Phase 7: AI-assisted query authoring (autocomplete, explain, fix, optimize, NL translation)

Adds a self-hosted (Ollama, qwen2.5-coder) model provider abstraction
with a pluggable opt-in cloud adapter, schema grounding, and a shared
cost/safety guard every AI-suggested query is assessed against --
compiling to and executing through the same unchanged Phase 2 IR/
compiler and Phase 4 tenant scoping as a hand-written query, no
parallel execution path.

Track A (built into the query bar): inline ghost-text autocomplete,
"Explain this query", "Fix this query" with a diff view, and a
rule-based "Optimize" suggestion. Track B: natural-language-to-query
translation, always a separate review step from execution, with
`sentryctl query --nl` requiring explicit confirmation to run.
Every accepted/dismissed translate-fix-optimize interaction is logged
into the same append-only audit_log table Phase 4 built.

Two real product bugs were found and fixed via live browser
verification (a Svelte effect re-running on every keystroke that
silently cancelled the ghost-text debounce; a ghost-text widget
positioned at document offset 0 instead of the cursor), and a real
costguard logic bug (unbounded-aggregation vs. raw-row) was caught by
its own test suite. New integration tests wire a real Ollama client
through the real HTTP handler against a mock server matching Ollama's
wire contract (hack/mock-ollama), keeping model-quality verification
out of CI as a disclosed, periodic human-run check instead.

