Skip to main content

ADR-005: Multi-Provider LLM Registry

Status: Accepted Date: 2026-02-24 Deciders: OpsFlow LLC (DTCC/FINOS Hackathon team)

Context​

CALMGuard uses LLMs for all 4 compliance agents. The choice of LLM provider affects cost, latency, output quality, and vendor lock-in. Financial services organizations (like DTCC) have specific vendor approval processes that may require flexibility.

Options considered:

  • A) Single provider (Gemini only) — simplest; creates hard dependency on Google
  • B) Multi-provider with manual switching — flexible but operational burden
  • C) Multi-provider via Vercel AI SDK — unified API across providers; provider swapped via config

Decision​

Use Vercel AI SDK with a multi-provider registry in src/lib/ai/provider-registry.ts. Supported providers: Google Gemini (default), Anthropic Claude, OpenAI GPT, xAI Grok. Provider selection is controlled by the AI_PROVIDER environment variable (or defaults to Gemini).

All agent calls go through the shared provider registry — no provider SDK is imported directly in agent files.

Consequences​

Good:

  • Financial services customers can swap to their approved LLM vendor without code changes
  • Vercel AI SDK abstracts generateObject across all providers — no agent rewrite needed per provider
  • Gemini 2.0 Flash is the fastest/cheapest option for hackathon demo; Claude Opus available for higher-quality output
  • Reduces vendor lock-in risk — a key concern for FINOS open-source adoption

Neutral:

  • 4 provider packages in production dependencies (adds ~200KB to bundle). All are optional — only the active provider's API key is required.

Bad:

  • Provider-specific model capabilities differ (context windows, structured output reliability). Each provider requires validation testing when switching. Not yet automated.