Vercel AI SDK — the multi-provider LLM abstraction (streaming, tool calling, structured output). Powers the chatbot and the Graph RAG pipeline. See the ai-tools skill and docs/blueprint/ai/.
Why was it chosen?
- One API across providers — chat routes Groq / OpenAI / Google Gemini behind it; connect or swap a provider under Admin → AI → Models without touching call sites (
src/lib/server/ai/connections.tsbuilds the registry from the saved rows). Embeddings use Google Gemini (gemini-embedding-001, 1536-dim) over the same saved Google connection. There is no image-generation provider. - Handles RAG natively (Retrieve → Augment → Generate) — no LangChain needed since queries aren't LLM-generated.
Stack integration: AI SDK is the LLM abstraction; Neo4j holds the knowledge graph (Graph RAG); PostgreSQL stores embeddings.
Known limitations
- Pin exact versions — breaking changes land between major versions (the project pins AI SDK v6).
- AI SDK RSC is deprecated; use AI SDK UI.
Related
- ../data/neo4j.md - Graph database for RAG
- ../ops/deployment.md - Deployment targets
- ../ops/logging.md - AI request logging