[PLAT-334] AI reports #8673

pjain1 · 2026-01-20T13:33:32Z

Add AI-powered scheduled insight reports with new format (format: ai_session) on existing report.
Implement AI resolver that creates AI sessions by default with analyst agent for completing user prompt

On reconciliation, an AI resolver is created which generates an AI session with analyst_agent and a slightly modified system prompt for scheduled report, upon competition it include a summary and send out the session id link to the recipient.

Report modes behaviour

Recipient mode - In this mode, AI report is run with each recipient attributes and a separate session is created for them.

Creator mode - In this mode, a single session is created with owners attribute (without any charts as it required mv agg for rendering), this session is a shared session and a magic token is used for viewing this session (this mgc token has no access just used for authentication), for continuing conversation user should be logged in and conversation should be forked.

Format

Reports can now use format: ai_session to generate AI-powered insights instead of traditional query exports. Example YAML:

  type: report
  format: ai_session
  data:
    ai:
      agent: analyst_agent
      prompt: 
      time_range:
        iso_duration: P7D
        time_zone: UTC
      comparison_time_range:
        iso_duration: P7D
        iso_offset: P7D
      context:
        explore: my_dashboard

Checklist:

Covered by tests
Ran it and it works as intended
Reviewed the diff before requesting a review
Checked for unhandled edge cases
Linked the issues it closes
Checked if the docs need to be updated. If so, create a separate Linear DOCS issue
Intend to cherry-pick into the release branch
I'm proud of this work!

fix fix refactor

pjain1 · 2026-01-20T17:56:02Z

runtime/ai/ai.go

 }

-const llmRequestTimeout = 60 * time.Second
+const llmRequestTimeout = 90 * time.Second


After all the tool calls, sometimes it would take more than 1 minute for LLM to synthesize the final output. So had to increase it.

begelundmuller · 2026-01-27T17:43:49Z

proto/rill/runtime/v1/resources.proto

  bool intervals_check_unclosed = 15;
+  // AI report configuration
+  string format = 17; // "query" (default) or "ai_session"
+  AIReportConfig ai_config = 18; // Configuration for AI-powered reports (only used when format = "ai_session")


If you look at alerts, they have generic resolver and resolver_properties fields:

rill/proto/rill/admin/v1/api.proto

Lines 3340 to 3345 in fd7a010

string resolver = 13;

google.protobuf.Struct resolver_properties = 14;

// DEPRECATED: Use resolver and resolver_properties instead.

string query_name = 3;

// DEPRECATED: Use resolver and resolver_properties instead.

string query_args_json = 4;

rill/proto/rill/runtime/v1/resources.proto

Lines 649 to 650 in fd7a010

string resolver = 22;

google.protobuf.Struct resolver_properties = 23;

We also intended to do the same for reports, but we never got around to it. However, with this change, rather than hard-code the AI resolver properties here, can we take a generic approach similar to alerts?

made the change for ai report, need to fix the export request and download api for standard report

begelundmuller · 2026-01-27T17:47:04Z

proto/rill/runtime/v1/resources.proto

+// AITimeRange defines a time range using ISO 8601 duration strings.
+// The time range is resolved at report execution time.
+message AITimeRange {
+  string iso_duration = 1; // ISO 8601 duration (e.g., "P7D" for 7 days, "P1M" for 1 month)
+  string iso_offset = 2; // Optional ISO 8601 offset for comparison ranges (e.g., "P7D" to offset by 7 days)
+  string time_zone = 3; // IANA timezone (e.g., "America/New_York")
+}


This appears not to support Rill time expressions? I believe the UI only uses those for new reports (or otherwise, they will probably migrate to that soon).

Instead of a custom type, can it use our main TimeRange proto type, which we already have utils for converting into a concrete time range? Thinking about this one:

rill/proto/rill/runtime/v1/queries.proto

Lines 610 to 627 in fd7a010

// 2 of the (start, end, iso_duration) should be set

message TimeRange {

// Optional. Defaults to min

google.protobuf.Timestamp start = 1;

// Optional. Defaults to max

google.protobuf.Timestamp end = 2;

// Optional, ie PT1M

string iso_duration = 3;

// Optional, ie PT1M

string iso_offset = 4;

TimeGrain round_to_grain = 5;

// Optional. IANA format, ie Europe/Copenhagen. Defaults to UTC

string time_zone = 6;

// Optional. Rill format time range. Should only be used for alerts and reports.

