peerbench 0.0.1 → 0.0.2-alpha.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/README.md +308 -2
- package/dist/abstract-Dec9Sc5O.d.ts +12 -0
- package/dist/benchmarks/index.d.ts +1698 -0
- package/dist/benchmarks/index.js +915 -0
- package/dist/benchmarks/index.js.map +1 -0
- package/dist/catalogs/index.d.ts +75 -0
- package/dist/catalogs/index.js +88 -0
- package/dist/catalogs/index.js.map +1 -0
- package/dist/chunk-22HU24QF.js +8 -0
- package/dist/chunk-22HU24QF.js.map +1 -0
- package/dist/chunk-232PY7K3.js +50 -0
- package/dist/chunk-232PY7K3.js.map +1 -0
- package/dist/chunk-7TREBPSJ.js +26 -0
- package/dist/chunk-7TREBPSJ.js.map +1 -0
- package/dist/chunk-DUBKY73H.js +128 -0
- package/dist/chunk-DUBKY73H.js.map +1 -0
- package/dist/chunk-GVF4YZF3.js +15 -0
- package/dist/chunk-GVF4YZF3.js.map +1 -0
- package/dist/chunk-HJH3SW3L.js +103 -0
- package/dist/chunk-HJH3SW3L.js.map +1 -0
- package/dist/chunk-IUN2IUCS.js +58 -0
- package/dist/chunk-IUN2IUCS.js.map +1 -0
- package/dist/chunk-PZ5AY32C.js +10 -0
- package/dist/chunk-PZ5AY32C.js.map +1 -0
- package/dist/chunk-VBOM2YEG.js +47 -0
- package/dist/chunk-VBOM2YEG.js.map +1 -0
- package/dist/chunk-ZJWSK4VO.js +11 -0
- package/dist/chunk-ZJWSK4VO.js.map +1 -0
- package/dist/data-BmN5WjZ4.d.ts +57 -0
- package/dist/generic-array-DLHWSvf1.d.ts +22 -0
- package/dist/index-WiPjF2AL.d.ts +15 -0
- package/dist/index.d.ts +38 -3845
- package/dist/index.js +40 -3557
- package/dist/index.js.map +1 -1
- package/dist/llm-DNj_tp2T.d.ts +22 -0
- package/dist/llm-judge-DIG1f1Az.d.ts +67 -0
- package/dist/provider-BDjGp2y-.d.ts +10 -0
- package/dist/providers/index.d.ts +72 -0
- package/dist/providers/index.js +263 -0
- package/dist/providers/index.js.map +1 -0
- package/dist/rate-limiter-CSmVIRsM.d.ts +60 -0
- package/dist/schemas/extensions/index.d.ts +14 -0
- package/dist/schemas/extensions/index.js +13 -0
- package/dist/schemas/extensions/index.js.map +1 -0
- package/dist/schemas/index.d.ts +233 -0
- package/dist/schemas/index.js +27 -0
- package/dist/schemas/index.js.map +1 -0
- package/dist/schemas/llm/index.d.ts +98 -0
- package/dist/schemas/llm/index.js +37 -0
- package/dist/schemas/llm/index.js.map +1 -0
- package/dist/scorers/index.d.ts +63 -0
- package/dist/scorers/index.js +494 -0
- package/dist/scorers/index.js.map +1 -0
- package/dist/simple-system-prompt-CzPYuvo0.d.ts +49 -0
- package/dist/system-prompt--0FdPWqK.d.ts +58 -0
- package/dist/utilities-BrRH32rD.d.ts +30 -0
- package/package.json +39 -21
- package/LICENSE +0 -21
|
@@ -0,0 +1,233 @@
|
|
|
1
|
+
import { I as IdGenerator } from '../index-WiPjF2AL.js';
|
|
2
|
+
export { b as IdSchema } from '../index-WiPjF2AL.js';
|
|
3
|
+
import z__default, { z } from 'zod';
|
|
4
|
+
|
|
5
|
+
declare const BaseTestCaseSchemaV1: z.ZodObject<{
|
|
6
|
+
id: z.ZodString;
|
|
7
|
+
kind: z.ZodString;
|
|
8
|
+
schemaVersion: z.ZodNumber;
|
|
9
|
+
metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
|
|
10
|
+
}, z.core.$strip>;
|
|
11
|
+
type BaseTestCaseV1 = z.infer<typeof BaseTestCaseSchemaV1>;
|
|
12
|
+
declare const defineTestCaseSchema: <TBaseSchema extends z.ZodObject<Omit<{
|
|
13
|
+
id: z.ZodString;
|
|
14
|
+
kind: z.ZodString;
|
|
15
|
+
schemaVersion: z.ZodNumber;
|
|
16
|
+
metadata: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
|
|
17
|
+
}, "kind" | "schemaVersion"> & {
|
|
18
|
+
kind: z.ZodString | z.ZodLiteral<string>;
|
|
19
|
+
schemaVersion: z.ZodNumber | z.ZodLiteral<number>;
|
|
20
|
+
}, z.core.$strip>, TKind extends string | undefined = undefined, TSchemaVersion extends number | undefined = undefined, TFields extends Readonly<{
|
|
21
|
+
[k: string]: z.core.$ZodType<unknown, unknown, z.core.$ZodTypeInternals<unknown, unknown>>;
|
|
22
|
+
}> = {}>(config: {
|
|
23
|
+
baseSchema: TBaseSchema;
|
|
24
|
+
kind?: TKind | undefined;
|
|
25
|
+
schemaVersion?: TSchemaVersion | undefined;
|
|
26
|
+
fields?: TFields | undefined;
|
|
27
|
+
}) => (TBaseSchema extends z.ZodObject<infer U extends Readonly<{
