Easy8 AI Knowledge assistant PRO / Easy8 AI Helpdesk - Self-Hosted AI
Use this guide when:
- the customer uses Easy8 AI Knowledge assistant PRO or Easy8 AI Helpdesk
- ChromaDB runs on the customer infrastructure
- the LLM and embeddings services also run on customer infrastructure
Easy8 AI Knowledge assistant PRO and Easy8 AI Helpdesk share the same infrastructure requirements.
Required services
This scenario requires:
chaincelerychroma- local LLM endpoint
- local embeddings endpoint
Environment configuration
Set the common variables from Common configuration, then add:
LLM_URL=http://vllm-host:8081
EMBEDDINGS_URL=http://embeddings-host:8085
CHROMADB_URL=http://chroma:8000
VLLM_API_KEY=your-api-key
If your VLLM endpoint (or reverse proxy in front of it, such as nginx) requires authentication, set VLLM_API_KEY. The value will be sent as Authorization: Bearer <VLLM_API_KEY> header.
Recommendation: We recommend setting the API key in the reverse proxy (e.g., nginx) rather than in VLLM itself. VLLM only protects certain URLs with authentication, but not all endpoints. A reverse proxy can enforce authentication consistently across all requests.
Supported self-hosted LLM
Easy8 AI Knowledge assistant PRO and Easy8 AI Helpdesk currently support:
gemma-4-26b- served through
vLLM
The endpoint configured in LLM_URL must expose gemma-4-26b from a reachable vLLM deployment.
Supported self-hosted embeddings
For self-hosted retrieval, the supported embeddings setup is:
BAAI/bge-m3- served through a TEI endpoint configured by
EMBEDDINGS_URL
The endpoint configured in EMBEDDINGS_URL must expose BAAI/bge-m3 from a reachable TEI deployment.
Embeddings hardware requirements
For BAAI/bge-m3 served through TEI:
- only NVIDIA GPU deployment is supported
- minimum supported GPU memory is
4 GB VRAM - CPU-only deployment is not supported
8 GB VRAMor more is recommended for faster indexing and higher throughput
For the detailed hardware reference, see Embeddings hardware recommendations.
Note that you can run embeddings alongside LLM on a single GPU. See our docker compose example that limits LLM to 90% of GPU memory.
Installation steps
- If the customer is preparing a new GPU host for self-hosted AI, follow Self-hosted AI server setup first.
- Deploy the local LLM endpoint by following the vLLM deployment example.
- Deploy the local embeddings endpoint by following the TEI deployment example.
- Make sure the local LLM and embeddings endpoints are running and reachable from the
chaincontainer. - Add the required environment variables to the Easy8 deployment
.envfile. - Verify the package includes
docker-compose_easy_ai.yml. - Run:
- Confirm
app,chain,chroma,celery, andredisare running. - Configure the reverse proxy for browser access to Easy8 AI before testing activation.
Use the reverse proxy rule from Common configuration.
-
Verify both public chain URLs work through the reverse proxy:
-
https://your.easy.host/ai/chain/sys/monitoring.json https://your.easy.host/ai/chain/api/info
The runtime info endpoint should report the expected local vLLM provider and the embeddings model. See Verification and troubleshooting.
Notes
LLM_URLshould point to a reachablevLLMendpoint servinggemma-4-26bEMBEDDINGS_URLshould point to a reachable TEI endpoint servingBAAI/bge-m3- the TEI deployment for
BAAI/bge-m3should run on NVIDIA GPU infrastructure with at least8 GB VRAM