> ## Documentation Index > Fetch the complete documentation index at: https://docs.costgraph.ai/llms.txt > Use this file to discover all available pages before exploring further. # Get AI serving deployment metrics > Returns latency quantile, capacity and throughput time-series per model for one serving deployment, sourced from the metrics store. Each series states its own unit because gateways disagree: vLLM publishes seconds and APISIX milliseconds, and a histogram cannot be rescaled after recording. Signals the deployment's gateway does not publish are named in missing_signals rather than returned as zero. A range of 24h or less is answered from the per-replica series; a longer range is answered from the deployment-grain series, which carries no replica identity, so a replica filter beyond 24h is rejected. The response also states how the deployment's pods map to accelerators, and never divides a card's figure between pods. cost_per_mtok is the accelerator spend behind a million output tokens, at each point of the window; it is returned only where the deployment's accelerator rate is known and the deployment is serving output tokens, and is named in missing_signals otherwise. gateway_connection says why a deployment has no series, since the three causes need three different actions: not_connected when no model server is connected for this namespace, selector_mismatch when one is but this deployment's pods do not carry the labels it is read by, naming that selector, and connected when the deployment is being read and any silence is the gateway's own. missing_signal_notes explains a missing signal the connected server only started publishing in a later release, so a request-stage breakdown an older build never emits reads as a version gap rather than as a broken scrape. ## OpenAPI ````yaml /api-reference/costgraph/openapi.json get /api/v1/tenant/ai/deployments/{id}/metrics openapi: 3.0.0 info: description: Read and manage your CostGraph organization, spend, alerts, and settings. title: CostGraph API contact: {} version: '1.0' servers: - url: https://api.costgraph.ai security: [] tags: - name: ai x-group: AI - name: ai-serving x-group: AI serving - name: anomalies x-group: Anomalies - name: auth x-group: Auth - name: billing x-group: Billing - name: billing-export x-group: Billing export - name: budgets x-group: Budgets - name: ci x-group: CI - name: compute-recommendations x-group: Compute recommendations - name: config x-group: Config - name: cost x-group: Cost - name: gpus x-group: GPUs - name: graphai x-group: Graph AI - name: infracost x-group: Infracost - name: integrations x-group: Integrations - name: invitations x-group: Invitations - name: kubernetes-clusters x-group: Kubernetes clusters - name: marketplace x-group: Marketplace - name: network-requests x-group: Network requests - name: notifications x-group: Notifications - name: oauth x-group: OAuth - name: oauth-clients x-group: OAuth clients - name: opencost x-group: OpenCost - name: organization x-group: Audit log - name: organizations x-group: Organizations - name: placement-alternatives x-group: Placement alternatives - name: reports x-group: Reports - name: service-map x-group: Service map - name: settings x-group: Settings - name: sso x-group: Single sign-on - name: tenants x-group: Tenants - name: user x-group: Users - name: virtual-machines x-group: Virtual machines - name: virtual-tags x-group: Virtual tags - name: workflows x-group: Workflows paths: /api/v1/tenant/ai/deployments/{id}/metrics: get: tags: - ai-serving summary: Get AI serving deployment metrics description: >- Returns latency quantile, capacity and throughput time-series per model for one serving deployment, sourced from the metrics store. Each series states its own unit because gateways disagree: vLLM publishes seconds and APISIX milliseconds, and a histogram cannot be rescaled after recording. Signals the deployment's gateway does not publish are named in missing_signals rather than returned as zero. A range of 24h or less is answered from the per-replica series; a longer range is answered from the deployment-grain series, which carries no replica identity, so a replica filter beyond 24h is rejected. The response also states how the deployment's pods map to accelerators, and never divides a card's figure between pods. cost_per_mtok is the accelerator spend behind a million output tokens, at each point of the window; it is returned only where the deployment's accelerator rate is known and the deployment is serving output tokens, and is named in missing_signals otherwise. gateway_connection says why a deployment has no series, since the three causes need three different actions: not_connected when no model server is connected for this namespace, selector_mismatch when one is but this deployment's pods do not carry the labels it is read by, naming that selector, and connected when the deployment is being read and any silence is the gateway's own. missing_signal_notes explains a missing signal the connected server only started publishing in a later release, so a request-stage breakdown an older build never emits reads as a version gap rather than as a broken scrape. parameters: - description: Tenant ID name: X-CostGraph-Tenant-ID in: header required: true schema: type: string - description: AI serving deployment ID name: id in: path required: true schema: type: string - description: Time window (e.g. 1h, 6h, 24h, 7d). Default 1h name: range in: query schema: type: string - description: Resolution step (e.g. 30s, 1m, 5m) name: step in: query schema: type: string - description: >- Serving signals (comma-separated): all,time_to_first_token,inter_token,end_to_end,queue,prefill,decode,request_output_tokens,kv_cache_utilization,batch_size,queue_depth,preemptions,concurrency,tokens_per_second,cost_per_mtok. Default all name: signals in: query schema: type: string - description: Quantiles between 0 and 1 (comma-separated). Default 0.5,0.95,0.99 name: quantiles in: query schema: type: string - description: Restrict to one model name: model in: query schema: type: string - description: Restrict to one replica, as namespace/pod name: replica in: query schema: type: string responses: '200': description: OK content: application/json: schema: allOf: - $ref: '#/components/schemas/responses.SuccessResponse' - type: object properties: data: $ref: '#/components/schemas/aigateway.ServingLatencyMetrics' '400': description: Bad Request content: application/json: schema: $ref: '#/components/schemas/responses.ErrorResponse' '401': description: Unauthorized content: application/json: schema: $ref: '#/components/schemas/responses.ErrorResponse' '403': description: Forbidden content: application/json: schema: $ref: '#/components/schemas/responses.ErrorResponse' '404': description: Not Found content: application/json: schema: $ref: '#/components/schemas/responses.ErrorResponse' '500': description: Internal Server Error content: application/json: schema: $ref: '#/components/schemas/responses.ErrorResponse' security: - BearerAuth: [] components: schemas: responses.SuccessResponse: type: object required: - message - status properties: data: {} message: type: string example: some message status: type: string example: success aigateway.ServingLatencyMetrics: type: object properties: end: type: string gateway_connection: $ref: '#/components/schemas/aigateway.ServingGatewayConnection' gpu_attribution: $ref: '#/components/schemas/aigateway.ServingGPUAttribution' missing_signal_notes: type: object additionalProperties: type: string missing_signals: type: array items: type: string rate_window_seconds: type: integer series: type: array items: $ref: '#/components/schemas/aigateway.ServingLatencySeries' start: type: string step_seconds: type: integer responses.ErrorResponse: type: object required: - message - status properties: message: type: string example: some message status: type: string example: error aigateway.ServingGatewayConnection: type: object properties: gateway: type: string label_selector: type: string reason: type: string state: type: string aigateway.ServingGPUAttribution: type: object properties: devices: type: integer pods: type: integer reason: type: string state: type: string aigateway.ServingLatencySeries: type: object properties: kind: type: string model: type: string points: type: array items: $ref: '#/components/schemas/metrics.MetricPoint' quantile: type: number signal: type: string unit: type: string usage_kind: type: string metrics.MetricPoint: type: object required: - t - v properties: t: type: integer v: type: number securitySchemes: BearerAuth: description: Enter "Bearer {token}" type: apiKey name: Authorization in: header ```` This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.