Skip to content

============================================================================

ShadowHound Configuration for vLLM on Thor

============================================================================

This configuration is optimized for using vLLM running on Thor (Jetson).

NOTE: The agent will AUTO-DETECT that you're using a local LLM (non-OpenAI

base URL) and automatically use local embeddings. You don't NEED to set

USE_LOCAL_EMBEDDINGS=true, but you can if you want to be explicit.

============================================================================

Agent Configuration - vLLM on Thor

AGENT_BACKEND=openai OPENAI_BASE_URL=http://192.168.10.116:8000/v1 OPENAI_MODEL=Qwen/Qwen2.5-Coder-7B-Instruct USE_PLANNING_AGENT=false

API key (required by DIMOS but not used with vLLM)

OPENAI_API_KEY=sk-dummy-key-for-vllm

Embeddings (OPTIONAL - auto-detected)

The agent automatically uses local embeddings when OPENAI_BASE_URL

is NOT api.openai.com. Uncomment to override:

USE_LOCAL_EMBEDDINGS=true

Optional RAG settings

RAG_QUERY_N=3 RAG_SIMILARITY_THRESHOLD=0.4

Robot configuration

ROBOT_IP=192.168.1.103 CONN_TYPE=webrtc MOCK_ROBOT=false

ROS configuration

ROS_DOMAIN_ID=0 RMW_IMPLEMENTATION=rmw_cyclonedds_cpp

Web interface

ENABLE_WEB_INTERFACE=true WEB_PORT=8080 WEB_HOST=0.0.0.0

Logging

LOG_LEVEL=DEBUG DIMOS_DEBUG=true

============================================================================

Troubleshooting

============================================================================

Error: "ValueError: No embedding data received"

→ Solution: Set USE_LOCAL_EMBEDDINGS=true (vLLM doesn't support embeddings)

Error: "Connection refused"

→ Verify vLLM running: curl http://192.168.10.116:8000/health

Error: "401 Unauthorized" during vLLM startup

→ Run: huggingface-cli login

→ Accept: https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct

============================================================================