An open-source agentic RAG that shows its work
Elysia introduces a transparent, agentic RAG framework that goes beyond “text in, text out” by combining a decision-tree architecture, dynamic data displays, and automatic data awareness. Built as a pip-installable, end-to-end app that connects to Weaviate Cloud, it reduces cost and latency via per-user feedback few-shotting, multi-model routing, and chunk-on-demand retrieval. Strategically, it’s designed to replace Verba and serve as a flexible foundation for production-grade, domain-specific AI agents.
Points clés
- August 12, 2025: the Elysia launch post is authored by Victoria Slocum, Danny Williams, and Edward Schmuhl.
- Elysia is an open-source agentic RAG framework with a FastAPI backend and a full web UI; it installs with pip install elysia-ai and runs via elysia start.
- Its core is a decision-tree of “decision agents” with global context, explicit reasoning, robust error handling (including an “impossible flag”), and a hard pass limit to avoid loops.
- Elysia connects to Weaviate Cloud for retrieval, auto-generates smart filters, and supports hybrid search (vector, keyword, aggregations), with a free 14-day sandbox to get started.
- The frontend can render seven dynamic display types: generic data, tables, e‑commerce product cards, tickets, conversations/messages, documents, and charts.
- An LLM pre-analyzes collections to create summaries, metadata, and display mappings, while users can edit the generated metadata and property mappings.
- The feedback system stores per-user examples in Weaviate, retrieves similar queries via vector similarity, and uses few-shot prompts so smaller models can replicate large-model quality.
- Chunk-on-demand dynamically splits long documents at query time, storing quantized chunks cross-referenced to originals to cut storage while improving retrieval over time.
- A multi-model strategy routes tasks by complexity; Gemini is the default test model for its large context and speed, and tools can auto-trigger at a 50,000-token context threshold.
- Real-world demo: Glowe’s AI skincare chat runs on Elysia with three custom tools (query agent, stack generation, similar products), and Elysia is slated to replace Verba as Weaviate’s RAG app.
À retenir
Want results without a PhD in prompt whispering? Start with the Weaviate sandbox, pip install elysia-ai, and let the decision tree do the heavy lifting. Flip on the feedback setting so your future answers get faster and cheaper (your wallet will high‑five you), and lean on chunk‑on‑demand to avoid turning your storage bill into a horror story. Finally, watch the tree view when things go sideways—seeing the agent’s reasoning beats guessing, and yes, you can still customize the blob for style points.
Sources