|
|
28
|
+
[k: string]: z.core.$ZodType<unknown, unknown, z.core.$ZodTypeInternals<unknown, unknown>>;
|
|
29
|
+
}>, z.core.$strip> ? z.ZodObject<Omit<U, "kind" | "schemaVersion"> & TFields & {
|
|
30
|
+
kind: TKind extends undefined ? U["kind"] : z.ZodLiteral<TKind>;
|
|
31
|
+
schemaVersion: TSchemaVersion extends undefined ? U["schemaVersion"] : z.ZodLiteral<TSchemaVersion>;
|
|
32
|
+
}, z.core.$strip> : never) & {
|
|
33
|
+
new: (input: Omit<z.core.output<TBaseSchema extends z.ZodObject<infer U extends Readonly<{
|
|
34
|
+
[k: string]: z.core.$ZodType<unknown, unknown, z.core.$ZodTypeInternals<unknown, unknown>>;
|
|
35
|
+
}>, z.core.$strip> ? z.ZodObject<Omit<U, "kind" | "schemaVersion"> & TFields & {
|
|
36
|
+
kind: TKind extends undefined ? U["kind"] : z.ZodLiteral<TKind>;
|
|
37
|
+
schemaVersion: TSchemaVersion extends undefined ? U["schemaVersion"] : z.ZodLiteral<TSchemaVersion>;
|
|
38
|
+
}, z.core.$strip> : never>, "kind" | "schemaVersion">) => z.core.output<TBaseSchema extends z.ZodObject<infer U extends Readonly<{
|
|
39
|
+
[k: string]: z.core.$ZodType<unknown, unknown, z.core.$ZodTypeInternals<unknown, unknown>>;
|
|
40
|
+
}>, z.core.$strip> ? z.ZodObject<Omit<U, "kind" | "schemaVersion"> & TFields & {
|
|
41
|
+
kind: TKind extends undefined ? U["kind"] : z.ZodLiteral<TKind>;
|
|
42
|
+
schemaVersion: TSchemaVersion extends undefined ? U["schemaVersion"] : z.ZodLiteral<TSchemaVersion>;
|
|
43
|
+
}, z.core.$strip> : never>;
|
|
44
|
+
newWithId(input: Omit<z.core.output<TBaseSchema extends z.ZodObject<infer U extends Readonly<{
|
|
45
|
+
[k: string]: z.core.$ZodType<unknown, unknown, z.core.$ZodTypeInternals<unknown, unknown>>;
|
|
46
|
+
}>, z.core.$strip> ? z.ZodObject<Omit<U, "kind" | "schemaVersion"> & TFields & {
|
|
47
|
+
kind: TKind extends undefined ? U["kind"] : z.ZodLiteral<TKind>;
|
|
48
|
+
schemaVersion: TSchemaVersion extends undefined ? U["schemaVersion"] : z.ZodLiteral<TSchemaVersion>;
|
|
49
|
+
}, z.core.$strip> : never>, "kind" | "schemaVersion" | "id">, generator: IdGenerator): Promise<z.core.output<TBaseSchema extends z.ZodObject<infer U extends Readonly<{
|
|
50
|
+
[k: string]: z.core.$ZodType<unknown, unknown, z.core.$ZodTypeInternals<unknown, unknown>>;
|
|
51
|
+
}>, z.core.$strip> ? z.ZodObject<Omit<U, "kind" | "schemaVersion"> & TFields & {
|
|
52
|
+
kind: TKind extends undefined ? U["kind"] : z.ZodLiteral<TKind>;
|
|
53
|
+
schemaVersion: TSchemaVersion extends undefined ? U["schemaVersion"] : z.ZodLiteral<TSchemaVersion>;
|
|
54
|
+
}, z.core.$strip> : never>>;
|
|
55
|
+
};
|
|
56
|
+
|
|
57
|
+
declare const BaseResponseSchemaV1: z__default.ZodObject<{
|
|
58
|
+
id: z__default.ZodString;
|
|
59
|
+
kind: z__default.ZodString;
|
|
60
|
+
schemaVersion: z__default.ZodNumber;
|
|
61
|
+
startedAt: z__default.ZodNumber;
|
|
62
|
+
completedAt: z__default.ZodNumber;
|
|
63
|
+
testCaseId: z__default.ZodString;
|
|
64
|
+
metadata: z__default.ZodOptional<z__default.ZodRecord<z__default.ZodString, z__default.ZodUnknown>>;
|
|
65
|
+
}, z__default.core.$strip>;
|
|
66
|
+
type BaseResponseV1 = z__default.infer<typeof BaseResponseSchemaV1>;
|
|
67
|
+
declare const defineResponseSchema: <TBaseSchema extends z__default.ZodObject<Omit<{
|
|
68
|
+
id: z__default.ZodString;
|
|
69
|
+
kind: z__default.ZodString;
|
|
70
|
+
schemaVersion: z__default.ZodNumber;
|
|
71
|
+
startedAt: z__default.ZodNumber;
|
|
72
|
+
completedAt: z__default.ZodNumber;
|
|
73
|
+
testCaseId: z__default.ZodString;
|
|
74
|
+
metadata: z__default.ZodOptional<z__default.ZodRecord<z__default.ZodString, z__default.ZodUnknown>>;
|
|
75
|
+
}, "kind" | "schemaVersion"> & {
|
|
76
|
+
kind: z__default.ZodString | z__default.ZodLiteral<string>;
|
|
77
|
+
schemaVersion: z__default.ZodNumber | z__default.ZodLiteral<number>;
|
|
78
|
+
}, z__default.core.$strip>, TKind extends string | undefined = undefined, TSchemaVersion extends number | undefined = undefined, TFields extends Readonly<{
|
|
79
|
+
