Observability

GET /v1/observability returns one snapshot of a date range: how many sessions ran, how they ended, how fast the agent answered and what it cost. It is what the dashboard’s Observability page charts. For one call at full resolution, use calls.

curl -G https://api.talqing.com/v1/observability \
  -H "Authorization: Bearer $TALQING_API_KEY" \
  --data-urlencode "start_date=2026-08-24" \
  --data-urlencode "end_date=2026-08-30" \
  --data-urlencode "timezone=Asia/Kolkata"

const report = await talqing.observability.get({
  start_date: "2026-08-24",
  end_date: "2026-08-30",
  timezone: "Asia/Kolkata",
});
console.log(report.summary.total_sessions, report.summary.total_cost);

report = talqing.observability.get(
    start_date="2026-08-24", end_date="2026-08-30", timezone="Asia/Kolkata"
)
print(report["summary"]["total_sessions"], report["summary"]["total_cost"])

Parameters

Parameter	Default	Meaning
`start_date`	`end_date` minus 6 days	First day, `YYYY-MM-DD`.
`end_date`	Today in `timezone`	Last day, inclusive.
`timezone`	`UTC`	IANA zone; sets day boundaries and every `bucket` timestamp. Unknown zone is a `400`.
`agent_id`	—	Only this agent’s sessions.
`channel`	—	`voice`, `video` or `text`.

A range over seven days is a 400 (date range cannot exceed 7 days); so is start_date after end_date. Query each week separately to compare weeks. There is no granularity parameter: a single-day range returns 24 hourly buckets (future hours as zeroes); anything wider returns daily buckets. range.granularity says which. Every series is zero-filled across every bucket.

Response

Section	Holds
`range`	The resolved `start_date`, `end_date`, `timezone`, `granularity`.
`summary`	Totals: session counts by status, `completion_rate`, runtime, `total_cost`/`provider_cost`/`platform_fee`, `priced_sessions`/`pending_sessions`/`unpriceable_sessions`, `average_e2e_latency_ms`, and `by_channel` (always all three channels, zero when idle).
`timeline`	One row per bucket: sessions, completed, incomplete, the three cost figures, talk time, recorded, analysed and transferred counts, `cache_hit_ratio`.
`volume`	Eight breakdowns over the same buckets (below).
`endings`	Each distinct `close_reason` with its `bucket`, `sessions` and `failed_sessions`.
`latency`	`basis`, `channels`, `measured_sessions`, per-stage `metrics` with per-bucket points, `slow_sessions` over `slow_response_threshold_ms`, and a `distribution` across sessions.
`cost`	Spend by kind (`llm_cost`, `stt_cost`, `tts_cost`, `realtime_cost`, `avatar_cost`, plus `analysis_cost` and `priority_cost`), per-model `lines`, `by_provider`, LLM cache split, and `unpriceable_reasons`. See costs.
`agents`	Per agent and channel: sessions, `completion_rate`, `success_rate`, average duration, total and average cost, average latency.

Each volume breakdown is {series, other_count}; each series has a stable key, label, total and points.

Breakdown	Cut by
`agent_data`	Agent, by the name recorded on the session
`phone_number_data`	Phone number, labelled with the E.164
`type_data`	`SIP_INBOUND`, `SIP_OUTBOUND`, `WHATSAPP_INBOUND`, `STREAM`, `WEB`, `TEXT`
`status_data`	`completed`, `failed`, `canceled`, `running`, `queued`
`close_reason_data`	Close-reason bucket, not the raw string — see close reasons
`recording_data`	`stored`, `pending`, `failed`, `deleted`, `none`
`analysis_data`	`completed`, `failed`, `skipped`, `pending`, `none`
`transfer_data`	`connected`, `failed`

Each breakdown keeps the eight largest series; the rest are summed into one series keyed __other__, and other_count says how many values it holds.

Reading the numbers correctly

Latency differs by channel. latency.basis is per_turn for voice and video (end-to-end, weighted over measured turns) and per_call for text (only llm_node_ttft, the model’s time to first token). Without a channel filter, latency covers voice and video only; latency.channels says which. Never average the two.
Rates have their own denominators. success_rate is over sessions post-call analysis judged; cost figures over priced_sessions; latency over measured_sessions. Dividing by total_sessions gives a number that means nothing.

pending_sessions are still being priced and resolve themselves. unpriceable_sessions are sessions billing could not price; the reason is in cost.unpriceable_reasons. Agent task runs are not sessions and do not appear here; their cost is on the task’s runs. There is no per-turn, tool-call or webhook-delivery metric here; open the call instead.

In the dashboard

Observability has Today, Yesterday and 7-day presets, a date picker that cannot exceed seven days, and agent and channel filters.

Calls

One call: transcript, tool calls, latency, cost.

Debugging agents

Symptom-first playbook.

Get started

Build

Guides

Parameters

Response

Reading the numbers correctly

In the dashboard

Calls

Debugging agents

​Parameters

​Response

​Reading the numbers correctly

​In the dashboard

Calls

Debugging agents

Parameters

Response

Reading the numbers correctly

In the dashboard