> For the full page index, fetch https://echonova.ai/llms.txt # Transcripts and summaries What a conversation records for every turn, when the summary above the transcript appears, and why a summary is sometimes missing. A conversation records two things. The transcript is written as the conversation runs: one message for each side of every turn, with the numbers measured on that turn. The summary is written once, at the end, when the agent wraps the conversation up. Read this page to know what each one contains, what each one is for, and why a conversation can have a complete transcript and no summary. ## What every turn records A turn is one exchange: the customer speaks or types, the agent answers. Each turn adds to the conversation's running totals, and those totals are what the conversation page and [Agent analytics](/docs/agents/analytics) show. | Recorded on the conversation | From the turn | | ----------------------------------- | ------------------------------------------------------------------------------------------------------------------------- | | Turns | One more turn | | Prompt tokens and completion tokens | What the model read and wrote on this turn, or zero when the turn measured none | | Text to speech characters | Characters the agent spoke | | Speech to text seconds | Seconds of caller audio transcribed | | Credits charged | The price of the turn, worked out from that usage. [What a turn costs](/docs/agents/what-a-turn-costs) has the arithmetic | | Model share | The model's part of that price, a whole number of credits, added up across turns | | Tool calls and tools used | How many tool calls the turn made and which tools they were | | Model | The model that served the turn, recorded once per conversation and again only if it changes | | Customer name | Filled from the turn only when the conversation has none. A name from the channel's own profile always wins | Credits are charged against the workspace balance after every turn. When the balance runs out mid-conversation, the conversation ends and its status reads `out_of_credits`. The turns already charged stay charged. A turn counts and is charged whether or not it carried words. ## What a transcript message carries Each turn writes up to two messages: one for the caller's text and one for the agent's text, each only when that side said something. Transcript messages are written for every conversation, whichever channel carried it, including sessions started from the playground. | Field | Values | | --------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Direction | `inbound` from the caller, `outbound` from the agent | | Sender | `customer` or `bot` | | Type | `text`, `card`, `quick_reply`, `menu`, `link`, `warning` or `error`. Rich types keep their structured content beside the text | | Text | What was said. The text always carries a readable fallback, even for a rich type | | Delivery status | `received`, `sent`, `delivered`, `read` or `failed` | | Measurements | Per-turn numbers, present only when the turn measured them: the model's own time, per-stage latency, audio durations, a label for the turn, the credits the turn cost and the model's share of them | Two of those measurements are visible on the conversation page. The model's own time appears beside an agent message as `LLMS N Sec`, and the model's credit share is what the **LLM cost** chart plots. A conversation whose turns measured no credit share shows that chart's empty state rather than a flat zero line. Every message also moves the conversation's own counters: messages in, messages out, failed sends, and the average response time. The last message and its time are what the conversation list shows under the customer's name. ## When the summary appears The summary is the agent's own wrap-up of the conversation. The agent writes it once, after the last turn, and it appears under **Summary** in the conversation inspector. It is written for a person skimming a list of calls: two to four sentences covering what the caller wanted, what happened and what is still open, with no instructions, no tool payloads and no secrets in it. A conversation carries at most one summary, of at most 4000 characters, with leading and trailing space trimmed. The summary changes nothing else: the status, the credits, the duration and the transcript are all recorded turn by turn while the conversation runs, and they are complete before the summary exists. ## Why a summary can be missing A missing summary tells you nothing about the conversation. It ran, it was transcribed and it was charged either way. The **Summary** section is empty when: - **The agent wrote no summary.** The agent skips the summary when it has nothing to summarise. - **The summary was blank.** A summary that is empty after trimming is not kept. - **The summary did not reach the conversation.** The transcript and the numbers are recorded turn by turn as the conversation runs; the summary is delivered once at the end, so a conversation can be complete in every other way and still have no summary. - **The conversation ended before 14 September 2026.** Conversations from before that date have no summary, whatever the agent wrote at the time, and they are not backfilled. > [!NOTE] > A conversation whose summary is missing still has everything else. Read the > transcript, the metadata cards and the **Total credits** row on the > conversation for what actually happened. ## Get it right The summary is free text written by a model. Treat it as a reading aid, and take the record from the fields that are measured. **Recommended** - Read the status chip for how the conversation closed, **Total credits** for what it cost, and the transcript for what was said. - Use the summary to decide which conversations are worth opening. **Less effective** - Treating the summary as the resolution record, or counting conversations by what their summaries say. The summary is not a status, not a resolution, and not an input to any number on [Agent analytics](/docs/agents/analytics). It is not sent to your outbound webhooks. ## FAQ ### Is call audio kept? No. A voice conversation records how many seconds of audio were transcribed and the words that came out of it. There is no recording. See [Data retention](/docs/agents/data-retention). ### Can I edit or regenerate a summary? No. The summary is written once by the agent at the end of the conversation. Tags are the label you control on a conversation; see [Tag a conversation](/docs/agents/conversation-tags). ### Why does an agent message show no LLMS time? That turn measured no model time. The mark appears only on agent messages that carried the measurement. ### Does a playground session get a transcript and a summary? Yes. A session started from **Test** is recorded like any other conversation. See [Test your agent](/docs/agents/test-your-agent). ### Can I search transcript text? The conversation list searches the customer name and the contact id. Open a conversation to read its transcript. See [Find and read a conversation](/docs/agents/conversations). ## Related - [Find and read a conversation](/docs/agents/conversations): where the transcript and the inspector are - [Tag a conversation](/docs/agents/conversation-tags): the label you control on a conversation - [Agent analytics](/docs/agents/analytics): how the per-turn numbers become metrics - [What a turn costs](/docs/agents/what-a-turn-costs): how the measured usage becomes credits - [What is an agent](/docs/agents/what-is-an-agent): what a conversation is - [Tools overview](/docs/agents/tools-overview): what the tools a turn used refer to