[k: string]: z__default.core.$ZodType<unknown, unknown, z__default.core.$ZodTypeInternals<unknown, unknown>>;
|
|
80
|
+
}> = {}>(config: {
|
|
81
|
+
baseSchema: TBaseSchema;
|
|
82
|
+
kind?: TKind | undefined;
|
|
83
|
+
schemaVersion?: TSchemaVersion | undefined;
|
|
84
|
+
fields?: TFields | undefined;
|
|
85
|
+
}) => (TBaseSchema extends z__default.ZodObject<infer U extends Readonly<{
|
|
86
|
+
[k: string]: z__default.core.$ZodType<unknown, unknown, z__default.core.$ZodTypeInternals<unknown, unknown>>;
|
|
87
|
+
}>, z__default.core.$strip> ? z__default.ZodObject<Omit<U, "kind" | "schemaVersion"> & TFields & {
|
|
88
|
+
kind: TKind extends undefined ? U["kind"] : z__default.ZodLiteral<TKind>;
|
|
89
|
+
schemaVersion: TSchemaVersion extends undefined ? U["schemaVersion"] : z__default.ZodLiteral<TSchemaVersion>;
|
|
90
|
+
}, z__default.core.$strip> : never) & {
|
|
91
|
+
new: (input: Omit<z__default.core.output<TBaseSchema extends z__default.ZodObject<infer U extends Readonly<{
|
|
92
|
+
[k: string]: z__default.core.$ZodType<unknown, unknown, z__default.core.$ZodTypeInternals<unknown, unknown>>;
|
|
93
|
+
}>, z__default.core.$strip> ? z__default.ZodObject<Omit<U, "kind" | "schemaVersion"> & TFields & {
|
|
94
|
+
kind: TKind extends undefined ? U["kind"] : z__default.ZodLiteral<TKind>;
|
|
95
|
+
schemaVersion: TSchemaVersion extends undefined ? U["schemaVersion"] : z__default.ZodLiteral<TSchemaVersion>;
|
|
96
|
+
}, z__default.core.$strip> : never>, "kind" | "schemaVersion">) => z__default.core.output<TBaseSchema extends z__default.ZodObject<infer U extends Readonly<{
|
|
97
|
+
[k: string]: z__default.core.$ZodType<unknown, unknown, z__default.core.$ZodTypeInternals<unknown, unknown>>;
|
|
98
|
+
}>, z__default.core.$strip> ? z__default.ZodObject<Omit<U, "kind" | "schemaVersion"> & TFields & {
|
|
99
|
+
kind: TKind extends undefined ? U["kind"] : z__default.ZodLiteral<TKind>;
|
|
100
|
+
schemaVersion: TSchemaVersion extends undefined ? U["schemaVersion"] : z__default.ZodLiteral<TSchemaVersion>;
|
|
101
|
+
}, z__default.core.$strip> : never>;
|
|
102
|
+
newWithId(input: Omit<z__default.core.output<TBaseSchema extends z__default.ZodObject<infer U extends Readonly<{
|
|
103
|
+
[k: string]: z__default.core.$ZodType<unknown, unknown, z__default.core.$ZodTypeInternals<unknown, unknown>>;
|
|
104
|
+
}>, z__default.core.$strip> ? z__default.ZodObject<Omit<U, "kind" | "schemaVersion"> & TFields & {
|
|
105
|
+
kind: TKind extends undefined ? U["kind"] : z__default.ZodLiteral<TKind>;
|
|
106
|
+
schemaVersion: TSchemaVersion extends undefined ? U["schemaVersion"] : z__default.ZodLiteral<TSchemaVersion>;
|
|
107
|
+
}, z__default.core.$strip> : never>, "kind" | "schemaVersion" | "id">, generator: IdGenerator): Promise<z__default.core.output<TBaseSchema extends z__default.ZodObject<infer U extends Readonly<{
|
|
108
|
+
[k: string]: z__default.core.$ZodType<unknown, unknown, z__default.core.$ZodTypeInternals<unknown, unknown>>;
|
|
109
|
+
}>, z__default.core.$strip> ? z__default.ZodObject<Omit<U, "kind" | "schemaVersion"> & TFields & {
|
|
110
|
+
kind: TKind extends undefined ? U["kind"] : z__default.ZodLiteral<TKind>;
|
|
111
|
+
schemaVersion: TSchemaVersion extends undefined ? U["schemaVersion"] : z__default.ZodLiteral<TSchemaVersion>;
|
|
112
|
+
}, z__default.core.$strip> : never>>;
|
|
113
|
+
};
|
|
114
|
+
|
|
115
|
+
declare const BaseScoreSchemaV1: z__default.ZodObject<{
|
|
116
|
+
id: z__default.ZodString;
|
|
117
|
+
kind: z__default.ZodString;
|
|
118
|
+
schemaVersion: z__default.ZodNumber;
|
|
119
|
+
value: z__default.ZodNumber;
|
|
120
|
+
responseId: z__default.ZodString;
|
|
121
|
+
explanation: z__default.ZodOptional<z__default.ZodString>;
|
|
122
|
+
metadata: z__default.ZodOptional<z__default.ZodRecord<z__default.ZodString, z__default.ZodUnknown>>;
|
|
123
|
+
scoringMethod: z__default.ZodEnum<{
|
|
124
|
+
readonly ai: "ai";
|
|
125
|
+
readonly human: "human";
|
|
126
|
+
readonly algo: "algo";
|
|
127
|
+
}>;
|
|
128
