bitfab 0.65.11 → 0.66.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/{http-DdlELWcL.js → http-D91qP2J-.js} +2 -2
- package/dist/http-D91qP2J-.js.map +1 -0
- package/dist/{http-CFwCJQyF.cjs → http-DIz8mEMq.cjs} +2 -2
- package/dist/http-DIz8mEMq.cjs.map +1 -0
- package/dist/{index-CUstKXuW.d.cts → index-DiYFr-KV.d.cts} +224 -61
- package/dist/{index-CUstKXuW.d.ts → index-DiYFr-KV.d.ts} +224 -61
- package/dist/index.cjs +3 -2
- package/dist/index.d.cts +2 -2
- package/dist/index.d.ts +2 -2
- package/dist/index.js +3 -3
- package/dist/node.cjs +3 -2
- package/dist/node.d.cts +2 -2
- package/dist/node.d.ts +2 -2
- package/dist/node.js +3 -3
- package/dist/{registryCommand-Cr7m6Vif.js → registryCommand-Dk4Y1ZgW.js} +2 -2
- package/dist/registryCommand-Dk4Y1ZgW.js.map +1 -0
- package/dist/replayCli.js +1 -1
- package/dist/seedCli.js +1 -1
- package/dist/{src-lTOsAt47.js → src-CL5eQlNk.js} +114 -46
- package/dist/src-CL5eQlNk.js.map +1 -0
- package/dist/{src-mV-7DNLd.cjs → src-rJcHXJEz.cjs} +119 -45
- package/dist/src-rJcHXJEz.cjs.map +1 -0
- package/package.json +1 -1
- package/dist/http-CFwCJQyF.cjs.map +0 -1
- package/dist/http-DdlELWcL.js.map +0 -1
- package/dist/registryCommand-Cr7m6Vif.js.map +0 -1
- package/dist/src-lTOsAt47.js.map +0 -1
- package/dist/src-mV-7DNLd.cjs.map +0 -1
|
@@ -1536,6 +1536,9 @@ type ExperimentTally = {
|
|
|
1536
1536
|
stillPassing: number;
|
|
1537
1537
|
regressed: number;
|
|
1538
1538
|
stillFailing: number;
|
|
1539
|
+
/** Pairs whose original trace has no verdict, so the replay cannot be compared. */
|
|
1540
|
+
originalWithoutVerdict: number;
|
|
1541
|
+
/** @deprecated Use {@link ExperimentTally.originalWithoutVerdict}. */
|
|
1539
1542
|
originalUnlabeled: number;
|
|
1540
1543
|
unpaired: number;
|
|
1541
1544
|
classified: number;
|
|
@@ -1550,6 +1553,9 @@ type ExperimentTotals = {
|
|
|
1550
1553
|
succeeded: number;
|
|
1551
1554
|
failed: number;
|
|
1552
1555
|
pending: number;
|
|
1556
|
+
/** Replays that finished but have no verdict yet. */
|
|
1557
|
+
awaitingVerdicts: number;
|
|
1558
|
+
/** @deprecated Use {@link ExperimentTotals.awaitingVerdicts}. */
|
|
1553
1559
|
awaitingLabels: number;
|
|
1554
1560
|
errored: number;
|
|
1555
1561
|
withErrors: number;
|
|
@@ -1559,6 +1565,8 @@ type ExperimentTotals = {
|
|
|
1559
1565
|
};
|
|
1560
1566
|
type ExperimentCategoryTally = {
|
|
1561
1567
|
category: CategorySummary;
|
|
1568
|
+
verdicts: ExperimentTally;
|
|
1569
|
+
/** @deprecated Use {@link ExperimentCategoryTally.verdicts}. */
|
|
1562
1570
|
labels: ExperimentTally;
|
|
1563
1571
|
};
|
|
1564
1572
|
type ExperimentJitter = {
|
|
@@ -1569,6 +1577,9 @@ type ExperimentJitter = {
|
|
|
1569
1577
|
type ExperimentRollup = {
|
|
1570
1578
|
traces: ExperimentTally;
|
|
1571
1579
|
skippedTraces: number;
|
|
1580
|
+
/** Scored assertion and grader verdicts. */
|
|
1581
|
+
verdicts: ExperimentTally;
|
|
1582
|
+
/** @deprecated Use {@link ExperimentRollup.verdicts}. */
|
|
1572
1583
|
labels: ExperimentTally;
|
|
1573
1584
|
checks: ExperimentTally;
|
|
1574
1585
|
categories: ExperimentCategoryTally[];
|
|
@@ -1662,7 +1673,8 @@ type AssertionEvidenceParameter = {
|
|
|
1662
1673
|
name: string;
|
|
1663
1674
|
type: "string" | "number" | "boolean" | "object" | "array" | "null" | "unknown";
|
|
1664
1675
|
};
|
|
1665
|
-
|
|
1676
|
+
/** One suggested evidence candidate for an assertion verdict. Save only `{ spanId, text }` as evidence. */
|
|
1677
|
+
type AssertionVerdictEvidence = {
|
|
1666
1678
|
spanId: string;
|
|
1667
1679
|
text: string;
|
|
1668
1680
|
sourceField: string;
|
|
