@arizeai/phoenix-evals 2.4.0 → 2.5.0
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/CHANGELOG.md +256 -0
- package/README.md +6 -5
- package/dist/esm/__generated__/default_templates/COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG.d.ts +3 -0
- package/dist/esm/__generated__/default_templates/COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG.d.ts.map +1 -0
- package/dist/esm/__generated__/default_templates/COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG.js +93 -0
- package/dist/esm/__generated__/default_templates/COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG.js.map +1 -0
- package/dist/esm/__generated__/default_templates/index.d.ts +1 -0
- package/dist/esm/__generated__/default_templates/index.d.ts.map +1 -1
- package/dist/esm/__generated__/default_templates/index.js +1 -0
- package/dist/esm/__generated__/default_templates/index.js.map +1 -1
- package/dist/esm/llm/createCompletenessEvaluator.d.ts +46 -0
- package/dist/esm/llm/createCompletenessEvaluator.d.ts.map +1 -0
- package/dist/esm/llm/createCompletenessEvaluator.js +38 -0
- package/dist/esm/llm/createCompletenessEvaluator.js.map +1 -0
- package/dist/esm/llm/createDocumentRelevanceEvaluator.d.ts +16 -11
- package/dist/esm/llm/createDocumentRelevanceEvaluator.d.ts.map +1 -1
- package/dist/esm/llm/createDocumentRelevanceEvaluator.js +7 -11
- package/dist/esm/llm/createDocumentRelevanceEvaluator.js.map +1 -1
- package/dist/esm/llm/createRetrievalRelevanceEvaluator.d.ts +4 -5
- package/dist/esm/llm/createRetrievalRelevanceEvaluator.d.ts.map +1 -1
- package/dist/esm/llm/createRetrievalRelevanceEvaluator.js +4 -5
- package/dist/esm/llm/createRetrievalRelevanceEvaluator.js.map +1 -1
- package/dist/esm/llm/index.d.ts +1 -0
- package/dist/esm/llm/index.d.ts.map +1 -1
- package/dist/esm/llm/index.js +1 -0
- package/dist/esm/llm/index.js.map +1 -1
- package/dist/esm/tsconfig.esm.tsbuildinfo +1 -1
- package/dist/esm/utils/bindEvaluator.d.ts +4 -5
- package/dist/esm/utils/bindEvaluator.d.ts.map +1 -1
- package/dist/esm/utils/bindEvaluator.js +4 -5
- package/dist/esm/utils/bindEvaluator.js.map +1 -1
- package/dist/src/__generated__/default_templates/COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG.d.ts +3 -0
- package/dist/src/__generated__/default_templates/COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG.d.ts.map +1 -0
- package/dist/src/__generated__/default_templates/COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG.js +96 -0
- package/dist/src/__generated__/default_templates/COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG.js.map +1 -0
- package/dist/src/__generated__/default_templates/index.d.ts +1 -0
- package/dist/src/__generated__/default_templates/index.d.ts.map +1 -1
- package/dist/src/__generated__/default_templates/index.js +3 -1
- package/dist/src/__generated__/default_templates/index.js.map +1 -1
- package/dist/src/llm/createCompletenessEvaluator.d.ts +46 -0
- package/dist/src/llm/createCompletenessEvaluator.d.ts.map +1 -0
- package/dist/src/llm/createCompletenessEvaluator.js +49 -0
- package/dist/src/llm/createCompletenessEvaluator.js.map +1 -0
- package/dist/src/llm/createDocumentRelevanceEvaluator.d.ts +16 -11
- package/dist/src/llm/createDocumentRelevanceEvaluator.d.ts.map +1 -1
- package/dist/src/llm/createDocumentRelevanceEvaluator.js +7 -11
- package/dist/src/llm/createDocumentRelevanceEvaluator.js.map +1 -1
- package/dist/src/llm/createRetrievalRelevanceEvaluator.d.ts +4 -5
- package/dist/src/llm/createRetrievalRelevanceEvaluator.d.ts.map +1 -1
- package/dist/src/llm/createRetrievalRelevanceEvaluator.js +4 -5
- package/dist/src/llm/createRetrievalRelevanceEvaluator.js.map +1 -1
- package/dist/src/llm/index.d.ts +1 -0
- package/dist/src/llm/index.d.ts.map +1 -1
- package/dist/src/llm/index.js +1 -0
- package/dist/src/llm/index.js.map +1 -1
- package/dist/src/utils/bindEvaluator.d.ts +4 -5
- package/dist/src/utils/bindEvaluator.d.ts.map +1 -1
- package/dist/src/utils/bindEvaluator.js +4 -5
- package/dist/src/utils/bindEvaluator.js.map +1 -1
- package/dist/tsconfig.tsbuildinfo +1 -1
- package/package.json +12 -12
- package/src/__generated__/default_templates/COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG.ts +95 -0
- package/src/__generated__/default_templates/index.ts +1 -0
- package/src/llm/createCompletenessEvaluator.ts +74 -0
- package/src/llm/createDocumentRelevanceEvaluator.ts +16 -11
- package/src/llm/createRetrievalRelevanceEvaluator.ts +4 -5
- package/src/llm/index.ts +1 -0
- package/src/utils/bindEvaluator.ts +4 -5
package/CHANGELOG.md
ADDED
|
@@ -0,0 +1,256 @@
|
|
|
1
|
+
# @arizeai/phoenix-evals
|
|
2
|
+
|
|
3
|
+
## 2.5.0
|
|
4
|
+
|
|
5
|
+
### Minor Changes
|
|
6
|
+
|
|
7
|
+
- 63805d0: Add `createCompletenessEvaluator` to judge whether every active user request in a conversation was actually completed.
|
|
8
|
+
|
|
9
|
+
### Patch Changes
|
|
10
|
+
|
|
11
|
+
- d67ea3f: Deprecate `createDocumentRelevanceEvaluator` in favor of
|
|
12
|
+
`createRetrievalRelevanceEvaluator`. Rename `documentText` to `context` and the
|
|
13
|
+
`unrelated` label to `irrelevant`. The deprecated factory and its types will be
|
|
14
|
+
removed in the next major release.
|
|
15
|
+
|
|
16
|
+
## 2.4.0
|
|
17
|
+
|
|
18
|
+
### Minor Changes
|
|
19
|
+
|
|
20
|
+
- 0e3b2c2: Add a built-in PII detection evaluator (`createPiiDetectionEvaluator`) that screens a conversation record for personally identifiable information, including system instructions, tool calls and results, and retrieved content the end user may never have seen.
|
|
21
|
+
|
|
22
|
+
## 2.3.0
|
|
23
|
+
|
|
24
|
+
### Minor Changes
|
|
25
|
+
|
|
26
|
+
- 7efa7b2: Add a built-in retrieval relevance evaluator (createRetrievalRelevanceEvaluator) that checks whether the external information retrieved during a step is relevant to the request it was serving. Unlike document relevance, it is source-agnostic and scores the retrieved information holistically, whether it came from a vector search, a tool or MCP call, a web search, or content embedded in an LLM turn.
|
|
27
|
+
|
|
28
|
+
## 2.2.0
|
|
29
|
+
|
|
30
|
+
### Minor Changes
|
|
31
|
+
|
|
32
|
+
- 1aa1a84: Redesign `createHallucinationEvaluator` as a conversation-grounding evaluator.
|
|
33
|
+
|
|
34
|
+
This changes the evaluator input to `input` and `output`, where `input` is the conversation available to the assistant and `output` is the response being judged. A separate `context` field is no longer accepted.
|
|
35
|
+
|
|
36
|
+
Labels change from `factual`/`hallucinated` to `grounded`/`hallucinated`. Scores are minimized, with `hallucinated` equal to `1` and `grounded` equal to `0`. Existing stored evaluations, dashboards, thresholds, and label filters may require migration and should not be compared directly with results from the previous evaluator.