|
+
}, z__default.core.$strip>;
|
|
129
|
+
type BaseScoreV1 = z__default.infer<typeof BaseScoreSchemaV1>;
|
|
130
|
+
declare const defineScoreSchema: <TBaseSchema extends z__default.ZodObject<Omit<{
|
|
131
|
+
id: z__default.ZodString;
|
|
132
|
+
kind: z__default.ZodString;
|
|
133
|
+
schemaVersion: z__default.ZodNumber;
|
|
134
|
+
value: z__default.ZodNumber;
|
|
135
|
+
responseId: z__default.ZodString;
|
|
136
|
+
explanation: z__default.ZodOptional<z__default.ZodString>;
|
|
137
|
+
metadata: z__default.ZodOptional<z__default.ZodRecord<z__default.ZodString, z__default.ZodUnknown>>;
|
|
138
|
+
scoringMethod: z__default.ZodEnum<{
|
|
139
|
+
readonly ai: "ai";
|
|
140
|
+
readonly human: "human";
|
|
141
|
+
readonly algo: "algo";
|
|
142
|
+
}>;
|
|
143
|
+
}, "kind" | "schemaVersion"> & {
|
|
144
|
+
kind: z__default.ZodString | z__default.ZodLiteral<string>;
|
|
145
|
+
schemaVersion: z__default.ZodNumber | z__default.ZodLiteral<number>;
|
|
146
|
+
}, z__default.core.$strip>, TKind extends string | undefined = undefined, TSchemaVersion extends number | undefined = undefined, TFields extends Readonly<{
|
|
147
|
+
[k: string]: z__default.core.$ZodType<unknown, unknown, z__default.core.$ZodTypeInternals<unknown, unknown>>;
|
|
148
|
+
}> = {}>(config: {
|
|
149
|
+
baseSchema: TBaseSchema;
|
|
150
|
+
kind?: TKind | undefined;
|
|
151
|
+
schemaVersion?: TSchemaVersion | undefined;
|
|
152
|
+
fields?: TFields | undefined;
|
|
153
|
+
}) => (TBaseSchema extends z__default.ZodObject<infer U extends Readonly<{
|
|
154
|
+
[k: string]: z__default.core.$ZodType<unknown, unknown, z__default.core.$ZodTypeInternals<unknown, unknown>>;
|
|
155
|
+
}>, z__default.core.$strip> ? z__default.ZodObject<Omit<U, "kind" | "schemaVersion"> & TFields & {
|
|
156
|
+
kind: TKind extends undefined ? U["kind"] : z__default.ZodLiteral<TKind>;
|
|
157
|
+
schemaVersion: TSchemaVersion extends undefined ? U["schemaVersion"] : z__default.ZodLiteral<TSchemaVersion>;
|
|
158
|
+
}, z__default.core.$strip> : never) & {
|
|
159
|
+
new: (input: Omit<z__default.core.output<TBaseSchema extends z__default.ZodObject<infer U extends Readonly<{
|
|
160
|
+
[k: string]: z__default.core.$ZodType<unknown, unknown, z__default.core.$ZodTypeInternals<unknown, unknown>>;
|
|
161
|
+
}>, z__default.core.$strip> ? z__default.ZodObject<Omit<U, "kind" | "schemaVersion"> & TFields & {
|
|
162
|
+
kind: TKind extends undefined ? U["kind"] : z__default.ZodLiteral<TKind>;
|
|
163
|
+
schemaVersion: TSchemaVersion extends undefined ? U["schemaVersion"] : z__default.ZodLiteral<TSchemaVersion>;
|
|
164
|
+
}, z__default.core.$strip> : never>, "kind" | "schemaVersion">) => z__default.core.output<TBaseSchema extends z__default.ZodObject<infer U extends Readonly<{
|
|
165
|
+
[k: string]: z__default.core.$ZodType<unknown, unknown, z__default.core.$ZodTypeInternals<unknown, unknown>>;
|
|
166
|
+
}>, z__default.core.$strip> ? z__default.ZodObject<Omit<U, "kind" | "schemaVersion"> & TFields & {
|
|
167
|
+
kind: TKind extends undefined ? U["kind"] : z__default.ZodLiteral<TKind>;
|
|
168
|
+
schemaVersion: TSchemaVersion extends undefined ? U["schemaVersion"] : z__default.ZodLiteral<TSchemaVersion>;
|
|
169
|
+
}, z__default.core.$strip> : never>;
|
|
170
|
+
newWithId(input: Omit<z__default.core.output<TBaseSchema extends z__default.ZodObject<infer U extends Readonly<{
|
|
171
|
+
[k: string]: z__default.core.$ZodType<unknown, unknown, z__default.core.$ZodTypeInternals<unknown, unknown>>;
|
|
172
|
+
}>, z__default.core.$strip> ? z__default.ZodObject<Omit<U, "kind" | "schemaVersion"> & TFields & {
|
|
173
|
+
kind: TKind extends undefined ? U["kind"] : z__default.ZodLiteral<TKind>;
|
|
174
|
+
schemaVersion: TSchemaVersion extends undefined ? U["schemaVersion"] : z__default.ZodLiteral<TSchemaVersion>;
|
|
175
|
+
}, z__default.core.$strip> : never>, "kind" | "schemaVersion" | "id">, generator: IdGenerator): Promise<z__default.core.output<TBaseSchema extends z__default.ZodObject<infer U extends Readonly<{
