Provider Attempt Accounting & Telemetry
AVAILABLEInspect full-stack execution traces, hop-by-hop latency breakdowns, token costs, and automated failover recoveries for every request dispatched by Naagmani.
- Portal Location:
Organization > Provider Attempts(/attempts) - API Endpoint:
/v1/attempts
What is a Provider Attempt? #
When a client application sends a single logical request to /v1/chat/completions, Naagmani may execute one or more Upstream Provider Attempts to fulfill it (for example, if Candidate #1 experiences a timeout or 429 rate limit and fails over to Candidate #2).
Provider Attempt Accounting records the complete execution lifecycle for every hop:
Rendering diagram...
graph TD
Req["Logical Request: req_01J8F0A7B9C1D2E3F4G5H6..."]
Req --> Hop1["Hop #1: DeepSeek (deepseek-chat)"]
Hop1 -->|"550ms, HTTP 503 Overloaded (Failed)"| Cascade["Failover Cascade Triggered"]
Cascade --> Hop2["Hop #2: Anthropic (claude-3-5-sonnet)"]
Hop2 -->|"1250ms, HTTP 200 OK (Succeeded)"| Success["Final Stream Returned to Client"]
Telemetry Attributes Recorded per Attempt #
For every upstream attempt, Naagmani captures:
| Telemetry Field | Description | Example |
|---|---|---|
attempt_id | Unique identifier for this specific provider dispatch hop. | att_01J9H2X9F8E7... |
request_id | Parent logical request ID grouping multi-hop cascades. | req_01J9H2X9F8E7... |
provider | Target upstream vendor name. | deepseek, anthropic, openai |
model | Specific upstream model executed. | deepseek-chat, gpt-4o |
status_code | Raw HTTP response code from the provider. | 200, 429, 503 |
latency_ms | Total end-to-end duration of this attempt in milliseconds. | 450ms |
time_to_first_token_ms | Initial streaming TTFT latency. | 180ms |
prompt_tokens / completion_tokens | Exact billable token counts. | 142 prompt / 350 completion |
cost_cogs_usd | Exact provider cost of this hop calculated from active price books. | $0.00042 |
error_message | Raw error payload if the attempt failed. | upstream server overloaded |
Inspecting Traces in Developer Portal #
- Navigate to Organization > Provider Attempts (
/attempts). - Filter by:
- Date Range: Last 15 minutes, 24 hours, 7 days.
- Status: Successful only (
200), Failed only (4xx/5xx), or Multi-hop cascades. - Project / Environment: Filter by specific application.
- Click any row to expand the detailed trace drawer: view hop timing waterfall, exact token breakdowns, and raw header metadata.