|
|
37
|
+
|
|
38
|
+
## 2.1.0
|
|
39
|
+
|
|
40
|
+
### Minor Changes
|
|
41
|
+
|
|
42
|
+
- 899dece: Add a built-in toxicity evaluator (createToxicityEvaluator) for classifying text as toxic or non-toxic.
|
|
43
|
+
|
|
44
|
+
## 2.0.0
|
|
45
|
+
|
|
46
|
+
### Major Changes
|
|
47
|
+
|
|
48
|
+
- 5a8f3f7: Upgrade `@arizeai/phoenix-evals` to AI SDK v7. Evaluator telemetry uses the AI SDK v7 telemetry API with a per-call `OpenTelemetry` integration from `@ai-sdk/otel`, while preserving all globally registered integrations (logging, metrics, and tracing to other backends); a Phoenix tracing integration is appended unless a global integration already traces with the same tracer. Evaluator spans now follow the OpenTelemetry `gen_ai.*` conventions emitted by `@ai-sdk/otel` instead of the AI SDK v6 `ai.*` span format. The `telemetry.tracer` and `telemetry.isEnabled` options keep working as before, and system messages in prompt templates continue to be supported. Requires Node.js >=22.12 and AI SDK v7-compatible model providers (e.g. `@ai-sdk/openai` v4).
|
|
49
|
+
|
|
50
|
+
## 1.2.0
|
|
51
|
+
|
|
52
|
+
### Minor Changes
|
|
53
|
+
|
|
54
|
+
- f94067b: Add px setup script for agent onboarding
|
|
55
|
+
|
|
56
|
+
## 1.1.0
|
|
57
|
+
|
|
58
|
+
### Minor Changes
|
|
59
|
+
|
|
60
|
+
- 1f7d5f0: Add built-in code (non-LLM) evaluators for classification metrics: `createPrecisionEvaluator`, `createRecallEvaluator`, `createF1Evaluator`, `createFBetaEvaluator`, and the bundling helper `createPrecisionRecallFScoreEvaluators`, plus the underlying `computePrecisionRecallFScore` function. These support binary (via `positiveLabel`) and multi-class (`macro`/`micro`/`weighted` averaging) classification, mirroring the Python `PrecisionRecallFScore` evaluator.
|
|
61
|
+
|
|
62
|
+
## 1.0.3
|
|
63
|
+
|
|
64
|
+
### Patch Changes
|
|
65
|
+
|
|
66
|
+
- 5d14f23: Add LinkedIn link to the Community section of the README.
|
|
67
|
+
|
|
68
|
+
## 1.0.2
|
|
69
|
+
|
|
70
|
+
### Patch Changes
|
|
71
|
+
|
|
72
|
+
- 1028be5: Bundle curated package docs and examples into npm packages under `docs/`.
|
|
73
|
+
|
|
74
|
+
## 1.0.1
|
|
75
|
+
|
|
76
|
+
### Patch Changes
|
|
77
|
+
|
|
78
|
+
- b4ded15: Update package READMEs with latest capabilities
|
|
79
|
+
- phoenix-client: Add Sessions section documenting `listSessions`, `getSession`, and `addSessionAnnotation`
|
|
80
|
+
- phoenix-evals: Add full pre-built evaluators table (conciseness, correctness, document relevance, refusal, tool evaluators), fix import paths for `bindEvaluator`
|
|
81
|
+
- phoenix-mcp: Expand Tool Coverage section with complete, accurate tool list
|
|
82
|
+
- phoenix-config: Full documentation replacing the one-line placeholder (installation, environment variables, usage examples)
|
|
83
|
+
- phoenix-cli: Add `px annotation-config` command documentation
|
|
84
|
+
|
|
85
|
+
## 1.0.0
|
|
86
|
+
|
|
87
|
+
### Major Changes
|
|
88
|
+
|
|
89
|
+
- 03b10a8: feat: upgrade zod from v3 to v4
|
|
90
|
+
|
|
91
|
+
BREAKING CHANGE: Upgraded zod from v3 to v4. This changes inferred TypeScript types
|
|
92
|
+
for schemas using `z.looseObject()` (previously `.passthrough()`) which now include
|
|
93
|
+
`[x: string]: unknown` in their output types. Consumers using these types may need
|
|
94
|
+
to update their code. Additionally, `ZodError.errors` has been replaced with
|
|
95
|
+
`ZodError.issues`, `z.record()` now requires explicit key schemas, and
|
|
96
|
+
`zod-to-json-schema` has been replaced with native `z.toJSONSchema()`.
|
|
97
|
+
|
|
98
|
+
## 0.10.0
|
|
99
|
+
|
|
100
|
+
### Minor Changes
|
|
101
|
+
|
|
102
|
+
- b18325b: feat: upgrade AI SDK to v6
|
|
103
|
+
- d43c4ee: feat: add conciseness evaluator
|
|
104
|
+
|
|
105
|
+
## 0.9.0
|
|
106
|
+
|
|
107
|
+
### Minor Changes
|
|
108
|
+
|
|
109
|
+
- 0738cb9: Add tool response handling llm evaluator.
|
|
110
|
+
- d6d1953: Add createToolSelectionEvaluator for evaluating LLM tool selection decisions. This evaluator determines whether the correct tool was selected for a given context, checking if the LLM chose the best available tool for the user query.
|
|
111
|
+
|
|
112
|
+
## 0.8.0
|
|
113
|
+
|
|
114
|
+
### Minor Changes
|
|
115
|
+
|
|
116
|
+
- d1c89bf: Add createToolInvocationEvaluator for evaluating LLM tool invocations
|
|
117
|
+
|
|
118
|
+
## 0.7.0
|
|
119
|
+
|
|
120
|
+
### Minor Changes
|
|
121
|
+
|
|
122
|
+
- da13ad5: add generic correctness evaluator
|
|
123
|
+
|
|
124
|
+
## 0.6.5
|
|
125
|
+
|
|
126
|
+
### Patch Changes
|
|
127
|
+
|
|
128
|
+
- 4208604: trigger changeset publish
|
|
129
|
+
|
|
130
|
+
## 0.6.4
|
|
131
|
+
|
|
132
|
+
### Patch Changes
|
|
133
|
+
|
|
134
|
+
- c96475c: trigger changeset publish
|
|
135
|
+
|
|
136
|
+
## 0.6.3
|
|
137
|
+
|
|
138
|
+
### Patch Changes
|
|
139
|
+
|
|
140
|
+
- 857b617: add links to packages
|
|
141
|
+
|
|
142
|
+
## 0.6.2
|
|
143
|
+
|
|
144
|
+
### Patch Changes
|
|
145
|
+
|
|
146
|
+
- 0589fc3: create evaluator helper
|
|
147
|
+
|
|
148
|
+
## 0.6.1
|
|
149
|
+
|
|
150
|
+
### Patch Changes
|
|
151
|
+
|
|
152
|
+
- 5cacc22: create evaluator helper function
|
|
153
|
+
|
|
154
|
+
## 0.6.0
|
|
155
|
+
|
|
156
|
+
### Minor Changes
|
|
157
|
+
|
|
158
|
+
- 84877d6: add data mapping helpers for evaluators
|
|
159
|
+
|
|
160
|
+
## 0.5.1
|
|
161
|
+
|
|
162
|
+
### Patch Changes
|
|
163
|
+
|
|
164
|
+
- 9408838: add message template support
|
|
165
|
+
|
|
166
|
+
## 0.5.0
|
|
167
|
+
|
|
168
|
+
### Minor Changes
|
|
169
|
+
|
|
170
|
+
- aa5dc26: serialize nested values in templates as JSON
|
|
171
|
+
|
|
172
|
+
## 0.4.0
|
|
173
|
+
|
|
174
|
+
### Minor Changes
|
|
175
|
+
|
|
176
|
+
- 885be2a: make phoenix-client be able to take in phoenix evals directly
|
|
177
|
+
|
|
178
|
+
## 0.3.1
|
|
179
|
+
|
|
180
|
+
### Patch Changes
|
|
181
|
+
|
|
182
|
+
- a756a95: fix Mustache.render by disabling HTML escape
|
|
183
|
+
|
|
184
|
+
## 0.3.0
|
|
185
|
+
|
|
186
|
+
### Minor Changes
|
|
187
|
+
|
|
188
|
+
- f66c5b6: refactor: change source to kind for evaluators
|
|
189
|
+
|
|
190
|
+
## 0.2.2
|
|
191
|
+
|
|
192
|
+
### Patch Changes
|
|
193
|
+
|
|
194
|
+
- c7cc7d9: feat: Add createOrGetDataset helper function to phoenix-client
|
|
195
|
+
|
|
196
|
+
Additionally clean up build artifacts and type-checking amongst example scripts.
