Qwen Audio Agent

English

简体中文

English

简体中文

Appearance

LightRAG Knowledge Base ​

qwen-audio-agent can use an independently operated LightRAG instance as frontend knowledge. LightRAG owns document parsing, embeddings, indexing, and retrieval; the Gateway calls it only through the KnowledgeProvider boundary.

When to use it ​

  • You want richer semantic and graph retrieval than the built-in keyword library.
  • You already use LightRAG and want the voice frontend to reuse its documents and model setup.
  • You want the knowledge system to evolve or be replaced independently of Realtime and clients.

Prepare LightRAG ​

LightRAG requires an independently configured LLM and embedding model. Either can run locally or through an OpenAI-compatible service. Install the Server with uv:

bash
uv tool install "lightrag-hku[api]"

Create .env in the LightRAG working directory with this general shape:

dotenv
LLM_BINDING=openai
LLM_BINDING_HOST=https://your-openai-compatible-service.example/v1
LLM_BINDING_API_KEY=your_llm_key
LLM_MODEL=your_llm_model

EMBEDDING_BINDING=openai
EMBEDDING_BINDING_HOST=https://your-openai-compatible-service.example/v1
EMBEDDING_BINDING_API_KEY=your_embedding_key
EMBEDDING_MODEL=your_embedding_model
EMBEDDING_DIM=1024

LIGHTRAG_API_KEY=your_lightrag_api_key

The embedding dimension must match the model. Follow the official LightRAG documentation for the complete configuration. Start a localhost-only service:

bash
lightrag-server --host 127.0.0.1 --port 9621

Run the integration example ​

From a qwen-audio-agent source checkout:

bash
cp examples/lightrag/.env.example examples/lightrag/.env.local

Set the voice frontend and LightRAG connection values:

dotenv
QWEN_AUDIO_REALTIME_API_KEY=your_dashscope_api_key
LIGHTRAG_URL=http://127.0.0.1:9621
LIGHTRAG_API_KEY=your_lightrag_api_key
LIGHTRAG_WORKSPACE=
LIGHTRAG_QUERY_MODE=mix

Start the example:

bash
node --env-file=examples/lightrag/.env.local examples/lightrag/gateway.mjs

Open http://127.0.0.1:3101, import a file from the Knowledge Library panel, wait for indexing, then query it by voice or text. This isolated example uses frontend-only mode and does not start a backend Agent.

QWEN_AUDIO_REALTIME_API_KEY belongs only to the voice frontend. LightRAG keeps its LLM and embedding configuration in its own process. qwen-audio-agent does not install, start, or modify LightRAG.

Data and task boundaries ​

  • The Gateway obtains raw chunks from /query/data; the voice frontend generates the answer.
  • LightRAG track_id values, graph objects, and HTTP responses do not reach the client or model.
  • Upload and deletion are asynchronous inside LightRAG, but the Gateway reports success only after real completion.
  • Cancelling a Gateway ingestion task stops waiting without invoking a global pipeline cancel that could affect other documents.
  • Users remain responsible for LightRAG documents, indexes, credentials, and workspace storage.

See examples/lightrag for the full code and configuration reference.

Authors and acknowledgements ​

  • Thanks to the LightRAG project and its contributors for open-sourcing the document processing, knowledge graph, and retrieval capabilities.
  • Li Xu designed the replaceable KnowledgeProvider boundary and implemented this integration example.