@@ -1671,6 +1683,8 @@ type AssertionLabelEvidence = {
|
|
|
1671
1683
|
spanType: "llm" | "agent" | "function" | "guardrail" | "handoff" | "custom";
|
|
1672
1684
|
isMocked: boolean;
|
|
1673
1685
|
};
|
|
1686
|
+
/** @deprecated Use {@link AssertionVerdictEvidence}. Same type. */
|
|
1687
|
+
type AssertionLabelEvidence = AssertionVerdictEvidence;
|
|
1674
1688
|
type SaveAssertion = {
|
|
1675
1689
|
/** Omit to preserve the category, or pass null to clear it. */
|
|
1676
1690
|
category_id?: string | null;
|
|
@@ -1992,14 +2006,20 @@ declare class TracesClient {
|
|
|
1992
2006
|
saveScenario({ traceId, ...edit }: SaveScenarioParams): Promise<Scenario>;
|
|
1993
2007
|
}
|
|
1994
2008
|
//#endregion
|
|
1995
|
-
//#region src/
|
|
1996
|
-
|
|
1997
|
-
type
|
|
1998
|
-
|
|
1999
|
-
type
|
|
2009
|
+
//#region src/verdicts.d.ts
|
|
2010
|
+
/** How sure the judge is of a verdict. */
|
|
2011
|
+
type VerdictConfidence = "VeryLow" | "Low" | "Medium" | "High" | "VeryHigh";
|
|
2012
|
+
/** What a write did. `no-active-verdict` means an archive found nothing to clear. */
|
|
2013
|
+
type VerdictAction = "set" | "archived" | "no-active-verdict" | "skipped" | "abstained";
|
|
2014
|
+
/** Whether a trace or assertion has a verdict: `judged`, `skipped` (withheld on purpose), or `not-judged`. */
|
|
2015
|
+
type VerdictStatus = "judged" | "skipped" | "not-judged";
|
|
2016
|
+
/** Who wrote an assertion verdict. */
|
|
2017
|
+
type VerdictSource = "human" | "agent";
|
|
2000
2018
|
/** An assertion verdict. `abstained` means the judge could not find the evidence the assertion needs; it counts as a fail. */
|
|
2001
2019
|
type Verdict = "pass" | "fail" | "abstained";
|
|
2002
|
-
|
|
2020
|
+
/** A verdict a person or a grader can give. Neither abstains. */
|
|
2021
|
+
type PassFailVerdict = "pass" | "fail";
|
|
2022
|
+
type VerdictAddress = {
|
|
2003
2023
|
traceId: string;
|
|
2004
2024
|
originalTraceId?: never;
|
|
2005
2025
|
attempt?: never;
|
|
@@ -2008,63 +2028,117 @@ type LabelAddress = {
|
|
|
2008
2028
|
attempt?: number;
|
|
2009
2029
|
traceId?: never;
|
|
2010
2030
|
};
|
|
2011
|
-
|
|
2031
|
+
/** The assertion a verdict scores, on an original trace (`traceId`) or a replay (`originalTraceId` plus optional `attempt`). */
|
|
2032
|
+
type VerdictTarget = VerdictAddress & {
|
|
2012
2033
|
assertionId: string;
|
|
2013
2034
|
};
|
|
2014
|
-
|
|
2035
|
+
/** Like {@link VerdictTarget}, but `assertionId` may be omitted to clear an old whole-trace verdict. */
|
|
2036
|
+
type ArchiveVerdictTarget = VerdictAddress & {
|
|
2015
2037
|
assertionId?: string;
|
|
2016
2038
|
};
|
|
2017
|
-
type
|
|
2018
|
-
|
|
2039
|
+
type VerdictJudgment = {
|
|
2040
|
+
verdict: PassFailVerdict;
|
|
2019
2041
|
annotation: string;
|
|
2020
|
-
confidence?:
|
|
2042
|
+
confidence?: VerdictConfidence;
|
|
2021
2043
|
evidence?: Justification | null;
|
|
2044
|
+
label?: never;
|
|
2045
|
+
abstain?: never;
|
|
2046
|
+
skip?: never;
|
|
2047
|
+
archive?: never;
|
|
2022
2048
|
} | {
|
|
2023
|
-
|
|
2049
|
+
verdict: "abstained";
|
|
2024
2050
|
annotation?: string;
|
|
2051
|
+
label?: never;
|
|
2052
|
+
abstain?: never;
|
|
2053
|
+
skip?: never;
|
|
2054
|
+
archive?: never;
|
|
2025
2055
|
} | {
|
|
2056
|
+
/** @deprecated Use `verdict: "pass" | "fail"`. Sent as `verdict` on the wire. */
|
|
2057
|
+
label: boolean;
|
|
2058
|
+
annotation: string;
|
|
2059
|
+
confidence?: VerdictConfidence;
|
|
2060
|
+
evidence?: Justification | null;
|
|
2061
|
+
verdict?: never;
|
|
2062
|
+
skip?: never;
|
|
2063
|
+
archive?: never;
|
|
2064
|
+
} | {
|
|
2065
|
+