|
|
197
|
+
|
|
198
|
+
## 0.2.1
|
|
199
|
+
|
|
200
|
+
### Patch Changes
|
|
201
|
+
|
|
202
|
+
- c85780b: Add support for generics across evals and experiments
|
|
203
|
+
|
|
204
|
+
## 0.2.0
|
|
205
|
+
|
|
206
|
+
### Minor Changes
|
|
207
|
+
|
|
208
|
+
- f92dab7: add name, optimization direction, etc.
|
|
209
|
+
|
|
210
|
+
## 0.1.0
|
|
211
|
+
|
|
212
|
+
### Minor Changes
|
|
213
|
+
|
|
214
|
+
- f11969d: capture source and optimization direction
|
|
215
|
+
|
|
216
|
+
## 0.0.8
|
|
217
|
+
|
|
218
|
+
### Patch Changes
|
|
219
|
+
|
|
220
|
+
- 7b204a5: bump the ai sdk to 5
|
|
221
|
+
|
|
222
|
+
## 0.0.7
|
|
223
|
+
|
|
224
|
+
### Patch Changes
|
|
225
|
+
|
|
226
|
+
- 83748e6: add type exports and better documentation across packages
|
|
227
|
+
|
|
228
|
+
## 0.0.6
|
|
229
|
+
|
|
230
|
+
### Patch Changes
|
|
231
|
+
|
|
232
|
+
- a57ec81: add telemetry controls
|
|
233
|
+
|
|
234
|
+
## 0.0.5
|
|
235
|
+
|
|
236
|
+
### Patch Changes
|
|
237
|
+
|
|
238
|
+
- 20db91d: Add tracing to evals, add tracing controls
|
|
239
|
+
|
|
240
|
+
## 0.0.4
|
|
241
|
+
|
|
242
|
+
### Patch Changes
|
|
243
|
+
|
|
244
|
+
- 2609fcd: bump target JS to es2017 for native async
|
|
245
|
+
|
|
246
|
+
## 0.0.3
|
|
247
|
+
|
|
248
|
+
### Patch Changes
|
|
249
|
+
|
|
250
|
+
- a6a4ca8: add docs to the readme
|
|
251
|
+
|
|
252
|
+
## 0.0.2
|
|
253
|
+
|
|
254
|
+
### Patch Changes
|
|
255
|
+
|
|
256
|
+
- b3e30db: Alpha version of evals
|
package/README.md
CHANGED
|
@@ -98,9 +98,10 @@ All pre-built evaluators are available from the `@arizeai/phoenix-evals/llm` mod
|
|
|
98
98
|
| Evaluator | Function | Description |
|
|
99
99
|
| ---------------------- | ------------------------------------- | --------------------------------------------------------------------------------- |
|
|
100
100
|
| Faithfulness | `createFaithfulnessEvaluator` | Detects hallucinations — checks if the output is grounded in the provided context |
|
|
101
|
+
| Completeness | `createCompletenessEvaluator` | Checks whether every active user request in a conversation was actually completed |
|
|
101
102
|
| Conciseness | `createConcisenessEvaluator` | Evaluates whether the response is appropriately concise |
|
|
102
103
|
| Correctness | `createCorrectnessEvaluator` | Checks if the output is factually correct given the input |
|
|
103
|
-
|
|
|
104
|
+
| Retrieval Relevance | `createRetrievalRelevanceEvaluator` | Measures how relevant retrieved information is to a request |
|
|
104
105
|
| Refusal | `createRefusalEvaluator` | Detects whether the model refused to answer |
|
|
105
106
|
| Tool Invocation | `createToolInvocationEvaluator` | Evaluates whether the correct tool was invoked with the right arguments |
|
|
106
107
|
| Tool Selection | `createToolSelectionEvaluator` | Checks whether the right tool was selected for the task |
|
|
@@ -112,7 +113,7 @@ import {
|
|
|
112
113
|
createFaithfulnessEvaluator,
|
|
113
114
|
createConcisenessEvaluator,
|
|
114
115
|
createCorrectnessEvaluator,
|
|
115
|
-
|
|
116
|
+
createRetrievalRelevanceEvaluator,
|
|
116
117
|
createRefusalEvaluator,
|
|
117
118
|
} from "@arizeai/phoenix-evals/llm";
|
|
118
119
|
import { openai } from "@ai-sdk/openai";
|
|
@@ -138,11 +139,11 @@ const correctnessResult = await correctnessEvaluator.evaluate({
|
|
|
138
139
|
console.log(correctnessResult);
|
|
139
140
|
// Output: { label: "correct", score: 1, explanation: "..." }
|
|
140
141
|
|
|
141
|
-
//
|
|
142
|
-
const relevanceEvaluator =
|
|
142
|
+
// Retrieval Relevance: checks if retrieved information is relevant to the request
|
|
143
|
+
const relevanceEvaluator = createRetrievalRelevanceEvaluator({ model });
|
|
143
144
|
const relevanceResult = await relevanceEvaluator.evaluate({
|
|
144
145
|
input: "What is the capital of France?",
|
|
145
|
-
|
|
146
|
+
context: "Paris is the capital of France and a major European city.",
|
|
146
147
|
});
|
|
147
148
|
console.log(relevanceResult);
|
|
148
149
|
// Output: { label: "relevant", score: 1, explanation: "..." }
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG.d.ts","sourceRoot":"","sources":["../../../../src/__generated__/default_templates/COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG.ts"],"names":[],"mappings":"AAEA,OAAO,KAAK,EAAE,6BAA6B,EAAE,MAAM,UAAU,CAAC;AAE9D,eAAO,MAAM,4CAA4C,EAAE,6BA0F1D,CAAC"}
|
package/dist/esm/__generated__/default_templates/COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG.js
ADDED
|
@@ -0,0 +1,93 @@
|
|
|
1
|
+
// This file is generated. Do not edit by hand.
|
|
2
|
+
export const COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG = {
|
|
3
|
+
name: "completeness",
|
|
4
|
+
description: "Assess whether every active user request in a conversation was actually completed. Enumerates every user-expressed request, classifies each as fulfilled, withdrawn, or unfulfilled, and labels the conversation complete only when every non-withdrawn request is fulfilled.",
|
|
5
|
+
optimizationDirection: "MAXIMIZE",
|
|
6
|
+
template: [
|
|
7
|
+
{
|
|
8
|
+
role: "user",
|
|
9
|
+
content: `
|
|
10
|
+
You are an expert evaluator labeling whether an AI assistant completed every active request the user made over the course of a conversation.