// For dashboard call ResolveTimeRanges.

string expression = 7;

string time_dimension = 8; // Optional. If not specified, falls back to the primary time dimension in the metrics view spec

}

Aa far as I remember, at the time of coding UI was not using expressions so used this, related to this during tool calls LLM would most of the times use expressions in wrong way like adding other fields along with it, causing validation errors. Was not aware its the way going forward, I can make that change and possibly add some concrete examples with expression so that it uses them correctly

This would always lead to failed tool calls thus wasting tokens

runtime/ai/analyst_agent.go

begelundmuller · 2026-01-27T17:55:46Z

runtime/ai/analyst_agent.go

+	// Add scheduled insight mode context
+	data["is_scheduled_insight"] = args.IsScheduledInsight
+	data["is_scheduled_insight_user_prompt"] = args.IsScheduledInsight && !(strings.EqualFold(strings.TrimSpace(args.Prompt), "Generate the scheduled insight report."))


Doing a strings.EqualFold seems kind of unsafe in case we change that text upstream. Would it be possible to keep the prompt empty (so we can do if eq .prompt ""), and inject the "Generate ..." prompt in userPrompt when prompt is empty and IsScheduledInsight is true?

Nit: Can you move this up to the initial data := statement like the other props?

Adding a custom flag to indicate if user has provided custom prompt. The reason for not injecting userPrompt is because it won't be visible in session messages then, if we want in session msgs then we will have add router and agent call msgs along with their full json args.

It won't matter much anyways as per discussion with Nishant, we would always want a prompt here so I think as user keep on selecting time period, measures and dims in the report modal, UI can dynamically create prompt like it does for Explain feature - prompt can be like this for comparative analysis

You are doing comparative analysis between two time periods <t1, t2> in scheduled insight report mode, your analysis should: 1. Compare current period to the comparison period for <all key> or <m1, m2...> measures 2. Identify which measures changed significantly (>10%) 3. For each significant change, identify the dimensional drivers 4. Highlight any ranking changes in top dimensions 5. Generate 3-5 key insights with supporting charts Focus areas: - **Overall changes**: Which measures changed the most between periods? - **Drivers of change**: Which dimensions contributed most to increases/decreases? - **Ranking shifts**: Did any top dimensions change rank significantly? - **Anomalies**: Any unusual patterns unique to one period?

For single period analysis -

You are doing analysis on time period <t1> in scheduled insight report mode, your analysis should: 1. Show totals for the <most impactful> or <m1, m2...> measures in the period 2. Identify interesting trends within the time range (use time series) 3. Find anomalies - unusual spikes, drops, or outliers 4. Highlight top performers and notable dimension values 5. Generate 3-5 key insights with supporting charts Focus areas: - **Totals**: What are the key numbers for this period? - **Trends**: How did metrics change over the period? Any acceleration/deceleration? - **Anomalies**: Are there any unusual data points that stand out? - **Distribution**: Which dimensions dominate? Any concentration issues?

begelundmuller · 2026-01-27T18:33:29Z

runtime/reconcilers/report.go

 		case "email":
 			recipients := pbutil.ToSliceString(notifier.Properties.AsMap()["recipients"])
-			for _, recipient := range recipients {
-				opts := &email.ScheduledReport{
-					ToEmail:        recipient,
-					ToName:         "",
-					DisplayName:    rep.Spec.DisplayName,
-					ReportTime:     t,
-					DownloadFormat: formatExportFormat(rep.Spec.ExportFormat),
-				}
-				urls, ok := meta.RecipientURLs[recipient]
-				if !ok {
-					return false, fmt.Errorf("failed to get recipient URLs for %q", recipient)
-				}
-				opts.OpenLink = urls.OpenURL
-				u, err := createExportURL(urls.ExportURL, t)
-				if err != nil {
-					return false, err
+			if rep.Spec.Format == "ai_session" {
+				reports := make(map[string]*aiReport)
+				for _, recipient := range recipients {
+					urls, ok := meta.RecipientURLs[recipient]
+					if !ok {
+						return false, fmt.Errorf("failed to get recipient URLs for %q", recipient)


This logic is getting pretty nested and hard to follow. Can you think of a way to simplify the flows?

Depending on your thoughts on my earlier comment about having a single SendScheduledReport function, maybe it can build the messages and send the messages in two separate steps, and become simpler that way?

runtime/reconcilers/report.go

begelundmuller · 2026-01-27T18:44:24Z

runtime/reconcilers/report.go

+func buildAISessionURL(baseOpenURL, sessionID string) string {
+	// replace {session_id} in the baseOpenURL with the actual sessionID
+	return strings.ReplaceAll(baseOpenURL, "%7Bsession_id%7D", sessionID)
+}


This feels a little hacky, had to jump back to the URL generation to check things here. Would be nice at least with a comment about this feature in the URLs generation function.