/** @deprecated Use `verdict: "abstained"`, or {@link VerdictsClient.abstain}. */
|
|
2026
2066
|
abstain: true;
|
|
2027
2067
|
annotation?: string;
|
|
2068
|
+
verdict?: never;
|
|
2069
|
+
skip?: never;
|
|
2070
|
+
archive?: never;
|
|
2071
|
+
} | {
|
|
2072
|
+
skip: true;
|
|
2073
|
+
annotation?: string;
|
|
2074
|
+
verdict?: never;
|
|
2075
|
+
label?: never;
|
|
2076
|
+
abstain?: never;
|
|
2077
|
+
archive?: never;
|
|
2028
2078
|
};
|
|
2029
|
-
type
|
|
2079
|
+
type VerdictArchive = {
|
|
2030
2080
|
archive: true;
|
|
2031
|
-
|
|
2032
|
-
|
|
2081
|
+
verdict?: never;
|
|
2082
|
+
label?: never;
|
|
2083
|
+
abstain?: never;
|
|
2084
|
+
skip?: never;
|
|
2085
|
+
};
|
|
2086
|
+
/** One agent verdict write: a verdict, a skip, or an archive. */
|
|
2087
|
+
type VerdictUpdate = (VerdictTarget & VerdictJudgment) | (ArchiveVerdictTarget & VerdictArchive);
|
|
2088
|
+
/** What the server did with one agent write. `key` echoes the id the write addressed. */
|
|
2089
|
+
type VerdictOutcome = {
|
|
2033
2090
|
key: string;
|
|
2034
2091
|
traceId: string;
|
|
2035
|
-
action:
|
|
2092
|
+
action: VerdictAction;
|
|
2036
2093
|
};
|
|
2037
|
-
|
|
2094
|
+
/** A verdict a person decided, on one assertion. A person never abstains. */
|
|
2095
|
+
type HumanVerdictUpdate = {
|
|
2038
2096
|
traceId: string;
|
|
2039
2097
|
assertionId: string;
|
|
2040
|
-
label: boolean;
|
|
2041
2098
|
annotation: string;
|
|
2042
|
-
confidence?:
|
|
2099
|
+
confidence?: VerdictConfidence;
|
|
2043
2100
|
evidence?: Justification | null;
|
|
2044
|
-
}
|
|
2045
|
-
|
|
2101
|
+
} & ({
|
|
2102
|
+
verdict: PassFailVerdict;
|
|
2103
|
+
label?: never;
|
|
2104
|
+
} | {
|
|
2105
|
+
/** @deprecated Use `verdict: "pass" | "fail"`. Sent as `verdict` on the wire. */
|
|
2106
|
+
label: boolean;
|
|
2107
|
+
verdict?: never;
|
|
2108
|
+
});
|
|
2109
|
+
/** What the server did with one human verdict. */
|
|
2110
|
+
type HumanVerdictOutcome = {
|
|
2046
2111
|
traceId: string;
|
|
2047
2112
|
assertionId: string | null;
|
|
2113
|
+
verdict: PassFailVerdict;
|
|
2114
|
+
/** @deprecated Use {@link HumanVerdictOutcome.verdict}. True when the verdict is `pass`. */
|
|
2048
2115
|
label: boolean;
|
|
2049
2116
|
action: "set";
|
|
2050
2117
|
};
|
|
2118
|
+
/** One assertion's verdict on a trace. */
|
|
2051
2119
|
type AssertionVerdict = {
|
|
2052
2120
|
assertionId: string;
|
|
2053
2121
|
assertion: string | null;
|
|
2054
|
-
|
|
2055
|
-
label: boolean | null;
|
|
2056
|
-
/** The verdict, or null when there is none. `label` carries the same verdict as a boolean, false for both fail and abstained. */
|
|
2122
|
+
/** The verdict, or null when there is none. */
|
|
2057
2123
|
verdict: Verdict | null;
|
|
2124
|
+
verdictStatus: VerdictStatus;
|
|
2125
|
+
source: VerdictSource;
|
|
2058
2126
|
annotation: string | null;
|
|
2059
2127
|
evidence: Justification | null;
|
|
2060
|
-
confidence:
|
|
2061
|
-
labelSource: LabelSource;
|
|
2128
|
+
confidence: VerdictConfidence | null;
|
|
2062
2129
|
approved: boolean;
|
|
2063
|
-
}
|
|
2064
|
-
type TraceLabels = {
|
|
2065
|
-
traceId: string;
|
|
2130
|
+
/** @deprecated Use {@link AssertionVerdict.verdictStatus}. */
|
|
2066
2131
|
labelStatus: LabelStatus;
|
|
2132
|
+
/** @deprecated Use {@link AssertionVerdict.verdict}. False for both fail and abstained. */
|
|
2067
2133
|
label: boolean | null;
|
|
2134
|
+
/** @deprecated Use {@link AssertionVerdict.source}. */
|
|
2135
|
+
labelSource: VerdictSource;
|
|
2136
|
+
};
|
|
2137
|
+
/** A trace's effective verdict plus one row per scored assertion. */
|
|
2138
