Skip to content
Almathal Handbook

ADR-0021: AI/LLM Slice as First-Class Concern

Accepted

Date: 2026-06-10

Context

Enterprise software demand in 2025–2026 is dominated by AI-augmented applications: RAG over private corpora, AI-assisted workflows, intelligent ticket routing, document processing pipelines, agent-based automation. A platform that targets enterprise without first-class AI building blocks ships for the wrong decade.

Decision

The Component Library includes an AI/LLM slice alongside Java, TS/JS, and Python slices. Adapters in this slice include:

  • LLM provider SDKs (Anthropic, OpenAI, Bedrock, Vertex, Ollama, vLLM)
  • Vector databases (pgvector, Qdrant, Weaviate, Milvus, Chroma)
  • Embedding model wrappers (sentence-transformers, OpenAI embeddings, Cohere, Voyage)
  • Orchestration frameworks (LangChain, LlamaIndex, Haystack, DSPy)
  • Document processing (unstructured.io, PyMuPDF, Apache Tika)
  • Retrieval and reranking (BM25, Cohere Rerank, cross-encoders)
  • Evaluation (RAGAS, DeepEval, promptfoo)
  • Guardrails (Guardrails AI, NeMo Guardrails)
  • Agent frameworks (LangGraph, CrewAI, AutoGen)
  • Streaming/UX (Vercel AI SDK, SSE)

MVP targets ~30–40 AI Adapters, weighted toward the most-used.

Rationale

  • Enterprise demand is dominated by AI-augmented apps; ignoring this slice produces a misaligned product.
  • The AI ecosystem moves fast and has weirder licensing — exactly the environment where curation adds value.
  • AI Adapters anchor the RAG Chatbot MVP archetype (ADR-0022) and several follow-on archetypes.

Consequences

  • The Curation Pipeline must handle AI Adapter volatility (frequent releases, deprecations, license changes).
  • The Capability namespace includes AI-specific categories (llm-completion, vector-search, embedding, etc.).
  • Cross-language AI seams (Python ingestion → Java API → React UI) become a first-class concern in early Archetypes.
  • License vetting matters more here than in stable Java/TS/Python ecosystems.

References