All projects

AI · MCP

Nixa AI

Knowledge assistant for the NICE/CXone docs, hybrid search, rerank and a source in every answer. Now also an MCP server that plugs into ChatGPT, Claude and Cursor.

Role
Full-stack
Year
2026
Stack
TypeScriptNext.jsMCPRAGEmbeddingsGeminiOpenAIOllama

The story

Nixa AI was born to solve a simple problem: finding answers in the NICE/CXone documentation usually means browsing dozens of pages, articles and menus before reaching the right information. Instead, you ask a question in natural language and get a direct answer, always with the link to the official document it came from.

Behind it is a RAG pipeline built for technical documentation. A crawler sweeps the official sources and splits everything into passages with context. Search runs on two channels at once, embeddings for the vague question and lexical search for the exact endpoint name, and a rerank decides what rises, weighing page type, source authority and recency.

The rule is no guessing: every answer cites its source, and if the documentation doesn’t cover the topic, Nixa says so instead of making something up. You choose the model, Gemini, OpenAI or Ollama running locally, and the key stays in the browser; the server stores none.

Then it evolved into an MCP server. A read-only endpoint exposes the search_nice_docs tool, so the same knowledge base now serves ChatGPT, Claude, Cursor, Gemini CLI, VS Code and any MCP client: your assistant searches Nixa’s base and answers with its own model. Just point it to the endpoint URL, no key needed.

The same pipeline powers the chat, the API and the MCP, only who writes the final answer changes. And the project is open source: all the code is on GitHub.

What’s inside

  • MCP server with the search_nice_docs tool, works in ChatGPT, Claude, Cursor, Gemini CLI and VS Code
  • Hybrid search: embeddings for the vague question, lexical search for the exact endpoint name
  • Rerank by page type, source authority and recency before building the context
  • A source in every answer, linking the official document, and a warning when the docs don’t cover it
  • Multi-LLM: Gemini, OpenAI or local Ollama, switch whenever you want
  • Key stays in the browser; read-only MCP endpoint with no key required
  • Open source, running entirely on your machine

Gallery

Next projectVidaBR