A customer-service bot that grades its own homework

What it does

Simba is a self-hosted customer-service assistant built around a customizable RAG pipeline. You feed it your knowledge base, embed it into a website via an npm chat widget, and let users query an LLM that draws from your documents rather than generic training data. A dashboard lets you manage documents and monitor conversations.

The interesting bit

The project treats evaluation as a first-class feature, not an afterthought. It tracks retrieval precision, answer faithfulness, and latency out of the box—essentially shipping with built-in report cards so you can swap models or chunking strategies with actual numbers instead of vibes.

Key highlights

• Swapable pipeline components: vector stores (Qdrant, FAISS, Chroma), embeddings (OpenAI, HuggingFace, Cohere), LLMs (OpenAI, Anthropic, local), rerankers, and document parsers.
• Async ingestion via Celery and Redis, with streaming responses from a FastAPI backend.
• Drop-in React chat widget via simba-chat-widget.
• Docker deployment with CPU or NVIDIA GPU support.
• Conversation analytics include user satisfaction and resolution rates, though the README doesn’t explain how these are measured.

Caveats

• Multi-tenant support, advanced analytics, webhooks, and a fine-tuning pipeline are on the roadmap but remain unchecked.
• The README calls the stack “production-ready” with streaming and async processing, yet several roadmap features are still missing.

Verdict

Worth a look if you want a self-hosted support bot with visible RAG metrics. Skip it if you need multi-tenant SaaS features today or prefer a fully managed black box.