But also, I'm thinking if there's a way we can make this a bit more generic/clean. Another thing we need to support soon is precomputed report exports. Which is kind of similar to a pre-created AI session. I wonder if thinking about that use case as well could give us a cleaner path here.

For example, in alerts, we store the alert history in AlertState. Maybe we should keep similar history for reports? And then keep the existing reports/{name}/open link and let the UI redirect? Not sure.

Added a comment at url generation place. You meant to store resolver and its corresponding meta id like session id in case of ai session in ReportExecution proto?

runtime/resolvers/ai.go

begelundmuller · 2026-01-27T18:49:08Z

runtime/resolvers/ai.go

+// resolveTimeRange resolves the time range from ISO duration to actual timestamps.
+func (r *aiResolver) resolveTimeRange() (start, end time.Time, err error) {


This function seems to duplicate functionality from rilltime. It would be really nice if we could centralize on the rilltime implementation and not have time range resolution logic elsewhere.

See the code here for an example of how rilltime is used:

rill/runtime/metricsview/executor/executor_rewrite_time.go

Line 87 in fd7a010

rillTime, err := rilltime.Parse(tr.Expression, rilltime.ParseOptions{

pjain1 added 3 commits January 20, 2026 00:45

first version

179f9b6

features

fa50fca

fix

89b2fd3

fix fix refactor

pjain1 marked this pull request as draft January 20, 2026 13:33

pjain1 added 3 commits January 20, 2026 21:15

proto for api

2f9847f

mjml template

d2fd159

Merge branch 'main' into ai_reports

cc3b9ca

pjain1 changed the title ~~Ai reports~~ AI reports Jan 20, 2026

pjain1 changed the title ~~AI reports~~ [PLAT-334] AI reports Jan 20, 2026

tweaks

44cc84b

pjain1 commented Jan 20, 2026

View reviewed changes

pjain1 added 4 commits January 21, 2026 13:14

tweaks

a0700bd

tweak

51a98cc

refactor

01bf8f4

recipient mode support for ai reports

a7b4198

pjain1 requested a review from begelundmuller January 26, 2026 05:13

pjain1 marked this pull request as ready for review January 26, 2026 05:13

pjain1 added 5 commits January 26, 2026 10:47

comment

321f32f

lint

8e25370

fix

986722a

creator mode support

0204856

Merge branch 'main' into ai_reports

ebfe286

begelundmuller requested changes Jan 27, 2026

View reviewed changes

pjain1 added 4 commits January 29, 2026 22:36

review comments

0f6e8d1

Merge branch 'main' into ai_reports

cebf8e3

fix test

9312804

email template

dfb6720

pjain1 requested a review from begelundmuller January 29, 2026 18:45

	string resolver = 13;
	google.protobuf.Struct resolver_properties = 14;
	// DEPRECATED: Use resolver and resolver_properties instead.
	string query_name = 3;
	// DEPRECATED: Use resolver and resolver_properties instead.
	string query_args_json = 4;

	string resolver = 22;
	google.protobuf.Struct resolver_properties = 23;

	// 2 of the (start, end, iso_duration) should be set
	message TimeRange {
	// Optional. Defaults to min
	google.protobuf.Timestamp start = 1;
	// Optional. Defaults to max
	google.protobuf.Timestamp end = 2;
	// Optional, ie PT1M
	string iso_duration = 3;
	// Optional, ie PT1M
	string iso_offset = 4;
	TimeGrain round_to_grain = 5;
	// Optional. IANA format, ie Europe/Copenhagen. Defaults to UTC
	string time_zone = 6;
	// Optional. Rill format time range. Should only be used for alerts and reports.
	// For dashboard call ResolveTimeRanges.
	string expression = 7;
	string time_dimension = 8; // Optional. If not specified, falls back to the primary time dimension in the metrics view spec
	}

		// resolveTimeRange resolves the time range from ISO duration to actual timestamps.
		func (r *aiResolver) resolveTimeRange() (start, end time.Time, err error) {

[PLAT-334] AI reports #8673

Are you sure you want to change the base?

[PLAT-334] AI reports #8673

Uh oh!

Conversation

pjain1 commented Jan 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Report modes behaviour

Format

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

pjain1 Jan 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

pjain1 Jan 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

pjain1 commented Jan 20, 2026 •

edited

Loading

pjain1 Jan 29, 2026 •

edited

Loading

pjain1 Jan 29, 2026 •

edited

Loading