|
+
type TraceVerdicts = {
|
|
2139
|
+
traceId: string;
|
|
2140
|
+
verdict: Verdict | null;
|
|
2141
|
+
verdictStatus: VerdictStatus;
|
|
2068
2142
|
annotation: string | null;
|
|
2069
2143
|
evidence: Justification | null;
|
|
2070
2144
|
approved: boolean;
|
|
@@ -2073,73 +2147,159 @@ type TraceLabels = {
|
|
|
2073
2147
|
/** How many of the failed assertion verdicts are abstentions, where the judge could not find the evidence. */
|
|
2074
2148
|
abstained: number;
|
|
2075
2149
|
assertions: AssertionVerdict[];
|
|
2150
|
+
/** @deprecated Use {@link TraceVerdicts.verdictStatus}. */
|
|
2151
|
+
labelStatus: LabelStatus;
|
|
2152
|
+
/** @deprecated Use {@link TraceVerdicts.verdict}. */
|
|
2153
|
+
label: boolean | null;
|
|
2076
2154
|
};
|
|
2077
|
-
|
|
2155
|
+
/** @deprecated Use {@link VerdictConfidence}. Same type. */
|
|
2156
|
+
type LabelConfidence = VerdictConfidence;
|
|
2157
|
+
/** @deprecated Use {@link VerdictAction}. Archives that clear nothing now report `no-active-verdict`. */
|
|
2158
|
+
type LabelAction = VerdictAction;
|
|
2159
|
+
/** @deprecated Use {@link VerdictStatus}. Read by the deprecated `labelStatus` fields only. */
|
|
2160
|
+
type LabelStatus = "labeled" | "skipped" | "unlabeled";
|
|
2161
|
+
/** @deprecated Use {@link VerdictSource}. Same type. */
|
|
2162
|
+
type LabelSource = VerdictSource;
|
|
2163
|
+
/** @deprecated Use {@link VerdictTarget}. Same type. */
|
|
2164
|
+
type LabelTarget = VerdictTarget;
|
|
2165
|
+
/** @deprecated Use {@link ArchiveVerdictTarget}. Same type. */
|
|
2166
|
+
type ArchiveLabelTarget = ArchiveVerdictTarget;
|
|
2167
|
+
/** @deprecated Use {@link VerdictUpdate}. Same type. */
|
|
2168
|
+
type LabelUpdate = VerdictUpdate;
|
|
2169
|
+
/** @deprecated Use {@link VerdictOutcome}. Same type. */
|
|
2170
|
+
type LabelOutcome = VerdictOutcome;
|
|
2171
|
+
/** @deprecated Use {@link HumanVerdictUpdate}. Same type. */
|
|
2172
|
+
type HumanLabelUpdate = HumanVerdictUpdate;
|
|
2173
|
+
/** @deprecated Use {@link HumanVerdictOutcome}. Same type. */
|
|
2174
|
+
type HumanLabelOutcome = HumanVerdictOutcome;
|
|
2175
|
+
/** @deprecated Use {@link TraceVerdicts}. Same type. */
|
|
2176
|
+
type TraceLabels = TraceVerdicts;
|
|
2177
|
+
/** Writes and reads assertion verdicts: what a judge decided about each assertion on a trace. */
|
|
2178
|
+
declare class VerdictsClient {
|
|
2078
2179
|
private readonly httpClient;
|
|
2079
2180
|
constructor(httpClient: HttpClient);
|
|
2080
|
-
|
|
2081
|
-
|
|
2082
|
-
|
|
2181
|
+
/** Save one agent verdict. Pass `experimentId` when the trace is a replay addressed by `originalTraceId`. */
|
|
2182
|
+
save(update: VerdictUpdate, experimentId?: string): Promise<VerdictOutcome>;
|
|
2183
|
+
/** Save many agent verdicts in one all-or-nothing batch. */
|
|
2184
|
+
saveAll(updates: VerdictUpdate[], experimentId?: string): Promise<VerdictOutcome[]>;
|
|
2185
|
+
/** Withhold a verdict, with `annotation` saying why, for an attempt that crashed or an assertion that could not be checked. */
|
|
2186
|
+
skip(target: VerdictTarget & {
|
|
2083
2187
|
annotation?: string;
|
|
2084
|
-
}, experimentId?: string): Promise<
|
|
2188
|
+
}, experimentId?: string): Promise<VerdictOutcome>;
|
|
2085
2189
|
/**
|
|
2086
2190
|
* Record that the evidence does not show the part of the run this assertion
|
|
2087
2191
|
* checks, so there was nothing to judge. Counted as a fail in every pass
|
|
2088
2192
|
* rate, but stored and read back as an abstention, apart from a fail. Pass
|
|
2089
2193
|
* `annotation` to say what the evidence would have had to show.