|
|
176
|
+
[k: string]: z__default.core.$ZodType<unknown, unknown, z__default.core.$ZodTypeInternals<unknown, unknown>>;
|
|
177
|
+
}>, z__default.core.$strip> ? z__default.ZodObject<Omit<U, "kind" | "schemaVersion"> & TFields & {
|
|
178
|
+
kind: TKind extends undefined ? U["kind"] : z__default.ZodLiteral<TKind>;
|
|
179
|
+
schemaVersion: TSchemaVersion extends undefined ? U["schemaVersion"] : z__default.ZodLiteral<TSchemaVersion>;
|
|
180
|
+
}, z__default.core.$strip> : never>>;
|
|
181
|
+
};
|
|
182
|
+
|
|
183
|
+
declare const BaseBenchmarkSpecSchemaV1: z__default.ZodObject<{
|
|
184
|
+
kind: z__default.ZodString;
|
|
185
|
+
schemaVersion: z__default.ZodNumber;
|
|
186
|
+
metadata: z__default.ZodOptional<z__default.ZodRecord<z__default.ZodString, z__default.ZodUnknown>>;
|
|
187
|
+
}, z__default.core.$strip>;
|
|
188
|
+
type BaseBenchmarkSpecV1 = z__default.infer<typeof BaseBenchmarkSpecSchemaV1>;
|
|
189
|
+
declare const defineBenchmarkSpecSchema: <TBaseSchema extends z__default.ZodObject<Omit<{
|
|
190
|
+
kind: z__default.ZodString;
|
|
191
|
+
schemaVersion: z__default.ZodNumber;
|
|
192
|
+
metadata: z__default.ZodOptional<z__default.ZodRecord<z__default.ZodString, z__default.ZodUnknown>>;
|
|
193
|
+
}, "kind" | "schemaVersion"> & {
|
|
194
|
+
kind: z__default.ZodString | z__default.ZodLiteral<string>;
|
|
195
|
+
schemaVersion: z__default.ZodNumber | z__default.ZodLiteral<number>;
|
|
196
|
+
}, z__default.core.$strip>, TKind extends string | undefined = undefined, TSchemaVersion extends number | undefined = undefined, TFields extends Readonly<{
|
|
197
|
+
[k: string]: z__default.core.$ZodType<unknown, unknown, z__default.core.$ZodTypeInternals<unknown, unknown>>;
|
|
198
|
+
}> = {}>(config: {
|
|
199
|
+
baseSchema: TBaseSchema;
|
|
200
|
+
kind?: TKind | undefined;
|
|
201
|
+
schemaVersion?: TSchemaVersion | undefined;
|
|
202
|
+
fields?: TFields | undefined;
|
|
203
|
+
}) => (TBaseSchema extends z__default.ZodObject<infer U extends Readonly<{
|
|
204
|
+
[k: string]: z__default.core.$ZodType<unknown, unknown, z__default.core.$ZodTypeInternals<unknown, unknown>>;
|
|
205
|
+
}>, z__default.core.$strip> ? z__default.ZodObject<Omit<U, "kind" | "schemaVersion"> & TFields & {
|
|
206
|
+
kind: TKind extends undefined ? U["kind"] : z__default.ZodLiteral<TKind>;
|
|
207
|
+
schemaVersion: TSchemaVersion extends undefined ? U["schemaVersion"] : z__default.ZodLiteral<TSchemaVersion>;
|
|
208
|
+
}, z__default.core.$strip> : never) & {
|
|
209
|
+
new: (input: Omit<z__default.core.output<TBaseSchema extends z__default.ZodObject<infer U extends Readonly<{
|
|
210
|
+
[k: string]: z__default.core.$ZodType<unknown, unknown, z__default.core.$ZodTypeInternals<unknown, unknown>>;
|
|
211
|
+
}>, z__default.core.$strip> ? z__default.ZodObject<Omit<U, "kind" | "schemaVersion"> & TFields & {
|
|
212
|
+
kind: TKind extends undefined ? U["kind"] : z__default.ZodLiteral<TKind>;
|
|
213
|
+
schemaVersion: TSchemaVersion extends undefined ? U["schemaVersion"] : z__default.ZodLiteral<TSchemaVersion>;
|
|
214
|
+
}, z__default.core.$strip> : never>, "kind" | "schemaVersion">) => z__default.core.output<TBaseSchema extends z__default.ZodObject<infer U extends Readonly<{
|
|
215
|
+
[k: string]: z__default.core.$ZodType<unknown, unknown, z__default.core.$ZodTypeInternals<unknown, unknown>>;
|
|
216
|
+
}>, z__default.core.$strip> ? z__default.ZodObject<Omit<U, "kind" | "schemaVersion"> & TFields & {
|
|
217
|
+
kind: TKind extends undefined ? U["kind"] : z__default.ZodLiteral<TKind>;
|
|
218
|
+
schemaVersion: TSchemaVersion extends undefined ? U["schemaVersion"] : z__default.ZodLiteral<TSchemaVersion>;
|
|
219
|
+
}, z__default.core.$strip> : never>;
|
|
220
|
+
newWithId(input: Omit<z__default.core.output<TBaseSchema extends z__default.ZodObject<infer U extends Readonly<{
|
|
221
|
+
[k: string]: z__default.core.$ZodType<unknown, unknown, z__default.core.$ZodTypeInternals<unknown, unknown>>;
|
|
222
|
+
