PrivateGPT 1.0 turns local models into production AI backends with agentic RAG

PrivateGPT 1.0.0 is out, merging two years of work from Zylon.ai’s private fork back into the open-source project that has amassed 57.2k stars and 7.6k forks on GitHub. The original PrivateGPT was essentially a semantic search pipeline — ingest documents, embed chunks, retrieve by similarity, and pass context to a local model. That proved the concept; this new version is the full production application backend.

Key additions include a standard messages API, file and artifact ingestion, retrieval with citations, agentic RAG (retrieval-augmented generation where the AI can act), built-in tools mirroring those in the Claude API, custom tools support, MCP connectors, structured access to databases and CSVs, web search and extraction, and code execution. The architecture is now decoupled: PrivateGPT 1.0 acts as a middleware layer between your app/agent/UI and any OpenAI-compatible inference server (Ollama, vLLM, llama.cpp, etc.). Zylon.ai, the commercial product built on PrivateGPT, will now develop against this open-source repo, flowing commercial iterations back into the public project. Note: the existing PrivateGPT API is not forward-compatible — users must follow the migration guide.