|
|
2090
2194
|
*/
|
|
2091
|
-
abstain(target:
|
|
2195
|
+
abstain(target: VerdictTarget & {
|
|
2092
2196
|
annotation?: string;
|
|
2093
|
-
}, experimentId?: string): Promise<
|
|
2094
|
-
|
|
2095
|
-
|
|
2096
|
-
|
|
2097
|
-
|
|
2098
|
-
|
|
2197
|
+
}, experimentId?: string): Promise<VerdictOutcome>;
|
|
2198
|
+
/** Clear a previous verdict. `assertionId` may be omitted to clear an old whole-trace verdict. */
|
|
2199
|
+
archive(target: ArchiveVerdictTarget, experimentId?: string): Promise<VerdictOutcome>;
|
|
2200
|
+
/** Save one verdict a person decided. */
|
|
2201
|
+
saveHuman(update: HumanVerdictUpdate): Promise<HumanVerdictOutcome>;
|
|
2202
|
+
/** Save many verdicts a person decided, in one all-or-nothing batch. */
|
|
2203
|
+
saveHumanAll(updates: HumanVerdictUpdate[]): Promise<HumanVerdictOutcome[]>;
|
|
2204
|
+
/** Read one trace's verdicts, or null when the trace has none. */
|
|
2205
|
+
get(traceId: string): Promise<TraceVerdicts | null>;
|
|
2206
|
+
/** Read many traces' verdicts in one call. */
|
|
2207
|
+
getAll(traceIds: string[]): Promise<TraceVerdicts[]>;
|
|
2099
2208
|
/** Generate suggested evidence for an assertion on an original or replay trace. */
|
|
2100
|
-
|
|
2209
|
+
generateEvidence(traceId: string, assertionId: string): Promise<AssertionVerdictEvidence[]>;
|
|
2210
|
+
/** @deprecated Use {@link VerdictsClient.generateEvidence}. Same call. */
|
|
2211
|
+
generateLabelEvidence(traceId: string, assertionId: string): Promise<AssertionVerdictEvidence[]>;
|
|
2101
2212
|
}
|
|
2213
|
+
/**
|
|
2214
|
+
* @deprecated Use {@link VerdictsClient}. Same class.
|
|
2215
|
+
* @alias
|
|
2216
|
+
*/
|
|
2217
|
+
declare const LabelsClient: typeof VerdictsClient;
|
|
2218
|
+
/**
|
|
2219
|
+
* @deprecated Use {@link VerdictsClient}. Same class.
