CSV/OpenTelemetry importer — see your spend without a code change
Upload a CSV or OpenTelemetry span export of your existing LLM usage and see it in the real Cognocient dashboards, with no code changes and no production traffic touched.
Upload a CSV or an OpenTelemetry span export of usage you already have,
and see it populate the real Cognocient dashboards immediately — no
base_url change, no code change, and nothing in your production
traffic is touched.
This is historical data, not live monitoring
Every dashboard page shows a persistent, non-dismissible banner when an account has imported data — it's never presented as if it were real-time. See What each path actually gives you for how this compares to the proxy and the Python wrapper.
CSV format
There's no fixed pre-defined format to match exactly. After you upload a file, a preview step shows your actual column headers and lets you map each one to a Cognocient field — the same flexible-mapping design that makes this usable by a non-technical evaluator, not just an engineer.
| Field | Required | Notes |
|---|---|---|
timestamp | Yes | ISO 8601, YYYY-MM-DD [HH:MM:SS], or Unix epoch (seconds or ms) |
model | Yes | e.g. gpt-4o, claude-sonnet-4-6 |
provider | No | Inferred from the model name if omitted |
cost_usd | No | Calculated from prompt_tokens + completion_tokens + model if omitted, using the same pricing table the live proxy uses |
prompt_tokens / completion_tokens | No | Required only if cost_usd isn't provided |
tag_feature / tag_department / tag_user / tag_session | No | Same attribution model as the proxy's X-Cost-* headers |
status_code | No | Defaults to 200 |
OpenTelemetry span export
For prospects already using OTel instrumentation, upload a span export
JSON (resourceSpans[].scopeSpans[].spans[]). Span attributes are
mapped automatically using GenAI semantic-convention names —
gen_ai.request.model, gen_ai.usage.input_tokens,
gen_ai.usage.output_tokens, gen_ai.system, and a few common aliases
— with no manual mapping step, since OTel attribute names already
follow a standard convention.
What works, and what doesn't
Fully functional against imported data: Feature Intelligence, Token Maxing, Context Tax, Waste Detection, and Forecast. Retry detection and context-bloat detection are replayed against the imported batch so these dashboards show real, non-zero results, not just empty defaults.
Not available for imported data, ever: pre-call budget enforcement (block/degrade/alert-before-it-fires) and the Failure Loop Breaker's active state. There's no live traffic for either to act on — enforcement pages show an explicit "not currently enforcing" notice for imported-only accounts rather than a misleading blank or fake-active state.
Frequently asked questions
Related articles