Application
Application observability shows the traffic and cost of one application, and each request it sends through the Gateway. Use it to check that an application is healthy and to find out what happened to a specific request.
Required role
Applies to
Why use it
Overview
The Dashboard for one application
- Is the application healthy?
- What does it cost, and what made the cost change?
- Which models does it use?
Requests
One request at a time
- Why did this request fail, or which check blocked it?
- Which stage took the time?
- What did Guardrails, the Context Engine and Smart Router change?
To compare applications, or to see the whole organization, use Organization.
Overview
ProjectsDefaultSupport Assistant
Support AssistantActive
support-assistant
Shortcut: on the Projects page, switch the list to Applications.
The Overview tab opens first. It starts with a summary of the application's configuration. Below the summary, it shows the Dashboard for this application only. Each metric uses the definition in Metrics, filtered to this application.
Usage
Requests
721 +10.1%
Input tokens
1.42M +14.8%
Output tokens
312K +9.6%
Cache hit rate
38.2% +4.1%
- The Overview opens on Today. Pick the period at the top right. Each application keeps its own period. A change to the Dashboard period does not change it.
- The charts use the Dashboard's controls: Metric, Grouping and Bucket.
- There is no Application or Project grouping.
- Two sections show only the total over time: Governance & Security and Data Integrations. Their metrics split only by application, so these charts have no Grouping control.
Each section links to the definitions of its metrics:
- Usage: requests and tokens.
- Cost: spend and its rates.
- Reliability: outcomes and latency.
- Governance & Security: guardrail activity.
- Data Integrations: retrieval metrics.
- Tools: tool call metrics.
Requests
The Requests tab is next to the Overview. It lists each request that the application sent through the Gateway. Use the Status filter to select Success, Success with fallback, Failed or Blocked. Clear the status filter to show all requests in the selected period. Each row shows the model and provider when the request reached a model, then the latency, the tokens, the cost and the time.
Each row also shows the result of the request:
A Fallback row means that the first model failed and a fallback model returned the response.
Read a trace
Click a row to open the trace of the request. The trace shows each stage of the request as a bar on one timeline. Use the bars to find the stage that took the time. The top of the trace shows the trace ID, the provider key when the request made an LLM call, the endpoint, the time, the total latency, the cost and the number of events.
Gateway 2.84 s
- Status
- 200
- Method
- POST
- Endpoint
- /v1/responses
- Latency
- 2.84 s
Planner 2.83 s
- Smart routing
- No
- Model
- gpt-4.1-mini
- Provider
- OpenAI
- Latency
- 2.83 s
- Zero data retention
- Yes
- Block training providers
- Yes
Guardrails PII 38 ms
- Outcome
- Redacted
- Action
- Redact
- Stage
- Input
- PII entity types
- EMAIL_ADDRESS
- PII matches
- 1
Context build 62 ms
- Token change
- ~−18.4%
- Max input tokens
- 1,047,576
Smart Compression 56 ms
- Action
- Optimize
- Items modified
- 3
- Token change
- ~−18.4%
LLM Call 2.69 s
- Model
- gpt-4.1-mini
- Provider
- OpenAI
- Data residency
- EU
- Latency
- 2.69 s
- Provider key
- Optiak key
- Streaming
- No
- Input tokens
- 1,204
- Cache read
- 0
- Output tokens
- 312
- Total tokens
- 1,516
| Stage | What it shows |
|---|---|
| Gateway | The HTTP status, method, endpoint and latency. For a tool call, it shows the operation and latency. See Gateway. |
| Guardrails | One stage for each check, with its outcome, action, input or output, and block response. A Sensitive Data check also shows the entity types it found. See Guardrails. |
| Planner | The requested model and provider, and the governance rules that applied. See Data Governance. |
| Retrieval | The documents that the request retrieved, with the data source, file, page and chunk for each. See Data Integrations. |
| Context build | The messages and tokens before and after the Context Engine. Each component, such as Smart Compression, has its own stage. See Context Engine. |
| Smart routing | The time that Smart Router took to choose a model. See Smart Router. |
| LLM Call | The model, provider, data residency and provider key. It also shows the tokens, cache use and streaming, and the tools that the model saw, called and received. A BYOK or custom provider key shows its name. See Models. |
| Tool execution | The tool name, provider, outcome, attempt and latency of each tool call. See Tools. |
A trace shows only the stages that ran for that request.
Blocked, redacted and optimized requests
Blocked
An input check stops the request. No model receives it.
Redacted
The check masks personal data. The request continues.
Optimized
The Context Engine removes tokens before the LLM Call.
The trace also shows a request blocked with Refusal as blocked. See Actions. A component in Audit (no action) mode records its result but does not change the request. See Context Engine.
Related
- Metrics: the definition and formula of each number.
- Organization: the traffic of all your applications.
- Gateway overview: the stages a request goes through.
- Members and Access: which roles can see each application and its requests.