|
|
2220
|
+
* @alias
|
|
2221
|
+
*/
|
|
2222
|
+
type LabelsClient = VerdictsClient;
|
|
2102
2223
|
//#endregion
|
|
2103
2224
|
//#region src/graders.d.ts
|
|
2104
|
-
|
|
2105
|
-
type
|
|
2225
|
+
/** Who wrote a grader verdict: a person, or the grader itself. */
|
|
2226
|
+
type GraderVerdictSource = "human" | "live_grader";
|
|
2227
|
+
/** One grader's verdict on one trace. */
|
|
2228
|
+
type GraderVerdict = {
|
|
2106
2229
|
traceId: string;
|
|
2107
2230
|
graderId: string;
|
|
2108
2231
|
graderName: string | null;
|
|
2109
2232
|
graderStatus: string;
|
|
2110
2233
|
graderType: string;
|
|
2111
|
-
|
|
2112
|
-
|
|
2234
|
+
/** The grader's verdict, or null when it has not judged the trace. */
|
|
2235
|
+
verdict: PassFailVerdict | null;
|
|
2236
|
+
reason: string | null;
|
|
2113
2237
|
failureDiagnostic: string | null;
|
|
2114
|
-
|
|
2115
|
-
source:
|
|
2238
|
+
confidence: VerdictConfidence | null;
|
|
2239
|
+
source: GraderVerdictSource;
|
|
2116
2240
|
evaluatedAt: string | null;
|
|
2241
|
+
/** @deprecated Use {@link GraderVerdict.verdict}. True when the verdict is `pass`. */
|
|
2242
|
+
label: boolean | null;
|
|
2243
|
+
/** @deprecated Use {@link GraderVerdict.reason}. */
|
|
2244
|
+
labelReason: string | null;
|
|
2245
|
+
/** @deprecated Use {@link GraderVerdict.confidence}. */
|
|
2246
|
+
labelConfidence: VerdictConfidence | null;
|
|
2117
2247
|
};
|
|
2118
|
-
|
|
2248
|
+
/** Pass `traceIds`, `graderId`, or both. Passing neither throws. */
|
|
2249
|
+
type GetGraderVerdictsParams = {
|
|
2119
2250
|
traceIds?: string[];
|
|
2120
2251
|
graderId?: string;
|
|
2121
2252
|
limit?: number;
|
|
2122
2253
|
};
|
|
2123
|
-
|
|
2254
|
+
/** A verdict for an external grader on one trace. A grader never abstains. */
|
|
2255
|
+
type GraderVerdictUpdate = {
|
|
2124
2256
|
graderId: string;
|
|
2125
2257
|
traceId: string;
|
|
2126
|
-
label: boolean;
|
|
2127
2258
|
reason?: string;
|
|
2128
2259
|
failureDiagnostic?: string;
|
|
2129
|
-
confidence?:
|
|
2130
|
-
}
|
|
2131
|
-
|
|
2260
|
+
confidence?: VerdictConfidence;
|
|
2261
|
+
} & ({
|
|
2262
|
+
verdict: PassFailVerdict;
|
|
2263
|
+
label?: never;
|
|
2264
|
+
} | {
|
|
2265
|
+
/** @deprecated Use `verdict: "pass" | "fail"`. Sent as `verdict` on the wire. */
|
|
2266
|
+
label: boolean;
|
|
2267
|
+
verdict?: never;
|
|
2268
|
+
});
|
|
2269
|
+
/** What the server did with one grader verdict. */
|
|
2270
|
+
type GraderVerdictOutcome = {
|
|
2132
2271
|
graderId: string;
|
|
2133
2272
|
traceId: string;
|
|
2273
|
+
verdict: PassFailVerdict;
|
|
2274
|
+
/** @deprecated Use {@link GraderVerdictOutcome.verdict}. True when the verdict is `pass`. */
|
|
2134
2275
|
label: boolean;
|
|
2135
2276
|
action: "set";
|
|
2136
2277
|
};
|
|
2278
|
+
/** @deprecated Use {@link GraderVerdictSource}. Same type. */
|
|
2279
|
+
type GraderLabelSource = GraderVerdictSource;
|
|
2280
|
+
/** @deprecated Use {@link GraderVerdict}. Same type. */
|
|
2281
|
+
type GraderLabel = GraderVerdict;
|
|
2282
|
+
/** @deprecated Use {@link GetGraderVerdictsParams}. Same type. */
|
|
2283
|
+
type GetGraderLabelsParams = GetGraderVerdictsParams;
|
|
2284
|
+
/** @deprecated Use {@link GraderVerdictUpdate}. Same type. */
|
|
2285
|
+
type GraderLabelUpdate = GraderVerdictUpdate;
|
|
2286
|
+
/** @deprecated Use {@link GraderVerdictOutcome}. Same type. */
|
|
2287
|
+
type GraderLabelOutcome = GraderVerdictOutcome;
|
|
2137
2288
|
declare class GradersClient {
|
|
2138
2289
|
private readonly httpClient;
|
|
2139
2290
|
constructor(httpClient: HttpClient);
|
|
2140
|
-
|
|
2141
|
-
|
|
2142
|
-
|
|
2291
|
+
/** Read each grader's verdict, one row per grader per trace. */
|
|
2292
|
+
getVerdicts(params: GetGraderVerdictsParams): Promise<GraderVerdict[]>;
|
|
2293
|
+
/** Save one verdict for an external grader. */
|
|
2294
|
+
saveVerdict(update: GraderVerdictUpdate): Promise<GraderVerdictOutcome>;