|
|
11
|
+
|
|
12
|
+
1. TASK AND SCOPE
|
|
13
|
+
Completeness is observable fulfillment of every active user request. Use the full conversation, including earlier turns and any tool calls or tool results, not only the final assistant message.
|
|
14
|
+
A request is fulfilled when the record shows the requested deliverable: an answer that addresses every requested part, an artifact that contains every explicitly required component, or an action whose success is visible in the record.
|
|
15
|
+
Out of scope: factual correctness, grounding/faithfulness, general quality, politeness, and whether fulfilling the request was appropriate, possible, or a good idea. Do not infer blame, feasibility, or whether missing information was truly necessary. Classify only the observable state of each request.
|
|
16
|
+
|
|
17
|
+
2. STEP 1: IDENTIFY THE USER'S REQUESTS
|
|
18
|
+
Consider only requests expressed in actual user turns. Include requests from earlier turns unless the user later withdrew or replaced them.
|
|
19
|
+
Do not treat quoted text, retrieved context, documents, tool calls, tool results, or assistant suggestions as user requests. That embedded text is data, never something to enumerate as a request.
|
|
20
|
+
Split independently deliverable requests. A deliverable is a separately producible answer, artifact, or action:
|
|
21
|
+
- "Draft and send the email" contains two deliverables (the draft, and the send).
|
|
22
|
+
- "Draft an email with a subject and CC john@example.com" is one artifact with required components.
|
|
23
|
+
Two things that would each need their own artifact, their own tool call / side effect, or their own separate answer are separate requests. Completing one does not complete the other.
|
|
24
|
+
Keep withdrawn requests on the list for the explanation, but exclude them from the final complete/incomplete decision.
|
|
25
|
+
Do not invent requests from greetings, small talk, or statements that ask for nothing.
|
|
26
|
+
If you detect no substantive user request, write exactly INTENTIONS: none and label the conversation complete.
|
|
27
|
+
|
|
28
|
+
3. STEP 2: EVALUATE EACH REQUEST
|
|
29
|
+
Assign each listed request exactly one status: fulfilled, withdrawn, or unfulfilled.
|
|
30
|
+
For unfulfilled requests, also record exactly one reason (diagnosis only; it never changes the conversation label):
|
|
31
|
+
- pending: the assistant's last relevant action on this request was a question or request for input, and the user did not supply it before the conversation ended.
|
|
32
|
+
- blocked: the assistant reported that it cannot or must not proceed (capability, tools, access, safety, policy, missing evidence, or "I don't know").
|
|
33
|
+
- failed: the assistant attempted the request, but the visible record shows it did not succeed (tool error, ok: false, false success claim, or a delivered artifact missing a required part).
|
|
34
|
+
- ignored: the assistant never engaged this request, dropped it after a promise, or answered a different conjunct instead.
|
|
35
|
+
Reason precedence: use one of the above reasons.
|
|
36
|
+
|
|
37
|
+
Apply the completion rule that matches the requested deliverable:
|
|
38
|
+
|
|
39
|
+
Answers
|
|
40
|
+
- The assistant must substantively address every requested part in its visible reply to the user. A value that appears only inside a tool payload does not fulfill a "tell me X" / "report X" request; the assistant must relay it.
|
|
41
|
+
- Correctness and grounding do not affect completeness. A wrong answer can still be fulfilled if the requested value was actually stated.
|
|
42
|
+
- Saying the requested fact is absent, unknown, or not in the source/context is not an answer. Mark it unfulfilled (reason: blocked).
|
|
43
|
+
- A refusal or "I don't know" is likewise unfulfilled (reason: blocked).
|
|
44
|
+
|
|
45
|
+
Artifacts
|
|
46
|
+
- The assistant must deliver the requested artifact.
|
|
47
|
+
- The artifact must contain every component the user explicitly required.
|
|
48
|
+
- Do not judge general quality beyond those explicit requirements. A missing required component is unfulfilled (reason: failed).
|
|
49
|
+
|
|
50
|
+
Actions
|
|
51
|
+
- Use evidence specific to each requested action.
|
|
52
|
+
- If matching tool evidence is present, it overrides the assistant's claim: confirmed success is fulfilled; failure or missing confirmation is unfulfilled (reason: failed).
|
|
53
|
+
- If no matching tool evidence is present throughout the whole conversation, judge from the assistant's visible response.
|
|
54
|
+
- A promise, acknowledgment, clarifying question, or statement of intent is not completion.
|
|
55
|
+
|
|
56
|
+
Withdrawn requests
|
|
57
|
+
- Mark withdrawn when the user cancelled, replaced, or otherwise took the request off the table ("stop", "never mind", "don't do that", "I only need X instead"). Honor the latest active instructions.
|
|
58
|
+
- List withdrawn requests with status withdrawn and reason_code none. Do not treat Withdrawn work as incomplete.
|
|
59
|
+
- After a withdrawal, still evaluate every remaining active request. Confirming "I did not do the cancelled thing" is not fulfillment of a different remaining ask.
|
|
60
|
+
|
|
61
|
+
4. STEP 3: ASSIGN THE CONVERSATION LABEL
|
|
62
|
+
<rubric>
|
|
63
|
+
COMPLETE - Every non-withdrawn request is fulfilled. Exclude withdrawn requests from the decision. A conversation with no substantive user request is complete. A conversation in which every request was withdrawn is complete.
|
|
64
|
+
|
|
65
|
+
INCOMPLETE - At least one non-withdrawn request is unfulfilled, for any reason (pending, blocked, failed, or ignored).
|
|
66
|
+
</rubric>
|
|
67
|
+
Do not collapse the session into a single reason. A session may contain more than one unfulfilled requests and with different reasons each; report each request's reason on its own line.
|
|
68
|
+
|
|
69
|
+
5. OUTPUT FORMAT
|
|
70
|
+
Put your complete analysis in your explanation. List each request in the order it first appeared, one line per request, using this exact format:
|
|
71
|
+
INTENTIONS:
|
|
72
|
+
- intention: "<concise statement of what the user wanted>" | state: <fulfilled|withdrawn|unfulfilled> | reason_code: <none|pending|blocked|failed|ignored> | reason: "<brief evidence-based justification, quoting the relevant assistant or tool output where possible>"
|
|
73
|
+
Use reason_code: none for fulfilled and withdrawn requests; for unfulfilled requests give exactly one of pending, blocked, failed, or ignored.
|
|
74
|
+
If the conversation contains no substantive request, write exactly:
|
|
75
|
+
INTENTIONS: none
|
|
76
|
+
Do not omit, invent, or merge independently deliverable requests. After the list, return the final complete or incomplete label: ignore withdrawn; complete only if every other listed request is fulfilled.
|
|
77
|
+
|
|
78
|
+
<data>
|
|
79
|
+
<conversation>
|
|
80
|
+
{{conversation}}
|
|
81
|
+
</conversation>
|
|
82
|
+
</data>
|
|
83
|
+
|
|
84
|
+
Work through the conversation in the three steps above, then decide. Did the assistant fulfill every active user request (complete), or did any non-withdrawn request remain unfulfilled (incomplete)?
