Celeste + Idun RAG agent

A RAG pipeline that answers questions about internal HR documents (French PDFs). Uses Celeste AI for generation and embeddings, LangGraph for orchestration, and Idun Agent Platform to serve and manage the agent.

How it works

rewrite_query → retrieve → generate

1. Rewrites the user's question into a vector-search-friendly query via Celeste (gemini-3-flash-preview)
2. Searches an in-memory vector store built from the PDFs using batch Celeste embeddings (gemini-embedding-2-preview)
3. Generates a grounded answer from the retrieved document chunks

Project structure

main.py              # Graph nodes and wiring
state.py             # InputState, GraphState, OutputState
celeste_adapter.py   # CelesteEmbeddings adapter + call_celeste wrapper
docs/                # HR PDF documents (NexaTech)
docs/blog/           # Blog post
tests/               # Lightweight adapter tests

Setup

uv sync

Set your API keys in .env:

GOOGLE_API_KEY=...

The Celeste calls use model-based provider inference, so no provider-specific SDK setup is needed beyond the API key.

Run

idun agent serve --source=manager

The agent fetches its config from Idun Cloud and starts on http://localhost:8800.

Links

• Celeste AI docs
• Celeste AI PyPI
• Idun Agent Platform
• Idun Cloud
• LangGraph docs