LobeChat → OpenObserve

Capture per-request latency and status codes for every chat completion processed by LobeChat. LobeChat is a self-hosted ChatGPT-like application with built-in OpenTelemetry support. Pass OTLP environment variables at startup and traces flow directly into OpenObserve.

Prerequisites

Docker
An OpenObserve account (cloud or self-hosted)
Your OpenObserve organisation ID and Base64-encoded auth token
An API key for at least one LLM provider (e.g. OpenAI)

Installation

Pull the LobeChat image:

docker pull lobehub/lobe-chat:latest

Configuration

Start LobeChat with ENABLE_TELEMETRY=1 and the standard OTLP environment variables. The exporter appends /v1/traces to OTEL_EXPORTER_OTLP_ENDPOINT automatically, so set the endpoint without the path suffix.

docker run -d --name lobechat \
  -p 3210:3210 \
  -e OPENAI_API_KEY=your_openai_api_key \
  -e OPENAI_MODEL_LIST=gpt-4o-mini \
  -e ENABLE_TELEMETRY=1 \
  -e OTEL_EXPORTER_OTLP_ENDPOINT=http://host.docker.internal:5080/api/default \
  -e "OTEL_EXPORTER_OTLP_HEADERS=Authorization=Basic <your_base64_token>" \
  -e OTEL_SERVICE_NAME=lobechat \
  lobehub/lobe-chat:latest

For OpenObserve Cloud replace the endpoint and token:

OTEL_EXPORTER_OTLP_ENDPOINT=https://api.openobserve.ai/api/your_org_id
OTEL_EXPORTER_OTLP_HEADERS=Authorization=Basic <your_base64_token>

OPENAI_MODEL_LIST restricts the available models. Without it LobeChat may default to a model your account cannot access, causing all requests to fail with no spans emitted.

Viewing Traces

Open LobeChat at http://localhost:3210, select a model, and send a chat message. Each request appears in OpenObserve as a server span on the /webapi/chat/openai route.

Log in to OpenObserve and navigate to Traces
Filter by service_name = lobechat
Each span covers the full round-trip from the browser request to the LLM provider response
Use http_status_code to identify failed requests
Sort by duration to find slow completions

LobeChat trace in OpenObserve

What Gets Captured

LobeChat exports HTTP server spans — one span per chat request. These are infrastructure-level spans: they capture route, latency, and status but not LLM-specific fields like token counts or model name.

Attribute	Description
`operation_name`	`POST`
`http_target`	`/webapi/chat/openai`
`http_method`	`POST`
`http_status_code`	HTTP response code (e.g. `200`)
`http_status_text`	`OK` on success
`http_url`	Full request URL
`http_host`	Host and port
`http_request_content_length_uncompressed`	Request body size in bytes
`http_user_agent`	Browser or client user agent
`net_transport`	`ip_tcp`
`span_kind`	`Server`
`span_status`	`UNSET` on success, `ERROR` on failure
`duration`	End-to-end request latency
`service_service_version`	LobeChat version (e.g. `1.143.3`)
`service_process_runtime_name`	`nodejs`

Next Steps

With LobeChat instrumented, every chat request is recorded in OpenObserve. From here you can monitor response latency, track error rates by route, and set up alerts when the chat API degrades.