}>, z__default.core.$strip> ? z__default.ZodObject<Omit<U, "kind" | "schemaVersion"> & TFields & {
|
|
223
|
+
kind: TKind extends undefined ? U["kind"] : z__default.ZodLiteral<TKind>;
|
|
224
|
+
schemaVersion: TSchemaVersion extends undefined ? U["schemaVersion"] : z__default.ZodLiteral<TSchemaVersion>;
|
|
225
|
+
}, z__default.core.$strip> : never>, "kind" | "schemaVersion" | "id">, generator: IdGenerator): Promise<z__default.core.output<TBaseSchema extends z__default.ZodObject<infer U extends Readonly<{
|
|
226
|
+
[k: string]: z__default.core.$ZodType<unknown, unknown, z__default.core.$ZodTypeInternals<unknown, unknown>>;
|
|
227
|
+
}>, z__default.core.$strip> ? z__default.ZodObject<Omit<U, "kind" | "schemaVersion"> & TFields & {
|
|
228
|
+
kind: TKind extends undefined ? U["kind"] : z__default.ZodLiteral<TKind>;
|
|
229
|
+
schemaVersion: TSchemaVersion extends undefined ? U["schemaVersion"] : z__default.ZodLiteral<TSchemaVersion>;
|
|
230
|
+
}, z__default.core.$strip> : never>>;
|
|
231
|
+
};
|
|
232
|
+
|
|
233
|
+
export { BaseBenchmarkSpecSchemaV1, type BaseBenchmarkSpecV1, BaseResponseSchemaV1, type BaseResponseV1, BaseScoreSchemaV1, type BaseScoreV1, BaseTestCaseSchemaV1, type BaseTestCaseV1, defineBenchmarkSpecSchema, defineResponseSchema, defineScoreSchema, defineTestCaseSchema };
|
|
@@ -0,0 +1,27 @@
|
|
|
1
|
+
import {
|
|
2
|
+
BaseBenchmarkSpecSchemaV1,
|
|
3
|
+
BaseScoreSchemaV1,
|
|
4
|
+
BaseTestCaseSchemaV1,
|
|
5
|
+
defineBenchmarkSpecSchema,
|
|
6
|
+
defineScoreSchema,
|
|
7
|
+
defineTestCaseSchema
|
|
8
|
+
} from "../chunk-232PY7K3.js";
|
|
9
|
+
import "../chunk-ZJWSK4VO.js";
|
|
10
|
+
import {
|
|
11
|
+
BaseResponseSchemaV1,
|
|
12
|
+
IdSchema,
|
|
13
|
+
defineResponseSchema
|
|
14
|
+
} from "../chunk-IUN2IUCS.js";
|
|
15
|
+
import "../chunk-PZ5AY32C.js";
|
|
16
|
+
export {
|
|
17
|
+
BaseBenchmarkSpecSchemaV1,
|
|
18
|
+
BaseResponseSchemaV1,
|
|
19
|
+
BaseScoreSchemaV1,
|
|
20
|
+
BaseTestCaseSchemaV1,
|
|
21
|
+
IdSchema,
|
|
22
|
+
defineBenchmarkSpecSchema,
|
|
23
|
+
defineResponseSchema,
|
|
24
|
+
defineScoreSchema,
|
|
25
|
+
defineTestCaseSchema
|
|
26
|
+
};
|
|
27
|
+
//# sourceMappingURL=index.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"sources":[],"sourcesContent":[],"mappings":"","names":[]}
|
|
@@ -0,0 +1,98 @@
|
|
|
1
|
+
export { B as BaseSystemPromptSchemaV1, a as BaseSystemPromptV1, d as defineSystemPromptSchema } from '../../system-prompt--0FdPWqK.js';
|
|
2
|
+
export { S as SimpleSystemPromptSchemaV1, a as SimpleSystemPromptV1 } from '../../simple-system-prompt-CzPYuvo0.js';
|
|
3
|
+
import { I as IdGenerator } from '../../index-WiPjF2AL.js';
|
|
4
|
+
import z__default from 'zod';
|
|
5
|
+
|
|
6
|
+
/**
|
|
7
|
+
* Base response schema specialized for LLM chat responses.
|
|
8
|
+
*/
|
|
9
|
+
declare const BaseLLMChatResponseSchemaV1: z__default.ZodObject<Omit<{
|
|
10
|
+
id: z__default.ZodString;
|
|
11
|
+
kind: z__default.ZodString;
|
|
12
|
+
schemaVersion: z__default.ZodNumber;
|
|
13
|
+
startedAt: z__default.ZodNumber;
|
|
14
|
+
completedAt: z__default.ZodNumber;
|
|
15
|
+
testCaseId: z__default.ZodString;
|
|
16
|
+
metadata: z__default.ZodOptional<z__default.ZodRecord<z__default.ZodString, z__default.ZodUnknown>>;
|
|
17
|
+
}, "kind" | "schemaVersion"> & {
|
|
18
|
+
data: z__default.ZodString;
|
|
19
|
+
modelSlug: z__default.ZodString;
|
|
20
|
+
provider: z__default.ZodString;
|
|
21
|
+
systemPromptId: z__default.ZodOptional<z__default.ZodString>;
|
|
22
|
+
inputTokensUsed: z__default.ZodOptional<z__default.ZodNumber>;
|
|
23
|
+
outputTokensUsed: z__default.ZodOptional<z__default.ZodNumber>;
|
|
24
|
+
inputCost: z__default.ZodOptional<z__default.ZodString>;
|
|
25
|
+
outputCost: z__default.ZodOptional<z__default.ZodString>;
|
|
26
|
+
} & {
|
|
27
|
+
kind: z__default.ZodString;
|
|
28
|
+
schemaVersion: z__default.ZodNumber;
|
|
29
|
+
}, z__default.core.$strip> & {
|
|
30
|
+
new: (input: Omit<{
|
|
31
|
+
id: string;
|
|
32
|
+