|
|
85
|
+
`,
|
|
86
|
+
},
|
|
87
|
+
],
|
|
88
|
+
choices: {
|
|
89
|
+
"complete": 1,
|
|
90
|
+
"incomplete": 0
|
|
91
|
+
},
|
|
92
|
+
};
|
|
93
|
+
//# sourceMappingURL=COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG.js.map
|
package/dist/esm/__generated__/default_templates/COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG.js.map
ADDED
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG.js","sourceRoot":"","sources":["../../../../src/__generated__/default_templates/COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG.ts"],"names":[],"mappings":"AAAA,+CAA+C;AAI/C,MAAM,CAAC,MAAM,4CAA4C,GAAkC;IACzF,IAAI,EAAE,cAAc;IACpB,WAAW,EAAE,+QAA+Q;IAC5R,qBAAqB,EAAE,UAAU;IACjC,QAAQ,EAAE;QACR;YACE,IAAI,EAAE,MAAM;YACZ,OAAO,EAAE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA4Ed;SACI;KACF;IACD,OAAO,EAAE;QACT,UAAU,EAAE,CAAC;QACb,YAAY,EAAE,CAAC;KAChB;CACA,CAAC"}
|
|
@@ -1,3 +1,4 @@
|
|
|
1
|
+
export { COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG } from "./COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG.js";
|
|
1
2
|
export { CONCISENESS_CLASSIFICATION_EVALUATOR_CONFIG } from "./CONCISENESS_CLASSIFICATION_EVALUATOR_CONFIG.js";
|
|
2
3
|
export { CORRECTNESS_CLASSIFICATION_EVALUATOR_CONFIG } from "./CORRECTNESS_CLASSIFICATION_EVALUATOR_CONFIG.js";
|
|
3
4
|
export { DOCUMENT_RELEVANCE_CLASSIFICATION_EVALUATOR_CONFIG } from "./DOCUMENT_RELEVANCE_CLASSIFICATION_EVALUATOR_CONFIG.js";
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../../src/__generated__/default_templates/index.ts"],"names":[],"mappings":"AAEA,OAAO,EAAE,2CAA2C,EAAE,MAAM,+CAA+C,CAAC;AAC5G,OAAO,EAAE,2CAA2C,EAAE,MAAM,+CAA+C,CAAC;AAC5G,OAAO,EAAE,kDAAkD,EAAE,MAAM,sDAAsD,CAAC;AAC1H,OAAO,EAAE,4CAA4C,EAAE,MAAM,gDAAgD,CAAC;AAC9G,OAAO,EAAE,6CAA6C,EAAE,MAAM,iDAAiD,CAAC;AAChH,OAAO,EAAE,6CAA6C,EAAE,MAAM,iDAAiD,CAAC;AAChH,OAAO,EAAE,uCAAuC,EAAE,MAAM,2CAA2C,CAAC;AACpG,OAAO,EAAE,mDAAmD,EAAE,MAAM,uDAAuD,CAAC;AAC5H,OAAO,EAAE,+CAA+C,EAAE,MAAM,mDAAmD,CAAC;AACpH,OAAO,EAAE,sDAAsD,EAAE,MAAM,0DAA0D,CAAC;AAClI,OAAO,EAAE,8CAA8C,EAAE,MAAM,kDAAkD,CAAC;AAClH,OAAO,EAAE,wCAAwC,EAAE,MAAM,4CAA4C,CAAC;AACtG,OAAO,EAAE,6CAA6C,EAAE,MAAM,iDAAiD,CAAC"}
|
|
1
|
+
{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../../src/__generated__/default_templates/index.ts"],"names":[],"mappings":"AAEA,OAAO,EAAE,4CAA4C,EAAE,MAAM,gDAAgD,CAAC;AAC9G,OAAO,EAAE,2CAA2C,EAAE,MAAM,+CAA+C,CAAC;AAC5G,OAAO,EAAE,2CAA2C,EAAE,MAAM,+CAA+C,CAAC;AAC5G,OAAO,EAAE,kDAAkD,EAAE,MAAM,sDAAsD,CAAC;AAC1H,OAAO,EAAE,4CAA4C,EAAE,MAAM,gDAAgD,CAAC;AAC9G,OAAO,EAAE,6CAA6C,EAAE,MAAM,iDAAiD,CAAC;AAChH,OAAO,EAAE,6CAA6C,EAAE,MAAM,iDAAiD,CAAC;AAChH,OAAO,EAAE,uCAAuC,EAAE,MAAM,2CAA2C,CAAC;AACpG,OAAO,EAAE,mDAAmD,EAAE,MAAM,uDAAuD,CAAC;AAC5H,OAAO,EAAE,+CAA+C,EAAE,MAAM,mDAAmD,CAAC;AACpH,OAAO,EAAE,sDAAsD,EAAE,MAAM,0DAA0D,CAAC;AAClI,OAAO,EAAE,8CAA8C,EAAE,MAAM,kDAAkD,CAAC;AAClH,OAAO,EAAE,wCAAwC,EAAE,MAAM,4CAA4C,CAAC;AACtG,OAAO,EAAE,6CAA6C,EAAE,MAAM,iDAAiD,CAAC"}
|
|
@@ -1,4 +1,5 @@
|
|
|
1
1
|
// This file is generated. Do not edit by hand.
|
|
2
|
+
export { COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG } from "./COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG.js";
|
|
2
3
|
export { CONCISENESS_CLASSIFICATION_EVALUATOR_CONFIG } from "./CONCISENESS_CLASSIFICATION_EVALUATOR_CONFIG.js";
|
|
3
4
|
export { CORRECTNESS_CLASSIFICATION_EVALUATOR_CONFIG } from "./CORRECTNESS_CLASSIFICATION_EVALUATOR_CONFIG.js";
|
|
4
5
|
export { DOCUMENT_RELEVANCE_CLASSIFICATION_EVALUATOR_CONFIG } from "./DOCUMENT_RELEVANCE_CLASSIFICATION_EVALUATOR_CONFIG.js";
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"index.js","sourceRoot":"","sources":["../../../../src/__generated__/default_templates/index.ts"],"names":[],"mappings":"AAAA,+CAA+C;AAE/C,OAAO,EAAE,2CAA2C,EAAE,MAAM,+CAA+C,CAAC;AAC5G,OAAO,EAAE,2CAA2C,EAAE,MAAM,+CAA+C,CAAC;AAC5G,OAAO,EAAE,kDAAkD,EAAE,MAAM,sDAAsD,CAAC;AAC1H,OAAO,EAAE,4CAA4C,EAAE,MAAM,gDAAgD,CAAC;AAC9G,OAAO,EAAE,6CAA6C,EAAE,MAAM,iDAAiD,CAAC;AAChH,OAAO,EAAE,6CAA6C,EAAE,MAAM,iDAAiD,CAAC;AAChH,OAAO,EAAE,uCAAuC,EAAE,MAAM,2CAA2C,CAAC;AACpG,OAAO,EAAE,mDAAmD,EAAE,MAAM,uDAAuD,CAAC;AAC5H,OAAO,EAAE,+CAA+C,EAAE,MAAM,mDAAmD,CAAC;AACpH,OAAO,EAAE,sDAAsD,EAAE,MAAM,0DAA0D,CAAC;AAClI,OAAO,EAAE,8CAA8C,EAAE,MAAM,kDAAkD,CAAC;AAClH,OAAO,EAAE,wCAAwC,EAAE,MAAM,4CAA4C,CAAC;AACtG,OAAO,EAAE,6CAA6C,EAAE,MAAM,iDAAiD,CAAC"}
|
|
1
|
+
{"version":3,"file":"index.js","sourceRoot":"","sources":["../../../../src/__generated__/default_templates/index.ts"],"names":[],"mappings":"AAAA,+CAA+C;AAE/C,OAAO,EAAE,4CAA4C,EAAE,MAAM,gDAAgD,CAAC;AAC9G,OAAO,EAAE,2CAA2C,EAAE,MAAM,+CAA+C,CAAC;AAC5G,OAAO,EAAE,2CAA2C,EAAE,MAAM,+CAA+C,CAAC;AAC5G,OAAO,EAAE,kDAAkD,EAAE,MAAM,sDAAsD,CAAC;AAC1H,OAAO,EAAE,4CAA4C,EAAE,MAAM,gDAAgD,CAAC;AAC9G,OAAO,EAAE,6CAA6C,EAAE,MAAM,iDAAiD,CAAC;AAChH,OAAO,EAAE,6CAA6C,EAAE,MAAM,iDAAiD,CAAC;AAChH,OAAO,EAAE,uCAAuC,EAAE,MAAM,2CAA2C,CAAC;AACpG,OAAO,EAAE,mDAAmD,EAAE,MAAM,uDAAuD,CAAC;AAC5H,OAAO,EAAE,+CAA+C,EAAE,MAAM,mDAAmD,CAAC;AACpH,OAAO,EAAE,sDAAsD,EAAE,MAAM,0DAA0D,CAAC;AAClI,OAAO,EAAE,8CAA8C,EAAE,MAAM,kDAAkD,CAAC;AAClH,OAAO,EAAE,wCAAwC,EAAE,MAAM,4CAA4C,CAAC;AACtG,OAAO,EAAE,6CAA6C,EAAE,MAAM,iDAAiD,CAAC"}
|
|
@@ -0,0 +1,46 @@
|
|
|
1
|
+
import type { CreateClassificationEvaluatorArgs } from "../types/evals.js";
|
|
2
|
+
import type { ClassificationEvaluator } from "./ClassificationEvaluator.js";
|
|
3
|
+
export interface CompletenessEvaluatorArgs<RecordType extends Record<string, unknown> = CompletenessEvaluationRecord> extends Omit<CreateClassificationEvaluatorArgs<RecordType>, "promptTemplate" | "choices" | "optimizationDirection" | "name"> {
|
|
4
|
+
optimizationDirection?: CreateClassificationEvaluatorArgs<RecordType>["optimizationDirection"];
|
|
5
|
+
name?: CreateClassificationEvaluatorArgs<RecordType>["name"];
|
|
6
|
+
choices?: CreateClassificationEvaluatorArgs<RecordType>["choices"];
|
|
7
|
+
promptTemplate?: CreateClassificationEvaluatorArgs<RecordType>["promptTemplate"];
|
|
8
|
+
}
|
|
9
|
+
/**
|
|
10
|
+
* A conversation record to judge for completed user requests.
