Edit

Get analytics for your project

  • Article

Custom question answering uses Azure diagnostic logging to store the telemetry data and chat logs. Follow the below steps to run sample queries to get analytics on the usage of your custom question answering project.

  1. Enable diagnostics logging for your language resource with custom question answering enabled.

  2. In the previous step, select Trace in addition to Audit, RequestResponse and AllMetrics for logging

    Enable trace logging in custom question answering

Kusto queries

Chat log

// All QnA Traffic
AzureDiagnostics
| where ResourceProvider == "MICROSOFT.COGNITIVESERVICES"
| where OperationName=="CustomQuestionAnswering QueryKnowledgebases" // This OperationName is valid for custom question answering enabled resources
| extend answer_ = tostring(parse_json(properties_s).answer)
| extend question_ = tostring(parse_json(properties_s).question)
| extend score_ = tostring(parse_json(properties_s).score)
| extend kbId_ = tostring(parse_json(properties_s).kbId)
| project question_, answer_, score_, kbId_

Traffic count per project and user in a time period

// Traffic count per KB and user in a time period
let startDate = todatetime('2019-01-01');
let endDate = todatetime('2020-12-31');
AzureDiagnostics
| where ResourceProvider == "MICROSOFT.COGNITIVESERVICES"
| where OperationName=="CustomQuestionAnswering QueryKnowledgebases" // This OperationName is valid for custom question answering enabled resources
| where TimeGenerated <= endDate and TimeGenerated >=startDate
| extend kbId_ = tostring(parse_json(properties_s).kbId)
| extend userId_ = tostring(parse_json(properties_s).userId)
| summarize ChatCount=count() by bin(TimeGenerated, 1d), kbId_, userId_

Latency of GenerateAnswer API

// Latency of GenerateAnswer
AzureDiagnostics
| where ResourceProvider == "MICROSOFT.COGNITIVESERVICES"
| where OperationName=="Generate Answer"
| project TimeGenerated, DurationMs
| render timechart

Average latency of all operations

// Average Latency of all operations
AzureDiagnostics
| where ResourceProvider == "MICROSOFT.COGNITIVESERVICES"
| project DurationMs, OperationName
| summarize count(), avg(DurationMs) by OperationName
| render barchart

Unanswered questions

// All unanswered questions
AzureDiagnostics
| where ResourceProvider == "MICROSOFT.COGNITIVESERVICES"
| where OperationName=="CustomQuestionAnswering QueryKnowledgebases" // This OperationName is valid for custom question answering enabled resources
| extend answer_ = tostring(parse_json(properties_s).answer)
| extend question_ = tostring(parse_json(properties_s).question)
| extend score_ = tostring(parse_json(properties_s).score)
| extend kbId_ = tostring(parse_json(properties_s).kbId)
| where score_ == 0
| project question_, answer_, score_, kbId_

Prebuilt custom question answering inference calls

// Show logs from AzureDiagnostics table 
// Lists the latest logs in AzureDiagnostics table, sorted by time (latest first). 
AzureDiagnostics
| where OperationName == "CustomQuestionAnswering QueryText"
| extend answer_ = tostring(parse_json(properties_s).answer)
| extend question_ = tostring(parse_json(properties_s).question)
| extend score_ = tostring(parse_json(properties_s).score)
| extend requestid = tostring(parse_json(properties_s)["apim-request-id"])
| project TimeGenerated, requestid, question_, answer_, score_

Next steps