testCaseId: string;
|
|
33
|
+
startedAt: number;
|
|
34
|
+
completedAt: number;
|
|
35
|
+
data: string;
|
|
36
|
+
modelSlug: string;
|
|
37
|
+
provider: string;
|
|
38
|
+
kind: string;
|
|
39
|
+
schemaVersion: number;
|
|
40
|
+
metadata?: Record<string, unknown> | undefined;
|
|
41
|
+
systemPromptId?: string | undefined;
|
|
42
|
+
inputTokensUsed?: number | undefined;
|
|
43
|
+
outputTokensUsed?: number | undefined;
|
|
44
|
+
inputCost?: string | undefined;
|
|
45
|
+
outputCost?: string | undefined;
|
|
46
|
+
}, "kind" | "schemaVersion">) => {
|
|
47
|
+
id: string;
|
|
48
|
+
testCaseId: string;
|
|
49
|
+
startedAt: number;
|
|
50
|
+
completedAt: number;
|
|
51
|
+
data: string;
|
|
52
|
+
modelSlug: string;
|
|
53
|
+
provider: string;
|
|
54
|
+
kind: string;
|
|
55
|
+
schemaVersion: number;
|
|
56
|
+
metadata?: Record<string, unknown> | undefined;
|
|
57
|
+
systemPromptId?: string | undefined;
|
|
58
|
+
inputTokensUsed?: number | undefined;
|
|
59
|
+
outputTokensUsed?: number | undefined;
|
|
60
|
+
inputCost?: string | undefined;
|
|
61
|
+
outputCost?: string | undefined;
|
|
62
|
+
};
|
|
63
|
+
newWithId(input: Omit<{
|
|
64
|
+
id: string;
|
|
65
|
+
testCaseId: string;
|
|
66
|
+
startedAt: number;
|
|
67
|
+
completedAt: number;
|
|
68
|
+
data: string;
|
|
69
|
+
modelSlug: string;
|
|
70
|
+
provider: string;
|
|
71
|
+
kind: string;
|
|
72
|
+
schemaVersion: number;
|
|
73
|
+
metadata?: Record<string, unknown> | undefined;
|
|
74
|
+
systemPromptId?: string | undefined;
|
|
75
|
+
inputTokensUsed?: number | undefined;
|
|
76
|
+
outputTokensUsed?: number | undefined;
|
|
77
|
+
inputCost?: string | undefined;
|
|
78
|
+
outputCost?: string | undefined;
|
|
79
|
+
}, "id" | "kind" | "schemaVersion">, generator: IdGenerator): Promise<{
|
|
80
|
+
id: string;
|
|
81
|
+
testCaseId: string;
|
|
82
|
+
startedAt: number;
|
|
83
|
+
completedAt: number;
|
|
84
|
+
data: string;
|
|
85
|
+
modelSlug: string;
|
|
86
|
+
provider: string;
|
|
87
|
+
kind: string;
|
|
88
|
+
schemaVersion: number;
|
|
89
|
+
metadata?: Record<string, unknown> | undefined;
|
|
90
|
+
systemPromptId?: string | undefined;
|
|
91
|
+
inputTokensUsed?: number | undefined;
|
|
92
|
+
outputTokensUsed?: number | undefined;
|
|
93
|
+
inputCost?: string | undefined;
|
|
94
|
+
outputCost?: string | undefined;
|
|
95
|
+
}>;
|
|
96
|
+
};
|
|
97
|
+
|
|
98
|
+
export { BaseLLMChatResponseSchemaV1 };
|
|
@@ -0,0 +1,37 @@
|
|
|
1
|
+
import {
|
|
2
|
+
BaseLLMChatResponseSchemaV1
|
|
3
|
+
} from "../../chunk-7TREBPSJ.js";
|
|
4
|
+
import {
|
|
5
|
+
IdSchema,
|
|
6
|
+
buildSchemaDefiner
|
|
7
|
+
} from "../../chunk-IUN2IUCS.js";
|
|
8
|
+
import "../../chunk-PZ5AY32C.js";
|
|
9
|
+
|
|
10
|
+
// src/schemas/llm/system-prompt.ts
|
|
11
|
+
import { z } from "zod";
|
|
12
|
+
var BaseSystemPromptSchemaV1 = z.object({
|
|
13
|
+
id: IdSchema,
|
|
14
|
+
kind: z.string(),
|
|
15
|
+
schemaVersion: z.number(),
|
|
16
|
+
version: z.number(),
|
|
17
|
+
metadata: z.record(z.string(), z.unknown()).optional()
|
|
18
|
+
});
|
|
19
|
+
var defineSystemPromptSchema = buildSchemaDefiner();
|
|
20
|
+
|
|
21
|
+
// src/schemas/llm/simple-system-prompt.ts
|
|
22
|
+
import { z as z2 } from "zod";
|
|
23
|
+
var SimpleSystemPromptSchemaV1 = defineSystemPromptSchema({
|
|
24
|
+
baseSchema: BaseSystemPromptSchemaV1,
|
|
25
|
+
kind: "sys-prompt.simple",
|
|
26
|
+
schemaVersion: 1,
|
|
27
|
+
fields: {
|
|
28
|
+
content: z2.string()
|
|
29
|
+
}
|
|
30
|
+
});
|
|
31
|
+
export {
|
|
32
|
+
BaseLLMChatResponseSchemaV1,
|
|
33
|
+
BaseSystemPromptSchemaV1,
|
|
34
|
+
SimpleSystemPromptSchemaV1,
|
|
35
|
+
defineSystemPromptSchema
|
|
36
|
+
};
|
|
37
|
+