|
|
11
|
+
*/
|
|
12
|
+
export type CompletenessEvaluationRecord = {
|
|
13
|
+
/**
|
|
14
|
+
* The full conversation, including turns, tool calls, and tool results.
|
|
15
|
+
* Include tools for agent traces so action success can be verified.
|
|
16
|
+
*/
|
|
17
|
+
conversation: string;
|
|
18
|
+
[key: string]: unknown;
|
|
19
|
+
};
|
|
20
|
+
/**
|
|
21
|
+
* Creates a completeness evaluator.
|
|
22
|
+
*
|
|
23
|
+
* This function returns an evaluator that checks whether every active user
|
|
24
|
+
* request in a conversation was actually completed, not merely acknowledged.
|
|
25
|
+
*
|
|
26
|
+
* @param args - The arguments for creating the completeness evaluator.
|
|
27
|
+
* @param args.model - The model to use for classification.
|
|
28
|
+
* @param args.choices - The possible classification choices (defaults to complete/incomplete).
|
|
29
|
+
* @param args.promptTemplate - The prompt template to use (defaults to COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG.template).
|
|
30
|
+
* @param args.telemetry - The telemetry to use for the evaluator.
|
|
31
|
+
*
|
|
32
|
+
* @returns An evaluator function that takes a {@link CompletenessEvaluationRecord} and returns a classification result
|
|
33
|
+
* indicating whether the conversation is `complete` or `incomplete`.
|
|
34
|
+
*
|
|
35
|
+
* @example
|
|
36
|
+
* ```ts
|
|
37
|
+
* const evaluator = createCompletenessEvaluator({ model: openai("gpt-4o-mini") });
|
|
38
|
+
* const result = await evaluator.evaluate({
|
|
39
|
+
* conversation:
|
|
40
|
+
* "User: Reset my password and update the billing address.\nAssistant: Your password has been reset.",
|
|
41
|
+
* });
|
|
42
|
+
* console.log(result.label); // "incomplete"
|
|
43
|
+
* ```
|
|
44
|
+
*/
|
|
45
|
+
export declare function createCompletenessEvaluator<RecordType extends Record<string, unknown> = CompletenessEvaluationRecord>(args: CompletenessEvaluatorArgs<RecordType>): ClassificationEvaluator<RecordType>;
|
|
46
|
+
//# sourceMappingURL=createCompletenessEvaluator.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"createCompletenessEvaluator.d.ts","sourceRoot":"","sources":["../../../src/llm/createCompletenessEvaluator.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,iCAAiC,EAAE,MAAM,gBAAgB,CAAC;AACxE,OAAO,KAAK,EAAE,uBAAuB,EAAE,MAAM,2BAA2B,CAAC;AAGzE,MAAM,WAAW,yBAAyB,CACxC,UAAU,SAAS,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG,4BAA4B,CACzE,SAAQ,IAAI,CACZ,iCAAiC,CAAC,UAAU,CAAC,EAC7C,gBAAgB,GAAG,SAAS,GAAG,uBAAuB,GAAG,MAAM,CAChE;IACC,qBAAqB,CAAC,EAAE,iCAAiC,CAAC,UAAU,CAAC,CAAC,uBAAuB,CAAC,CAAC;IAC/F,IAAI,CAAC,EAAE,iCAAiC,CAAC,UAAU,CAAC,CAAC,MAAM,CAAC,CAAC;IAC7D,OAAO,CAAC,EAAE,iCAAiC,CAAC,UAAU,CAAC,CAAC,SAAS,CAAC,CAAC;IACnE,cAAc,CAAC,EAAE,iCAAiC,CAAC,UAAU,CAAC,CAAC,gBAAgB,CAAC,CAAC;CAClF;AAED;;GAEG;AACH,MAAM,MAAM,4BAA4B,GAAG;IACzC;;;OAGG;IACH,YAAY,EAAE,MAAM,CAAC;IACrB,CAAC,GAAG,EAAE,MAAM,GAAG,OAAO,CAAC;CACxB,CAAC;AAEF;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AACH,wBAAgB,2BAA2B,CACzC,UAAU,SAAS,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG,4BAA4B,EAEzE,IAAI,EAAE,yBAAyB,CAAC,UAAU,CAAC,GAC1C,uBAAuB,CAAC,UAAU,CAAC,CAerC"}
|
|
@@ -0,0 +1,38 @@
|
|
|
1
|
+
import { COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG } from "../__generated__/default_templates/index.js";
|
|
2
|
+
import { createClassificationEvaluator } from "./createClassificationEvaluator.js";
|
|
3
|
+
/**
|
|
4
|
+
* Creates a completeness evaluator.
|
|
5
|
+
*
|
|
6
|
+
* This function returns an evaluator that checks whether every active user
|
|
7
|
+
* request in a conversation was actually completed, not merely acknowledged.
|
|
8
|
+
*
|
|
9
|
+
* @param args - The arguments for creating the completeness evaluator.
|
|
10
|
+
* @param args.model - The model to use for classification.
|
|
11
|
+
* @param args.choices - The possible classification choices (defaults to complete/incomplete).
|
|
12
|
+
* @param args.promptTemplate - The prompt template to use (defaults to COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG.template).
|
|
13
|
+
* @param args.telemetry - The telemetry to use for the evaluator.
|
|
14
|
+
*
|
|
15
|
+
* @returns An evaluator function that takes a {@link CompletenessEvaluationRecord} and returns a classification result
|
|
16
|
+
* indicating whether the conversation is `complete` or `incomplete`.
