@dbcube/ai

Provider-agnostic AI layer for Dbcube. Configure a provider once and get a single, common interface for embeddings and chat/generation — your code never talks to a vendor SDK. This is the L0 foundation of the Dbcube Data + AI ORM (vector search + RAG).

Zero runtime dependencies — every driver uses the native fetch (Node 18+).

Configure once

import { AI } from "@dbcube/ai";

AI.configure({ provider: "openai", apiKey: process.env.OPENAI_API_KEY! });
// or:
AI.configure({ provider: "anthropic", apiKey: process.env.ANTHROPIC_API_KEY!, model: "claude-sonnet-5" });
AI.configure({ provider: "ollama",    url: "http://localhost:11434", model: "llama3" });

Use the common interface

const provider = AI.provider();

const vectors = await provider.embed(["hello world", "another chunk"]); // number[][]
const answer  = await provider.complete("Summarize the theory of relativity.");
const reply   = await provider.chat([
  { role: "system", content: "You are terse." },
  { role: "user",   content: "What is a vector index?" },
]);

Capabilities are explicit

Not every provider does everything. Branch on capabilities instead of guessing:

const p = AI.provider();
if (p.capabilities.embed) {
  await p.embed(["..."]);
}

Note: Anthropic (Claude) has no embeddings API. Its capabilities.embed is false and embed() throws — pair Claude (as the generation/"agent" provider) with an embedding provider (OpenAI or Ollama).

Providers

| Provider | embed | chat / complete | rerank | |---|---|---|---| | openai | ✅ | ✅ | ⬜ (planned) | | anthropic | ❌ (no API) | ✅ | ⬜ | | ollama (local) | ✅ | ✅ | ⬜ | | google (Gemini) | ⬜ planned | ⬜ planned | ⬜ |

License

MIT