We wired LightRAG into a full-duplex voice runtime — ask your knowledge base questions by voice

Hi LightRAG team & community — we maintain qwen-audio-agent (2.3k⭐), an open-source realtime full-duplex voice runtime that sits in front of AI coding agents (Qwen Code, Claude Code, Codex, OpenCode, DeepSeek Harness, and more).

We just formalized a replaceable KnowledgeProvider boundary in the Gateway, and LightRAG is our first official external knowledge base integration — shipped as a complete, runnable example: examples/lightrag + a bilingual scenario guide on our user manual site.

What it does

Speak a question → the Gateway retrieves from your independently operated LightRAG server (semantic + graph retrieval via /query/data) → the voice frontend generates the answer in realtime speech. Full-duplex: you can interrupt mid-answer, and documents can be imported from the Knowledge Library panel (upload/list/delete through the LightRAG REST API).

Boundary design (the part we think is interesting)

• LightRAG stays fully in charge of itself: its own LLM, embeddings, documents, indexes, and workspace. Our core npm package contains zero LightRAG or Python dependencies — we never install, start, or modify a LightRAG instance.
• Raw chunks only: the Gateway consumes raw retrieval chunks; track_id values, graph objects, and raw HTTP responses never reach the client or the model.
• Honest async semantics: upload/delete are asynchronous inside LightRAG, but the Gateway reports success only after real completion. Cancelling an ingestion task stops waiting without invoking a global pipeline cancel that could affect other documents.

This is what "replaceable knowledge base" should mean: the voice frontend and the knowledge system evolve independently, connected by one typed provider contract.

Huge thanks to the LightRAG team — a clean REST API with workspace isolation made this integration genuinely pleasant. If anyone runs LightRAG in production, we'd love feedback on the example.

Disclosure: we are qwen-audio-agent maintainers. The integration example is Apache-2.0, same as this repo's license.