|
|
17
|
+
*
|
|
18
|
+
* @example
|
|
19
|
+
* ```ts
|
|
20
|
+
* const evaluator = createCompletenessEvaluator({ model: openai("gpt-4o-mini") });
|
|
21
|
+
* const result = await evaluator.evaluate({
|
|
22
|
+
* conversation:
|
|
23
|
+
* "User: Reset my password and update the billing address.\nAssistant: Your password has been reset.",
|
|
24
|
+
* });
|
|
25
|
+
* console.log(result.label); // "incomplete"
|
|
26
|
+
* ```
|
|
27
|
+
*/
|
|
28
|
+
export function createCompletenessEvaluator(args) {
|
|
29
|
+
const { choices = COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG.choices, promptTemplate = COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG.template, optimizationDirection = COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG.optimizationDirection, name = COMPLETENESS_CLASSIFICATION_EVALUATOR_CONFIG.name, ...rest } = args;
|
|
30
|
+
return createClassificationEvaluator({
|
|
31
|
+
...rest,
|
|
32
|
+
promptTemplate,
|
|
33
|
+
choices,
|
|
34
|
+
optimizationDirection,
|
|
35
|
+
name,
|
|
36
|
+
});
|
|
37
|
+
}
|
|
38
|
+
//# sourceMappingURL=createCompletenessEvaluator.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"createCompletenessEvaluator.js","sourceRoot":"","sources":["../../../src/llm/createCompletenessEvaluator.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,4CAA4C,EAAE,MAAM,oCAAoC,CAAC;AAGlG,OAAO,EAAE,6BAA6B,EAAE,MAAM,iCAAiC,CAAC;AA0BhF;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AACH,MAAM,UAAU,2BAA2B,CAGzC,IAA2C;IAE3C,MAAM,EACJ,OAAO,GAAG,4CAA4C,CAAC,OAAO,EAC9D,cAAc,GAAG,4CAA4C,CAAC,QAAQ,EACtE,qBAAqB,GAAG,4CAA4C,CAAC,qBAAqB,EAC1F,IAAI,GAAG,4CAA4C,CAAC,IAAI,EACxD,GAAG,IAAI,EACR,GAAG,IAAI,CAAC;IACT,OAAO,6BAA6B,CAAa;QAC/C,GAAG,IAAI;QACP,cAAc;QACd,OAAO;QACP,qBAAqB;QACrB,IAAI;KACL,CAAC,CAAC;AACL,CAAC"}
|
|
@@ -1,5 +1,11 @@
|
|
|
1
1
|
import type { CreateClassificationEvaluatorArgs } from "../types/evals.js";
|
|
2
2
|
import type { ClassificationEvaluator } from "./ClassificationEvaluator.js";
|
|
3
|
+
/**
|
|
4
|
+
* Arguments for the document relevance evaluator.
|
|
5
|
+
*
|
|
6
|
+
* @deprecated Use {@link RetrievalRelevanceEvaluatorArgs} instead. This type will be
|
|
7
|
+
* removed in the next major release.
|
|
8
|
+
*/
|
|
3
9
|
export interface DocumentRelevanceEvaluatorArgs<RecordType extends Record<string, unknown> = DocumentRelevanceEvaluationRecord> extends Omit<CreateClassificationEvaluatorArgs<RecordType>, "promptTemplate" | "choices" | "optimizationDirection" | "name"> {
|
|
4
10
|
optimizationDirection?: CreateClassificationEvaluatorArgs<RecordType>["optimizationDirection"];
|
|
5
11
|
name?: CreateClassificationEvaluatorArgs<RecordType>["name"];
|
|
@@ -8,6 +14,9 @@ export interface DocumentRelevanceEvaluatorArgs<RecordType extends Record<string
|
|
|
8
14
|
}
|
|
9
15
|
/**
|
|
10
16
|
* A record to be evaluated by the document relevance evaluator.
|
|
17
|
+
*
|
|
18
|
+
* @deprecated Use {@link RetrievalRelevanceEvaluationRecord} instead. Rename
|
|
19
|
+
* `documentText` to `context`. This type will be removed in the next major release.
|
|
11
20
|
*/
|
|
12
21
|
export interface DocumentRelevanceEvaluationRecord {
|
|
13
22
|
input: string;
|
|
@@ -17,6 +26,12 @@ export interface DocumentRelevanceEvaluationRecord {
|
|
|
17
26
|
/**
|
|
18
27
|
* Creates a document relevance evaluator function.
|
|
19
28
|
*
|
|
29
|
+
* @deprecated Use {@link createRetrievalRelevanceEvaluator} instead. Rename the
|
|
30
|
+
* `documentText` input field to `context`, and update code that checks for the
|
|
31
|
+
* `unrelated` label to check for `irrelevant`. Pass one document as `context` to
|
|
32
|
+
* preserve per-document evaluation. This function will be removed in the next major
|
|
33
|
+
* release.
|
|
34
|
+
*
|
|
20
35
|
* This function returns an evaluator that determines whether a given document text
|
|
21
36
|
* is relevant to a provided input question. The evaluator uses a classification model
|
|
22
37
|
* and a prompt template to make its determination.
|
|
@@ -27,18 +42,8 @@ export interface DocumentRelevanceEvaluationRecord {
|
|
|
27
42
|
* @param args.promptTemplate - The prompt template to use (defaults to DOCUMENT_RELEVANCE_TEMPLATE).
|
|
28
43
|
* @param args.telemetry - The telemetry to use for the evaluator.
|
|
29
44
|
*
|
|
30
|
-
* @returns An evaluator function that takes a {@link
|
|
45
|
+
* @returns An evaluator function that takes a {@link DocumentRelevanceEvaluationRecord} and returns a classification result
|
|
31
46
|
* indicating whether the document is relevant to the input question.