//# sourceMappingURL=index.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"sources":["../../../src/schemas/llm/system-prompt.ts","../../../src/schemas/llm/simple-system-prompt.ts"],"sourcesContent":["import { IdSchema } from \"../id\";\nimport { z } from \"zod\";\nimport { buildSchemaDefiner } from \"../schema-definer\";\n\nexport const BaseSystemPromptSchemaV1 = z.object({\n id: IdSchema,\n kind: z.string(),\n schemaVersion: z.number(),\n version: z.number(),\n metadata: z.record(z.string(), z.unknown()).optional(),\n});\nexport type BaseSystemPromptV1 = z.infer<typeof BaseSystemPromptSchemaV1>;\n\nexport const defineSystemPromptSchema =\n buildSchemaDefiner<typeof BaseSystemPromptSchemaV1.shape>();\n","import { z } from \"zod\";\nimport { defineSystemPromptSchema } from \"./system-prompt\";\nimport { BaseSystemPromptSchemaV1 } from \"./system-prompt\";\n\nexport const SimpleSystemPromptSchemaV1 = defineSystemPromptSchema({\n baseSchema: BaseSystemPromptSchemaV1,\n kind: \"sys-prompt.simple\",\n schemaVersion: 1,\n fields: {\n content: z.string(),\n },\n});\nexport type SimpleSystemPromptV1 = z.infer<typeof SimpleSystemPromptSchemaV1>;\n"],"mappings":";;;;;;;;;;AACA,SAAS,SAAS;AAGX,IAAM,2BAA2B,EAAE,OAAO;AAAA,EAC/C,IAAI;AAAA,EACJ,MAAM,EAAE,OAAO;AAAA,EACf,eAAe,EAAE,OAAO;AAAA,EACxB,SAAS,EAAE,OAAO;AAAA,EAClB,UAAU,EAAE,OAAO,EAAE,OAAO,GAAG,EAAE,QAAQ,CAAC,EAAE,SAAS;AACvD,CAAC;AAGM,IAAM,2BACX,mBAA0D;;;ACd5D,SAAS,KAAAA,UAAS;AAIX,IAAM,6BAA6B,yBAAyB;AAAA,EACjE,YAAY;AAAA,EACZ,MAAM;AAAA,EACN,eAAe;AAAA,EACf,QAAQ;AAAA,IACN,SAASC,GAAE,OAAO;AAAA,EACpB;AACF,CAAC;","names":["z","z"]}
|
|
@@ -0,0 +1,63 @@
|
|
|
1
|
+
import { A as AbstractScorer } from '../abstract-Dec9Sc5O.js';
|
|
2
|
+
export { B as BaseScorerResult } from '../abstract-Dec9Sc5O.js';
|
|
3
|
+
export { b as LLMJudgeCriterion, L as LLMJudgeScorer, d as LLMJudgeScorerParams, c as LLMJudgeScorerResult, a as MCQScorer, M as MCQScorerParams } from '../llm-judge-DIG1f1Az.js';
|
|
4
|
+
import '../llm-DNj_tp2T.js';
|
|
5
|
+
import '../provider-BDjGp2y-.js';
|
|
6
|
+
import 'openai/resources/shared';
|
|
7
|
+
import 'openai/resources/chat/completions';
|
|
8
|
+
import '../rate-limiter-CSmVIRsM.js';
|
|
9
|
+
|
|
10
|
+
type RegexPattern = {
|
|
11
|
+
/**
|
|
12
|
+
* The regex pattern to match against the response
|
|
13
|
+
*/
|
|
14
|
+
regex: RegExp;
|
|
15
|
+
/**
|
|
16
|
+
* The index of the capture group to extract (1-based, like match[1])
|
|
17
|
+
* If not provided, defaults to 1 (first capture group)
|
|
18
|
+
*/
|
|
19
|
+
captureGroupIndex?: number;
|
|
20
|
+
/**
|
|
21
|
+
* Optional function to transform the extracted value before validation
|
|
22
|
+
*/
|
|
23
|
+
transform?: (value: string) => string | undefined;
|
|
24
|
+
};
|
|
25
|
+
type RegexScorerParams = {
|
|
26
|
+
/**
|
|
27
|
+
* The input text to score
|
|
28
|
+
*/
|
|
29
|
+
input: string;
|
|
30
|
+
/**
|
|
31
|
+
* Array of regex patterns to try (in order, first match wins)
|
|
32
|
+
*/
|
|
33
|
+
patterns: RegexPattern[];
|
|
34
|
+
/**
|
|
35
|
+
* Expected value(s) to match against. Can be a record of expected values for named groups, or a validation function
|
|
36
|
+
*/
|
|
37
|
+
expectedValue: Record<string, string> | ((groupName: string, match: string) => boolean);
|
|
38
|
+
/**
|
|
39
|
+
* Optional: Which match to use when multiple matches are found
|
|
40
|
+
* Defaults to "last" (uses the last match found)
|
|
41
|
+
*/
|
|
42
|
+
matchPreference?: "first" | "last";
|
|
43
|
+
/**
|
|
44
|
+
* Optional: If true, allows partial scoring based on how many groups match
|
|
45
|
+
* For example, if 2 groups are expected and only 1 matches, score would be 0.5
|
|
46
|
+
* Defaults to false (all-or-nothing scoring)
|
|
47
|
+
*/
|
|
48
|
+
allowPartialScoring?: boolean;
|
|
49
|
+
};
|
|
50
|
+
/**
|
|
51
|
+
* Generic Regex scorer. It scores the given input against a set of regex patterns.
|
|
52
|
+
*/
|
|
53
|
+
declare class RegexScorer extends AbstractScorer {
|
|
54
|
+
readonly kind = "regex";
|
|
55
|
+
score(params: RegexScorerParams): Promise<{
|
|
56
|
+
value: number;
|
|
57
|
+
extractedAnswers: {
|
|
58
|
+
[k: string]: string;
|
|
59
|
+
};
|
|
60
|
+
}>;
|
|
61
|
+
}
|
|
62
|
+
|
|
63
|
+
export { AbstractScorer, type RegexPattern, RegexScorer, type RegexScorerParams };
|