Phase 5: LLM transforms and RAG data preparation

[netsirius]

Summary

Adds LLM-as-transformation capabilities and RAG data preparation transforms.

LLM Transforms

• LLMTransformPlugin: generic transform that calls any LLM API with prompt templating, structured JSON output, batching, content-hash caching, and retry with exponential backoff
• Supports claude, openai, local (Ollama), and any OpenAI-compatible API via custom baseUrl

RAG Transforms

• ChunkingPlugin: split documents into chunks (fixed, sentence, recursive strategies)
• EmbeddingPlugin: generate vector embeddings via OpenAI/Vertex AI/Cohere APIs
• GraphExtractionPlugin: extract entities and relationships using LLM

Local Model Support

All LLM plugins support provider: local for Ollama (http://localhost:11434). No API key required. Custom baseUrl supported for self-hosted endpoints.

Example YAML

transformations:
  - id: classify
    type: LLMTransform
    sources: [data]
    config:
      provider: local
      model: llama3
      prompt: "Classify: {text}"
      outputSchema: "category:string|confidence:double"
      batchSize: 5
      cache: true

Test plan

• sbt compile — All modules compile
• CI green
• Manual: LLMTransform with Ollama local model