Multi-tenant AI assistant with RAG, embeddings and per-organization knowledge base.
Tuki is a multi-tenant AI assistant platform that allows each organization to have its own assistant trained on its knowledge base, with vector embeddings, RAG over proprietary documents and an admin panel.
The challenge
Organizations need AI assistants that respond with their own current information, not just the general knowledge of LLMs. The primary use case was AAETAV: an assistant that could answer questions about courses, certifications, requirements and processes using the association's official documentation.
The technical challenge was building a multi-tenant RAG (Retrieval-Augmented Generation) system where each organization could upload their documents, embeddings would be stored in PostgreSQL with pgvector, and the LLM (Claude API) would respond citing sources from the tenant-specific knowledge base.
The solution
We developed Tuki with Next.js 14, pgvector in PostgreSQL for embedding storage, Voyage AI for high-quality embedding generation, and Claude (Anthropic) as the main LLM. The architecture includes: document ingestion with intelligent chunking, cosine similarity vector search, dynamic context per tenant, conversation history, admin panel for uploading documents and configuring the assistant, and an embeddable widget for integration on any site.
The AAETAV tenant was the first in production at tuki.noia.guru, with the association's complete knowledge base.