See /docs/phase-7-ai-design.md and /docs/phase-7-runbook.md.
This commit is contained in:
2026-08-16 18:06:27 -07:00
parent 661568085e
commit 7d316f92db
37 changed files with 5230 additions and 20 deletions
+141
View File
@@ -0,0 +1,141 @@
// Package costguard is the shared cost/safety check task 4 asked for --
// no such mechanism existed anywhere in Phase 2/3's compiler before this
// (confirmed by reading planner.go/sql.go before writing this: plan.TimeRange
// can be entirely unset, and nothing downstream rejects that). Built here
// as a standalone, pure function operating on the same ir.Plan every
// query -- hand-written or AI-generated -- already compiles to, so there
// is exactly one cost check, not one per code path.
//
// This package does not decide what a caller *does* with a Reject-level
// Assessment -- see /docs/phase-7-ai-design.md's "Cost/safety guard"
// section for how the AI tracks and the existing /query handler each
// apply this differently (AI suggestions withhold a Reject-level
// suggestion from being offered as directly runnable; the existing
// /query handler surfaces the same assessment as a non-blocking warning,
// deliberately not a new hard block on hand-written queries this phase
// didn't set out to change).
package costguard
import (
"regexp"
"strings"
"time"
"github.com/sentry/sentry/api/internal/querylang/ir"
)
type Level string
const (
LevelOK Level = "ok"
LevelWarn Level = "warn"
LevelReject Level = "reject"
)
type Assessment struct {
Level Level
Reasons []string
}
// maxReasonableSpan and the two below are first-pass heuristic
// thresholds, not benchmarked against a production-scale ClickHouse
// cluster -- this environment's own data is far smaller than what these
// numbers are meant to guard against. Flagged explicitly in
// /docs/phase-7-ai-design.md rather than presented as tuned. Revisit
// once there's real cluster-size data to check them against.
const maxReasonableSpan = 90 * 24 * time.Hour
// rawSQLTimestampRe is a best-effort, deliberately loose check for
// *some* mention of the timestamp column in a raw SQL statement's WHERE
// clause -- not a real SQL parser. A false negative here (a query that
// does filter by time in a way this regex doesn't recognize) just means
// an unnecessary Warn, not a Reject, so being loose-but-safe is the
// right failure direction. Raw SQL genuinely can't get the same
// structural guarantee the IR-based checks below get, and this package
// says so rather than pretending otherwise.
var rawSQLTimestampRe = regexp.MustCompile(`(?i)\btimestamp\b\s*[<>=]`)
// Assess evaluates one compiled plan. Never returns an error -- a plan
// that reached this point already parsed successfully; this is a
// judgment call about cost, not a correctness check.
func Assess(plan *ir.Plan) Assessment {
if plan.RawSQL != "" {
return assessRawSQL(plan.RawSQL)
}
return assessIR(plan)
}
func assessRawSQL(sql string) Assessment {
if rawSQLTimestampRe.MatchString(sql) {
return Assessment{Level: LevelOK}
}
return Assessment{
Level: LevelWarn,
Reasons: []string{
"no obvious timestamp filter found in this raw SQL -- this is a best-effort text check, not a real parse, so it may be wrong in either direction, but if this query has no time bound it could scan the full table's history",
},
}
}
func assessIR(plan *ir.Plan) Assessment {
var reasons []string
level := LevelOK
hasTimeBound := plan.TimeRange != nil && (!plan.TimeRange.From.IsZero() || !plan.TimeRange.To.IsZero())
if !hasTimeBound {
switch {
case plan.Aggregation != nil:
// Unlike a raw-row query, an aggregation gets no implicit
// row cap from the executor regardless of plan.Limit --
// see executor/sql.go's buildSQL: the defaultRowLimit
// safety net only applies `else if plan.Aggregation ==
// nil`. An unbounded aggregation is never merely
// "capped but slow" the way a raw-row fetch is.
level = LevelReject
reasons = append(reasons, "no time range filter, and this query aggregates -- every matching row across the table's entire history must be scanned to compute the aggregate, regardless of how small the output is")
default:
// A raw-row query with no explicit Limit still gets
// executor/sql.go's defaultRowLimit=100 safety net applied
// automatically -- it is not actually unbounded output,
// just potentially an expensive scan to find those rows
// without a time bound to narrow the search. Confirmed by
// reading buildSQL directly, not assumed: this is the same
// risk level whether plan.Limit is nil or explicitly set,
// so both cases share one Warn, not a Reject for one and a
// Warn for the other.
level = LevelWarn
reasons = append(reasons, "no time range filter -- results are capped (explicitly, or by the default 100-row limit), but ClickHouse may still need to scan well beyond that many rows to find them without a time bound to narrow the search")
}
if len(plan.TextSearch) > 0 {
reasons = append(reasons, "the free-text search stage is bounded by the existing 5,000-record Tantivy prefilter cap regardless of time range, which partially limits how bad this is, but doesn't remove the underlying ClickHouse-side cost")
}
} else if !plan.TimeRange.From.IsZero() && !plan.TimeRange.To.IsZero() {
span := plan.TimeRange.To.Sub(plan.TimeRange.From)
if span > maxReasonableSpan {
level = maxLevel(level, LevelWarn)
reasons = append(reasons, "time range spans more than 90 days -- this may be slow depending on data volume")
}
}
return Assessment{Level: level, Reasons: reasons}
}
func maxLevel(a, b Level) Level {
rank := map[Level]int{LevelOK: 0, LevelWarn: 1, LevelReject: 2}
if rank[b] > rank[a] {
return b
}
return a
}
// Summary renders an Assessment as one human-readable line, for
// embedding in an AI-suggestion response or a /query warnings entry --
// one shared rendering so the two callers don't independently invent
// slightly different phrasing for the same underlying reasons.
func Summary(a Assessment) string {
if a.Level == LevelOK || len(a.Reasons) == 0 {
return ""
}
return strings.Join(a.Reasons, "; ")
}
+100
View File
@@ -0,0 +1,100 @@
package costguard
import (
"testing"
"time"
"github.com/sentry/sentry/api/internal/querylang/ir"
)
// A raw-row (non-aggregation) query with no time range and no explicit
// Limit still gets executor/sql.go's defaultRowLimit=100 safety net
// applied automatically -- so this is a Warn (a possibly-expensive scan
// to find those 100 rows), not a Reject (genuinely unbounded output),
// which only an unbounded *aggregation* actually is. See the case
// immediately below for that contrast.
func TestAssessNoTimeRangeNoLimitRawRowWarns(t *testing.T) {
plan := &ir.Plan{Filters: []ir.FilterPredicate{{Field: "service", Op: "=", Value: "api"}}}
got := Assess(plan)
if got.Level != LevelWarn {
t.Errorf("Level = %v, want warn (executor applies a default row limit even with no explicit Limit)", got.Level)
}
if len(got.Reasons) == 0 {
t.Error("expected at least one reason")
}
}
func TestAssessNoTimeRangeWithAggregationRejects(t *testing.T) {
plan := &ir.Plan{
TimeRange: &ir.TimeRange{},
Aggregation: &ir.Aggregation{Funcs: []ir.AggFunc{{Func: "count", Alias: "count"}}},
}
got := Assess(plan)
if got.Level != LevelReject {
t.Errorf("Level = %v, want reject for an unbounded aggregation", got.Level)
}
}
func TestAssessNoTimeRangeWithLimitWarns(t *testing.T) {
plan := &ir.Plan{
TimeRange: &ir.TimeRange{},
Limit: &ir.Limit{N: 100},
}
got := Assess(plan)
if got.Level != LevelWarn {
t.Errorf("Level = %v, want warn (limited, no aggregation)", got.Level)
}
}
func TestAssessBoundedTimeRangeIsOK(t *testing.T) {
now := time.Now()
plan := &ir.Plan{
TimeRange: &ir.TimeRange{From: now.Add(-1 * time.Hour), To: now},
Limit: &ir.Limit{N: 100},
}
got := Assess(plan)
if got.Level != LevelOK {
t.Errorf("Level = %v, want ok, reasons: %v", got.Level, got.Reasons)
}
}
func TestAssessVeryLargeTimeRangeWarns(t *testing.T) {
now := time.Now()
plan := &ir.Plan{
TimeRange: &ir.TimeRange{From: now.Add(-200 * 24 * time.Hour), To: now},
Limit: &ir.Limit{N: 100},
}
got := Assess(plan)
if got.Level != LevelWarn {
t.Errorf("Level = %v, want warn for a 200-day range", got.Level)
}
}
func TestAssessRawSQLWithTimestampFilterIsOK(t *testing.T) {
plan := &ir.Plan{RawSQL: "SELECT count(*) FROM logs WHERE timestamp > now() - INTERVAL 1 HOUR"}
got := Assess(plan)
if got.Level != LevelOK {
t.Errorf("Level = %v, want ok, reasons: %v", got.Level, got.Reasons)
}
}
func TestAssessRawSQLWithoutTimestampFilterWarns(t *testing.T) {
plan := &ir.Plan{RawSQL: "SELECT service, count(*) FROM logs GROUP BY service"}
got := Assess(plan)
if got.Level != LevelWarn {
t.Errorf("Level = %v, want warn for raw SQL with no detectable time filter", got.Level)
}
}
func TestSummaryEmptyForOK(t *testing.T) {
if s := Summary(Assessment{Level: LevelOK}); s != "" {
t.Errorf("Summary(OK) = %q, want empty", s)
}
}
func TestSummaryJoinsReasons(t *testing.T) {
a := Assessment{Level: LevelWarn, Reasons: []string{"a", "b"}}
if s := Summary(a); s != "a; b" {
t.Errorf("Summary = %q, want %q", s, "a; b")
}
}