name: Argilla FinOps Framework
description: >
  FinOps considerations for teams using Argilla as a data annotation and
  dataset curation platform. Since Argilla is open-source and self-hosted,
  direct costs arise from infrastructure rather than software licensing.
  Deployments on Hugging Face Spaces incur compute costs billed by Hugging Face.
specificationVersion: "1.0"
provider: Argilla
focus:
  schema: FOCUS 1.0
  currency: USD
costCategories:
  - category: Infrastructure
    description: >
      Compute, storage, and networking costs for self-hosted Argilla deployments.
      Argilla requires a FastAPI server, a relational database (SQLite or
      PostgreSQL), and optionally ElasticSearch or OpenSearch for vector search.
    costDrivers:
      - name: Compute (Server)
        description: CPU/RAM for the Argilla FastAPI server process
        unit: instance-hours
        notes: Scales with concurrent users and dataset size
      - name: Database Storage
        description: PostgreSQL or SQLite storage for records, responses, and metadata
        unit: GB-month
        notes: Grows linearly with number of annotated records
      - name: Search Index
        description: ElasticSearch or OpenSearch cluster for vector search and filtering
        unit: GB-month
        notes: Optional; required for semantic search and large dataset filtering
      - name: Networking
        description: Egress bandwidth for API calls and Hugging Face Hub exports
        unit: GB
  - category: Hugging Face Spaces Deployment
    description: >
      When deploying Argilla via Hugging Face Spaces, compute costs are billed
      by Hugging Face at their standard hardware tier rates. No Argilla
      licensing fee applies.
    costDrivers:
      - name: Spaces CPU Hardware
        description: Free or paid CPU instances on Hugging Face Spaces
        unit: hours
        notes: Free tier available with limited vCPU and RAM
      - name: Spaces GPU Hardware
        description: GPU instances for accelerated annotation suggestion models
        unit: hours
        notes: Optional; billed per Hugging Face GPU pricing
  - category: Data Export and Integration
    description: >
      Costs associated with exporting datasets to Hugging Face Hub or other
      destinations and running suggestion/pre-annotation models.
    costDrivers:
      - name: Hugging Face Hub Storage
        description: Dataset storage on Hugging Face Hub after export via dataset.to_hub()
        unit: GB-month
        notes: Free tier available on Hugging Face Hub
      - name: Model Inference for Suggestions
        description: >
          Compute cost for running LLMs or NLP models to generate annotation
          suggestions (Distilabel integration or external inference endpoints)
        unit: tokens or inference-calls
optimization:
  - recommendation: >
      Use SQLite for small teams and single-node deployments to eliminate
      PostgreSQL infrastructure overhead.
  - recommendation: >
      Defer ElasticSearch/OpenSearch until dataset size exceeds tens of
      thousands of records; smaller datasets perform adequately with SQLite FTS.
  - recommendation: >
      Use Hugging Face Spaces free-tier hardware for evaluation and low-volume
      annotation projects to reduce compute spend.
  - recommendation: >
      Leverage the Distilabel library for offline batch pre-annotation rather
      than real-time inference to reduce per-record model inference costs.
notes: >
  Argilla is joining Hugging Face; future managed pricing models may consolidate
  infrastructure costs into a unified Hugging Face billing account.
  Monitor https://argilla.io/blog for announcements.