|
|
2295
|
+
/** Save up to 200 verdicts for external graders, all or nothing. */
|
|
2296
|
+
saveVerdictAll(updates: GraderVerdictUpdate[]): Promise<GraderVerdictOutcome[]>;
|
|
2297
|
+
/** @deprecated Use {@link GradersClient.getVerdicts}. Same call. */
|
|
2298
|
+
getLabels(params: GetGraderVerdictsParams): Promise<GraderVerdict[]>;
|
|
2299
|
+
/** @deprecated Use {@link GradersClient.saveVerdict}. Same call. */
|
|
2300
|
+
saveLabel(update: GraderVerdictUpdate): Promise<GraderVerdictOutcome>;
|
|
2301
|
+
/** @deprecated Use {@link GradersClient.saveVerdictAll}. Same call. */
|
|
2302
|
+
saveLabelAll(updates: GraderVerdictUpdate[]): Promise<GraderVerdictOutcome[]>;
|
|
2143
2303
|
}
|
|
2144
2304
|
//#endregion
|
|
2145
2305
|
//#region src/langgraph.d.ts
|
|
@@ -3639,7 +3799,10 @@ declare class Bitfab {
|
|
|
3639
3799
|
readonly organizationMembers: OrganizationMembersClient;
|
|
3640
3800
|
readonly traces: TracesClient;
|
|
3641
3801
|
readonly spans: SpansClient;
|
|
3642
|
-
|
|
3802
|
+
/** Assertion verdicts: what a judge decided about each assertion on a trace. */
|
|
3803
|
+
readonly verdicts: VerdictsClient;
|
|
3804
|
+
/** @deprecated Use {@link verdicts}. Same client. */
|
|
3805
|
+
readonly labels: VerdictsClient;
|
|
3643
3806
|
readonly graders: GradersClient;
|
|
3644
3807
|
readonly experiments: ExperimentsClient;
|
|
3645
3808
|
private readonly bamlClient;
|
|
@@ -4352,7 +4515,7 @@ interface CommitRef {
|
|
|
4352
4515
|
/**
|
|
4353
4516
|
* SDK version from package.json (injected at build time)
|
|
4354
4517
|
*/
|
|
4355
|
-
declare const __version__ = "0.
|
|
4518
|
+
declare const __version__ = "0.66.0";
|
|
4356
4519
|
//#endregion
|
|
4357
4520
|
//#region src/constants.d.ts
|
|
4358
4521
|
/**
|
|
@@ -4494,4 +4657,4 @@ declare function seedFromRegistry(registry: ReplayRegistry, pipeline: string, ca
|
|
|
4494
4657
|
run?: boolean;
|
|
4495
4658
|
}): Promise<SeedResult>;
|
|
4496
4659
|
//#endregion
|
|
4497
|
-
export { AdaptInputsFn as $,
|
|
4660
|
+
export { AdaptInputsFn as $, ExperimentRunBy as $n, NO_MOCK_OVERRIDE as $r, Verdict as $t, SpanMethodDecoratorContext as A, ScenarioData as An, BitfabClaudeAgentHandler as Ar, GraderVerdict as At, TraceResponse as B, TraceSearchFilters as Bn, CodeChangeFile as Br, HumanVerdictOutcome as Bt, NodeMethodDecorator as C, SaveScenarioData as Cn, SaveDatasetResult as Cr, BitfabLangGraphCallbackHandler as Ct, SeedCaseOptions as D, Scenario as Dn, VercelCallParams as Dr, GraderLabelOutcome as Dt, ReseedResult as E, SavedTraceAssertion as En, BitfabVercelAiHandler as Er, GraderLabel as Et, getCurrentSpan as F, ScenarioTrigger as Fn, Category as Fr, ArchiveLabelTarget as Ft, SpanListResult as G, TracesClient as Gn, TokenUsage as Gr, LabelSource as Gt, ListedSpan as H, TraceSearchResult as Hn, SpanLookup as Hr, LabelAction as Ht, getCurrentTrace as I, TraceAssertion as In, CategorySummary as Ir, ArchiveVerdictTarget as It, RivetContextLike as J, ExperimentCodeChangeTotals as Jn, MockOverride as Jr, LabelUpdate as Jt, SpansClient as K, Experiment as Kn, flushTraces as Kr, LabelStatus as Kt, CaptureSurface as L, TraceAssertionSource as Ln, SaveAssertionCategoryParams as Lr, AssertionVerdict as Lt, WrapBAMLOptions as M, ScenarioDataOrigin as Mn, AssertionCategory as Mr, GraderVerdictSource as Mt, WrappedBamlFn as N, ScenarioHistoryEntry as Nn, AssertionCategorySummary as Nr, GraderVerdictUpdate as Nt, SeedRunOptions as O, ScenarioAction as On, VercelGenerateResult as Or, GraderLabelSource as Ot, getCurrentReplayBranch as P, ScenarioSource as Pn, CategoriesClient as Pr, GradersClient as Pt, AdaptContext as Q, ExperimentRollupResult as Qn, MockValue as Qr, TraceVerdicts as Qt, ActiveSpanContext as R, TraceAssertionsResult as Rn, SaveCategoryParams as Rr, HumanLabelOutcome as Rt, DetachedTrace as S, SaveScenarioActualOutcome as Sn, SaveDatasetParams