Skip to main content
Handling a single conversation takes many calls to the language model, each with its own prompt. Every one of them is logged here, with the content sections retrieved and the action that followed. This is where you look to explain a reply.

One row is

one model call

Delivery

Daily delta on etl_updated_at

Columns

33

​
Where it lands

Each partition holds only the rows the datamart ingested that day, so union every etl_date= partition to reconstruct the table. One of them is etl_date=__initial__, the first load, holding everything that preceded the earliest daily partition. Check for _SUCCESS before reading a partition, as described in bucket layout.

​
Columns

​
bigint
Internal warehouse key. Not stable between deliveries — join on prompt_revision_id instead.
​
varchar(510)
Join key. Unique id of the decision.
​
varchar(510)
Join key. Conversation the decision belongs to. Matches cloudchat_id on claudia_conversation.
​
varchar(50)
Project the conversation belongs to.
​
varchar(510)
Environment the decision was made in.
​
varchar(50)
What the agent decided to do: NO_ACTION to keep talking, FORWARD_N2 to escalate, RESOLVE_TICKET or CLOSE_TICKET to end the conversation.
​
boolean
Whether the agent had exhausted its clarifying questions at this point.
​
varchar(65535)
The customer question this decision responds to, as interpreted.
​
varchar(255)
Which response strategy the agent chose. The value is a fully-qualified internal class name and its exact string is not part of the delivery contract: read the trailing segment, for example AcceptableConfidenceResponse or NoValidContentResponse.
​
timestamp
When the decision was made.
​
timestamp
When the row was last modified.
​
varchar(65535)
The answer the agent produced, before any formatting applied on the way out to the channel.
​
varchar(510)
Controlled flow active on the conversation when the decision was made.
​
varchar(510)
Run of that flow.
​
varchar(510)
Controlled flow this decision started or advanced.
​
varchar(510)
Run of the flow in response_bot_flow_id.
​
varchar(510)
Result of the flow run. Joins to result_id on eddie_result_outcomes.
​
varchar(510)
The content section the answer was built from. Joins to content_id on claudia_contents. This is the column that lets you attribute an answer to a piece of your base.
​
varchar(65535)
Written reasoning recorded for the decision.
​
varchar(510)
First of the candidate sections retrieved for this question, in relevance order. The candidates are what was available to answer with, not necessarily what was used.
​
varchar(510)
Second candidate section retrieved.
​
varchar(510)
Third candidate section retrieved.
​
smallint
Frustration estimated at this point in the conversation, from 0 to 100.
​
varchar(255)
frustration_score bucketed into a label.
​
varchar(510)
Written justification for the frustration classification.
​
varchar(65535)
Error recorded while producing the answer. NULL on decisions that succeeded, which is the vast majority.
​
varchar(510)
First candidate section considered when the agent asked a clarifying question instead of answering.
​
varchar(510)
Second candidate section considered for the clarifying question.
​
varchar(510)
Third candidate section considered for the clarifying question.
​
timestamp
When the datamart last wrote this row. This is the cursor for the daily delta: a partition holds the rows whose etl_updated_at fell on that day.
​
varchar(256)
Id of the tenant. NULL for rows with no tenant mapping.
​
varchar(256)
Name of the tenant. NULL for rows with no tenant mapping.
​
varchar
Partition column. The day whose rows this file holds, in YYYY-MM-DD. Not a date type, because the initial load carries the literal __initial__ instead of a day.