KQL functions¶

Stored functions defined in fabric/kql/02_policies.kql. The app-facing functions live in the FDA folder; the validation function in FDA/internal.

Function	Folder	Consumer
`SearchFdaInteractions`	FDA	Review app (results grid)
`GetFdaInteraction`	FDA	Review app (detail)
`FdaTuningSummary`	FDA	Dashboards / aggregate pages
`FdaFailures`	FDA	Review backlog / ops
`FdaTopQuestions`	FDA	Dashboard / tuning
`BuildFdaInteractionsKql`	FDA/internal	Validation / fallback

SearchFdaInteractions¶

Filtered search/list of curated interactions. Empty string / null = no filter.

SearchFdaInteractions(
    term:string = "",
    fromTime:datetime = datetime(null),
    toTime:datetime = datetime(null),
    agent:string = "",
    user:string = "",
    status:string = "",
    onlyMatched:bool = false,
    maxRows:long = 1000)

Parameter	Default	Effect
`term`	`""`	Free-text — matches `Question`, `Answer`, `GeneratedDax`, or `ExecutedDax` (`has`)
`fromTime` / `toTime`	last 30 days → now	Time window (null → `ago(30d)` … `now()`)
`agent`	`""`	Matches `AgentName` or `AgentId` exactly
`user`	`""`	`User has user`
`status`	`""`	Exact `Status` match (`Ok`/`Error`/`Unknown`)
`onlyMatched`	`false`	When true, drops `MatchConfidence == "Unmatched"` rows
`maxRows`	`1000`	Row cap (`take`)

Returns the review columns (timestamp, user, host, agent, model, question, answer, both DAX fields, DaxQueries, perf, status, error, match confidence, sources, thread, id), newest first.

GetFdaInteraction¶

Single interaction by id, for the detail pane.

GetFdaInteraction(id:string)

Returns the one FdaInteractions row where InteractionId == id.

FdaTuningSummary¶

Tuning KPIs over a window, grouped by agent and day — for dashboards.

FdaTuningSummary(fromTime:datetime = datetime(null), toTime:datetime = datetime(null))

Defaults to the last 7 days. Per AgentName × bin(Timestamp, 1d) it returns: Interactions, Users (dcount), Failures, FailureRatePct, UnmatchedDaxPct, P50DurationMs, P95DurationMs.

FdaFailures¶

Interactions needing attention.

FdaFailures(fromTime:datetime = datetime(null), toTime:datetime = datetime(null))

Defaults to the last 7 days. Returns rows where Status == "Error" or MatchConfidence == "Unmatched" or ExecutedDax is empty — projecting timestamp, user, agent, question, status, error, match confidence, and executed DAX, newest first.

FdaTopQuestions¶

Most frequent question patterns — tuning candidate list.

FdaTopQuestions(fromTime:datetime = datetime(null), toTime:datetime = datetime(null), topN:long = 50)

Defaults to the last 30 days. Normalizes questions (trim + tolower), groups by the normalized text, and returns the top topN by count with: Example (a representative original), Count, Users, FailureRatePct, AvgDurationMs. Feed the high-volume patterns back into the FDA as example queries.

BuildFdaInteractionsKql¶

KQL-native reproduction of the collector's correlation, for validation / spot-checks. The collector notebook is authoritative; this function lets you cross-check without re-running it.

BuildFdaInteractionsKql(lookback:timespan = 2d, windowSec:long = 90)

It pairs Graph prompts/responses by ConversationId, window-joins Raw_ExecutedDax by User == ExecutingUser within ±windowSec, computes MatchConfidence (Exact / Windowed / Unmatched), aggregates the per-turn DaxQueries array, and projects an FdaInteractions-shaped result. See Correlation model for the algorithm this mirrors.

Note

This validation path produces Status = "Unknown" and Sources = ["graph","monitoring"] — it does not apply the notebook's per-query status derivation. Use it to sanity-check the join, not as a replacement for the curated table.