as Sr, LangGraphIntegrationOptions as St, ReseedOptions as T, SaveScenarioParams as Tn, BitfabLanguageModelMiddleware as Tr, GetGraderVerdictsParams as Tt, SpanListField as U, TraceTarget as Un, SpanOccurrence as Ur, LabelConfidence as Ut, TracingProcessor as V, TraceSearchParams as Vn, HttpClient as Vr, HumanVerdictUpdate as Vt, SpanListParams as W, TraceTargetOccurrence as Wn, SpanType as Wr, LabelOutcome as Wt, RivetIntegrationOptions as X, ExperimentPage as Xn, MockOverrideInput as Xr, PassFailVerdict as Xt, RivetEffectType as Y, ExperimentJitter as Yn, MockOverrideCtx as Yr, LabelsClient as Yt, ReplayBranch as Z, ExperimentRollup as Zn, MockOverrideResolver as Zr, TraceLabels as Zt, BitfabConfig as _, Approver as _i, AssertionVerdictEvidence as _n, ListDatasetsParams as _r, ReplayInterruptSignal as _t, ReplayRegistryOptions as a, DbSnapshotProvider as ai, VerdictTarget as an, AddDatasetGradersResult as ar, ReplayExperiment as at, CurrentSpan as b, JustificationSpan as bi, SaveAssertionsAllParams as bn, RerunGradersOptions as br, BitfabOpenAIAgentHandler as bt, SeedResult as c, ReplayNodeIdentity as ci, ArchiveAssertionsParams as cn, DatasetGraderRef as cr, ReplayItemStartProgress as ct, seedFromRegistry as d, SelectiveReplayReport as di, AssertionGenerationError as dn, GraderRerun as dr, ReplayProgressItem as dt, NodeMatcher as ei, VerdictAction as en, ExperimentStatus as er, BITFAB_PROGRESS_PREFIX as et, finalizers as f, AllowedEnvVars as fi, AssertionGenerationStatus as fn, GraderRerunError as fr, ReplayResult as ft, Bitfab as g, ApprovalState as gi, AssertionRewriteStatus as gn, GraderRerunTimeoutError as gr, ReplayInterrupt as gt, CommitRef as h, ApprovalFields as hi, AssertionRewrite as hn, GraderRerunStatus as hr, serializeReplayResult as ht, ReplayRegistryContext as i, DbSnapshotConfig as ii, VerdictStatus as in, ListExperimentsParams as ir, ReplayError as it, SpanOptions as j, ScenarioDataAppend as jn, AssertionCategoriesClient as jr, GraderVerdictOutcome as jt, SpanMethodDecorator as k, ScenarioActualOutcome as kn, VercelStreamResult as kr, GraderLabelUpdate as kt, defineReplayRegistry as l, SelectiveReplayDecision as li, AssertionEvidenceParameter as ln, DatasetTraceIds as lr, ReplayOptions as lt, __version__ as m, ProviderDefinition as mi, AssertionLabelEvidence as mn, GraderRerunResult as mr, reportReplayProgress as mt, ReplayRegistration as n, BitfabError as ni, VerdictOutcome as nn, ExperimentTotals as nr, DbBranchReplayError as nt, ReseedRegistryResult as o, DbSnapshotRef as oi, VerdictUpdate as on, AddDatasetTracesResult as or, ReplayItem as ot, DEFAULT_SERVICE_URL as p, BamlExecutionResult as pi, AssertionGenerationTimeoutError as pn, GraderRerunProgress as pr, TraceIngestionType as pt, BitfabRivetIntegration as q, ExperimentCategoryTally as qn, DbBranchTimings as qr, LabelTarget as qt, ReplayRegistry as r, MixedTracingError as ri, VerdictSource as rn, ExperimentsClient as rr, MockStrategy as rt, SeedCase as s, SUPPORTED_PROVIDERS as si, VerdictsClient as sn, Dataset as sr, ReplayItemFinishProgress as st, ReplayOptionsFactory as t, SpanNodeMeta as ti, VerdictConfidence as tn, ExperimentTally as tr, DbBranchOptions as tt, reseedFromRegistry as u, SelectiveReplayOptions as ui, AssertionGeneration as un, DatasetsClient as ur, ReplayProgress as ut, BitfabFunction as v, Assignee as vi, GenerateAssertionsOptions as vn, RemoveDatasetGradersResult as vr, OrganizationMember as vt, NodeOptions as w, SaveScenarioHistoryEntry as wn, UpdateDatasetParams as wr, GetGraderLabelsParams as wt, CurrentTrace as x, SaveAssertionsParams as xn, RerunGradersResult as xr, BitfabLangGraphIntegration as xt, CaptureWhen as y, Justification as yi, SaveAssertion as yn, RemoveDatasetTracesResult as yr, OrganizationMembersClient as yt, BitfabOpenAITracingProcessor as z, TraceAssertionsUpdate as zn, CapturedSpan as zr, HumanLabelUpdate as zt };
|