reduced polling frequency
`storage.max_file_size_mb`	`500`	`200`	Smaller cap to limit CPU processing time
Video modality	Enabled	Disabled	Video analysis on CPU takes ~30 s per clip and is impractical

Video on CPU

The video_authenticity modality is commented out in the CPU config and the video-ftcn model is removed entirely. You can re-enable it, but expect latencies of 30 seconds or more per video.

Indexer Configuration

The indexer is configured entirely through CLI flags. It does not use a config file.

dfpn-indexer \
  --rpc-url https://api.devnet.solana.com \
  --index-path ./data/indexes \
  --bind 127.0.0.1:3030 \
  --log-level info \
  --content-registry GokivDYuQXPZCWRkwMhdH2h91KpDQXBEmpgBgs55bnpH \
  --analysis-marketplace 9xQeWvG816bUx9EPjHmaT23yvVM2ZWbrrpZb9PusVFin \
  --model-registry Fg6PaFpoGXkYsidMpWTK6W2BeZ7FEfcYkg476zPFsLnS \
  --worker-registry HmbTLCmaGvZhKnn1Zfa1JVnp7vkMV4DYVxPLWBVoN65L \
  --rewards 4uQeVj5tqViQh7yWWGStvkEG1Zmhx6uasJtWCJziofM

CLI Flags

Flag	Default	Description
`--rpc-url`, `-r`	`http://localhost:8899`	Solana JSON-RPC endpoint
`--index-path`, `-i`	`./data/indexes`	Directory for Tantivy search indexes
`--bind`, `-b`	`127.0.0.1:3030`	Address and port for the REST API server
`--log-level`	`info`	Log verbosity: `trace`, `debug`, `info`, `warn`, `error`
`--content-registry`	--	Content Registry program ID
`--analysis-marketplace`	--	Analysis Marketplace program ID
`--model-registry`	--	Model Registry program ID
`--worker-registry`	--	Worker Registry program ID
`--rewards`	--	Rewards program ID

Docker Environment Variables

When deploying the dashboard + indexer via Docker, program IDs and the RPC URL are passed as environment variables.

Variable	Description
`SOLANA_RPC_URL`	Solana JSON-RPC endpoint
`CONTENT_REGISTRY`	Content Registry program ID
`ANALYSIS_MARKETPLACE`	Analysis Marketplace program ID
`MODEL_REGISTRY`	Model Registry program ID
`WORKER_REGISTRY`	Worker Registry program ID
`REWARDS`	Rewards program ID

Docker Compose snippet

services:
  dfpn:
    image: dfpn-dashboard:latest
    ports:
      - "80:80"
    environment:
      SOLANA_RPC_URL: https://api.devnet.solana.com
      CONTENT_REGISTRY: GokivDYuQXPZCWRkwMhdH2h91KpDQXBEmpgBgs55bnpH
      ANALYSIS_MARKETPLACE: 9xQeWvG816bUx9EPjHmaT23yvVM2ZWbrrpZb9PusVFin
      MODEL_REGISTRY: Fg6PaFpoGXkYsidMpWTK6W2BeZ7FEfcYkg476zPFsLnS
      WORKER_REGISTRY: HmbTLCmaGvZhKnn1Zfa1JVnp7vkMV4DYVxPLWBVoN65L
      REWARDS: 4uQeVj5tqViQh7yWWGStvkEG1Zmhx6uasJtWCJziofM

Dashboard Configuration

The Vue.js dashboard uses Vite for development and nginx in production.

Development Proxy

During local development, Vite proxies API calls to the indexer:

// vite.config.ts
server: {
  proxy: {
    '/api': {
      target: 'http://127.0.0.1:3030',
      changeOrigin: true,
      rewrite: (path) => path.replace(/^\/api/, ''),
    },
  },
},

In production, nginx handles the same proxy rule (/api/ -> http://127.0.0.1:3030/).

Network Selector

The dashboard includes a network selector that switches between Solana clusters:

Network	RPC Endpoint	Description
Devnet	`https://api.devnet.solana.com`	Development and testing
Testnet	`https://api.testnet.solana.com`	Pre-production pilot
Mainnet	`https://api.mainnet-beta.solana.com`	Production (when available)

Switching networks updates the RPC connection and the set of program IDs used for on-chain queries.