|
|
32
|
-
*
|
|
33
|
-
* @example
|
|
34
|
-
* ```ts
|
|
35
|
-
* const evaluator = createDocumentRelevanceEvaluator({ model: openai("gpt-4o-mini") });
|
|
36
|
-
* const result = await evaluator.evaluate({
|
|
37
|
-
* input: "What is the capital of France?",
|
|
38
|
-
* documentText: "Paris is the capital and most populous city of France.",
|
|
39
|
-
* });
|
|
40
|
-
* console.log(result.label); // "relevant" or "unrelated"
|
|
41
|
-
* ```
|
|
42
47
|
*/
|
|
43
48
|
export declare function createDocumentRelevanceEvaluator<RecordType extends Record<string, unknown> = DocumentRelevanceEvaluationRecord>(args: DocumentRelevanceEvaluatorArgs<RecordType>): ClassificationEvaluator<RecordType>;
|
|
44
49
|
//# sourceMappingURL=createDocumentRelevanceEvaluator.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"createDocumentRelevanceEvaluator.d.ts","sourceRoot":"","sources":["../../../src/llm/createDocumentRelevanceEvaluator.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,iCAAiC,EAAE,MAAM,gBAAgB,CAAC;AACxE,OAAO,KAAK,EAAE,uBAAuB,EAAE,MAAM,2BAA2B,CAAC;AAGzE,MAAM,WAAW,8BAA8B,CAC7C,UAAU,SAAS,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GACxC,iCAAiC,CACnC,SAAQ,IAAI,CACZ,iCAAiC,CAAC,UAAU,CAAC,EAC7C,gBAAgB,GAAG,SAAS,GAAG,uBAAuB,GAAG,MAAM,CAChE;IACC,qBAAqB,CAAC,EAAE,iCAAiC,CAAC,UAAU,CAAC,CAAC,uBAAuB,CAAC,CAAC;IAC/F,IAAI,CAAC,EAAE,iCAAiC,CAAC,UAAU,CAAC,CAAC,MAAM,CAAC,CAAC;IAC7D,OAAO,CAAC,EAAE,iCAAiC,CAAC,UAAU,CAAC,CAAC,SAAS,CAAC,CAAC;IACnE,cAAc,CAAC,EAAE,iCAAiC,CAAC,UAAU,CAAC,CAAC,gBAAgB,CAAC,CAAC;CAClF;AAED
|
|
1
|
+
{"version":3,"file":"createDocumentRelevanceEvaluator.d.ts","sourceRoot":"","sources":["../../../src/llm/createDocumentRelevanceEvaluator.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,iCAAiC,EAAE,MAAM,gBAAgB,CAAC;AACxE,OAAO,KAAK,EAAE,uBAAuB,EAAE,MAAM,2BAA2B,CAAC;AAGzE;;;;;GAKG;AACH,MAAM,WAAW,8BAA8B,CAC7C,UAAU,SAAS,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GACxC,iCAAiC,CACnC,SAAQ,IAAI,CACZ,iCAAiC,CAAC,UAAU,CAAC,EAC7C,gBAAgB,GAAG,SAAS,GAAG,uBAAuB,GAAG,MAAM,CAChE;IACC,qBAAqB,CAAC,EAAE,iCAAiC,CAAC,UAAU,CAAC,CAAC,uBAAuB,CAAC,CAAC;IAC/F,IAAI,CAAC,EAAE,iCAAiC,CAAC,UAAU,CAAC,CAAC,MAAM,CAAC,CAAC;IAC7D,OAAO,CAAC,EAAE,iCAAiC,CAAC,UAAU,CAAC,CAAC,SAAS,CAAC,CAAC;IACnE,cAAc,CAAC,EAAE,iCAAiC,CAAC,UAAU,CAAC,CAAC,gBAAgB,CAAC,CAAC;CAClF;AAED;;;;;GAKG;AACH,MAAM,WAAW,iCAAiC;IAChD,KAAK,EAAE,MAAM,CAAC;IACd,YAAY,EAAE,MAAM,CAAC;IACrB,CAAC,GAAG,EAAE,MAAM,GAAG,OAAO,CAAC;CACxB;AAED;;;;;;;;;;;;;;;;;;;;;GAqBG;AACH,wBAAgB,gCAAgC,CAC9C,UAAU,SAAS,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GACxC,iCAAiC,EAEnC,IAAI,EAAE,8BAA8B,CAAC,UAAU,CAAC,GAC/C,uBAAuB,CAAC,UAAU,CAAC,CAerC"}
|
|
@@ -3,6 +3,12 @@ import { createClassificationEvaluator } from "./createClassificationEvaluator.j
|
|
|
3
3
|
/**
|
|
4
4
|
* Creates a document relevance evaluator function.
|
|
5
5
|
*
|
|
6
|
+
* @deprecated Use {@link createRetrievalRelevanceEvaluator} instead. Rename the
|
|
7
|
+
* `documentText` input field to `context`, and update code that checks for the
|
|
8
|
+
* `unrelated` label to check for `irrelevant`. Pass one document as `context` to
|
|
9
|
+
* preserve per-document evaluation. This function will be removed in the next major
|
|
10
|
+
* release.
|
|
11
|
+
*
|
|
6
12
|
* This function returns an evaluator that determines whether a given document text
|
|
7
13
|
* is relevant to a provided input question. The evaluator uses a classification model
|
|
8
14
|
* and a prompt template to make its determination.
|
|
@@ -13,18 +19,8 @@ import { createClassificationEvaluator } from "./createClassificationEvaluator.j
|
|
|
13
19
|
* @param args.promptTemplate - The prompt template to use (defaults to DOCUMENT_RELEVANCE_TEMPLATE).
|
|
14
20
|
* @param args.telemetry - The telemetry to use for the evaluator.
|
|
15
21
|
*
|
|
16
|
-
* @returns An evaluator function that takes a {@link
|
|
22
|
+
* @returns An evaluator function that takes a {@link DocumentRelevanceEvaluationRecord} and returns a classification result
|
|
17
23
|
* indicating whether the document is relevant to the input question.
|
|
18
|
-
*
|
|
19
|
-
* @example
|
|
20
|
-
* ```ts
|
|
21
|
-
* const evaluator = createDocumentRelevanceEvaluator({ model: openai("gpt-4o-mini") });
|
|
22
|
-
* const result = await evaluator.evaluate({
|
|
23
|
-
* input: "What is the capital of France?",
|
|
24
|
-
* documentText: "Paris is the capital and most populous city of France.",
|
|
25
|
-
* });
|
|
26
|
-
* console.log(result.label); // "relevant" or "unrelated"
|
|
27
|
-
* ```
|
|
28
24
|
*/
|
|
29
25
|
export function createDocumentRelevanceEvaluator(args) {
|
|
30
26
|
const { choices = DOCUMENT_RELEVANCE_CLASSIFICATION_EVALUATOR_CONFIG.choices, promptTemplate = DOCUMENT_RELEVANCE_CLASSIFICATION_EVALUATOR_CONFIG.template, optimizationDirection = DOCUMENT_RELEVANCE_CLASSIFICATION_EVALUATOR_CONFIG.optimizationDirection, name = DOCUMENT_RELEVANCE_CLASSIFICATION_EVALUATOR_CONFIG.name, ...rest } = args;
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"createDocumentRelevanceEvaluator.js","sourceRoot":"","sources":["../../../src/llm/createDocumentRelevanceEvaluator.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,kDAAkD,EAAE,MAAM,oCAAoC,CAAC;AAGxG,OAAO,EAAE,6BAA6B,EAAE,MAAM,iCAAiC,CAAC;
|
|
1
|
+
{"version":3,"file":"createDocumentRelevanceEvaluator.js","sourceRoot":"","sources":["../../../src/llm/createDocumentRelevanceEvaluator.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,kDAAkD,EAAE,MAAM,oCAAoC,CAAC;AAGxG,OAAO,EAAE,6BAA6B,EAAE,MAAM,iCAAiC,CAAC;AAiChF;;;;;;;;;;;;;;;;;;;;;GAqBG;AACH,MAAM,UAAU,gCAAgC,CAI9C,IAAgD;IAEhD,MAAM,EACJ,OAAO,GAAG,kDAAkD,CAAC,OAAO,EACpE,cAAc,GAAG,kDAAkD,CAAC,QAAQ,EAC5E,qBAAqB,GAAG,kDAAkD,CAAC,qBAAqB,EAChG,IAAI,GAAG,kDAAkD,CAAC,IAAI,EAC9D,GAAG,IAAI,EACR,GAAG,IAAI,CAAC;IACT,OAAO,6BAA6B,CAAa;QAC/C,GAAG,IAAI;QACP,cAAc;QACd,OAAO;QACP,qBAAqB;QACrB,IAAI;KACL,CAAC,CAAC;AACL,CAAC"}
|
|
@@ -26,11 +26,10 @@ export interface RetrievalRelevanceEvaluationRecord {
|
|
|
26
26
|
*
|
|
27
27
|
* This function returns an evaluator that determines whether the external
|
|
28
28
|
* information retrieved during a step is relevant to the request it was
|
|
29
|
-
* serving.
|
|
30
|
-
*
|
|
31
|
-
*
|
|
32
|
-
*
|
|
33
|
-
* search, or content embedded in an LLM turn.
|
|
29
|
+
* serving. It is source-agnostic and scores the retrieved information as a
|
|
30
|
+
* whole, whether it came from a vector search, a tool or MCP call, a web search,
|
|
31
|
+
* or content embedded in an LLM turn. Pass one document as `context` to evaluate
|
|
32
|
+
* documents individually.
|
|
34
33
|
*
|
|
35
34
|
* @param args - The arguments for creating the retrieval relevance evaluator.
|
|
36
35
|
* @param args.model - The model to use for classification.
|