JñānaSetu
A multi-tenant RAG platform: self-hosted embeddings and Qwen3-4B answer from a client’s own documents and show where every answer came from.
Clients upload their knowledge, embed one chat widget on their site, and customers get grounded answers with verified sources. Identity, ingestion, retrieval, inference and chat are separate modules in one Go API; the embedding worker and the LLM run self-hosted on a private network, with Supabase as the managed data layer.
Proves
- Go services
- React + TypeScript + Vite
- Vector search
- Multi-tenant security
- Self-hosted LLMs
Decisions that carry it
- Tenant trust stays server-side. Module 1 resolves the organization from the widget ID. A tenant ID sent by the browser is never treated as proof of anything.
- Atomic, versioned indexing. Every upload is an immutable version. Chunks are staged inactive and the complete version is activated in one transaction, so search never sees half an index.
- Grounded, or honest about it. Retrieval filters by tenant and active version. Weak evidence returns a configured fallback with no sources, and sources come only from retrieval metadata, never from model output.
- Private AI runtime. all-MiniLM-L6-v2 embeddings (384 dimensions) and Qwen3-4B on llama.cpp run self-hosted, never exposed publicly. Only the database, auth and storage are managed.
- Operable by design. Authenticated metrics, traces, durable audit events, and retention jobs that purge chats after 90 days and operational events after 180.
Built with
- Go
- React + Vite + TypeScript
- Supabase Postgres
- pgvector (HNSW)
- llama.cpp
- Qwen3-4B
- sentence-transformers
- FastAPI
- Docker Compose
- Prometheus + Grafana
- OpenTelemetry