新增人味文案评分和效果学习
This commit is contained in:
@@ -5,6 +5,7 @@ import { createManualPerformanceAdapter } from "../manual-adapter";
|
||||
import {
|
||||
createCalibrationEvent,
|
||||
GEO_RUBRIC_V1,
|
||||
scoreHumanCopyResult,
|
||||
scoreOptimizedArticle,
|
||||
} from "../scoring";
|
||||
|
||||
@@ -56,8 +57,10 @@ describe("calibration scoring", () => {
|
||||
const snapshot = await adapter.fetch({
|
||||
publication: {
|
||||
id: "pub_123",
|
||||
result_version_id: null,
|
||||
job_id: "job_123",
|
||||
revision: 2,
|
||||
publish_target: "official_site",
|
||||
platform: "official_site",
|
||||
url: "https://example.com/article",
|
||||
published_at: "2026-06-24T12:00:00.000Z",
|
||||
@@ -103,4 +106,39 @@ describe("calibration scoring", () => {
|
||||
expect(event.observations.join(" ")).toContain("询盘");
|
||||
expect(event.recommended_action).toContain("积累");
|
||||
});
|
||||
|
||||
it("scores human-copy result with a separate rubric", () => {
|
||||
const scoringRun = scoreHumanCopyResult({
|
||||
resultVersionId: "ver_1",
|
||||
result: {
|
||||
optimized_text: "我把这段文案顺了一下。",
|
||||
change_notes: [
|
||||
{
|
||||
original: "我把这段文案顺顺。",
|
||||
revised: "我把这段文案顺了一下。",
|
||||
reason: "修正重复表达。",
|
||||
confidence: "confident",
|
||||
revertible: false,
|
||||
},
|
||||
],
|
||||
ai_taste_checks: [
|
||||
{
|
||||
rule_id: "promotion_tone",
|
||||
status: "pass",
|
||||
evidence: "没有新增宣传腔。",
|
||||
suggestion: "",
|
||||
},
|
||||
],
|
||||
warnings: [],
|
||||
},
|
||||
});
|
||||
|
||||
expect(scoringRun).toMatchObject({
|
||||
result_version_id: "ver_1",
|
||||
case_type: "human_copy",
|
||||
rubric_version_id: "rubric_human_copy_v1",
|
||||
composite_score: expect.any(Number),
|
||||
});
|
||||
expect(scoringRun.composite_score).toBeGreaterThan(0);
|
||||
});
|
||||
});
|
||||
|
||||
@@ -1,6 +1,10 @@
|
||||
import { nanoid } from "nanoid";
|
||||
|
||||
import type { OptimizedArticle, QaReport } from "../domain/types";
|
||||
import type {
|
||||
CopyOptimizationResult,
|
||||
OptimizedArticle,
|
||||
QaReport,
|
||||
} from "../domain/types";
|
||||
import type {
|
||||
CalibrationContext,
|
||||
CalibrationDirection,
|
||||
@@ -56,16 +60,59 @@ export const GEO_RUBRIC_V1: RubricVersion = {
|
||||
],
|
||||
};
|
||||
|
||||
export const HUMAN_COPY_RUBRIC_V1: RubricVersion = {
|
||||
id: "rubric_human_copy_v1",
|
||||
version: "v1",
|
||||
name: "人味文案优化评分口径",
|
||||
dimensions: [
|
||||
{
|
||||
id: "restraint",
|
||||
label: "改动克制",
|
||||
weight: 0.2,
|
||||
description: "减少机械润色,不把短文案扩写成宣传稿。",
|
||||
},
|
||||
{
|
||||
id: "meaning_fidelity",
|
||||
label: "原意保真",
|
||||
weight: 0.25,
|
||||
description: "保留原文意图、事实和表达边界。",
|
||||
},
|
||||
{
|
||||
id: "natural_tone",
|
||||
label: "语气自然度",
|
||||
weight: 0.25,
|
||||
description: "读起来像真人表达,少套路句和格式痕迹。",
|
||||
},
|
||||
{
|
||||
id: "goal_fit",
|
||||
label: "目标匹配",
|
||||
weight: 0.15,
|
||||
description: "符合优化目标和发布场景。",
|
||||
},
|
||||
{
|
||||
id: "ai_taste_risk",
|
||||
label: "AI味风险",
|
||||
weight: 0.15,
|
||||
description: "宣传腔、套话、聊天痕迹和填充词风险低。",
|
||||
},
|
||||
],
|
||||
formula: "weighted_average_0_to_10",
|
||||
is_active: true,
|
||||
created_at: "2026-07-08T00:00:00.000Z",
|
||||
};
|
||||
|
||||
interface ScoreOptimizedArticleInput {
|
||||
jobId: string;
|
||||
article: OptimizedArticle;
|
||||
qaReport: QaReport;
|
||||
resultVersionId?: string | null;
|
||||
}
|
||||
|
||||
export function scoreOptimizedArticle({
|
||||
jobId,
|
||||
article,
|
||||
qaReport,
|
||||
resultVersionId = null,
|
||||
}: ScoreOptimizedArticleInput): ScoringRun {
|
||||
const combined = `${article.title}\n${article.summary}\n${article.body_markdown}`;
|
||||
const dimensionScores = {
|
||||
@@ -79,6 +126,8 @@ export function scoreOptimizedArticle({
|
||||
|
||||
return {
|
||||
id: `score_${nanoid(10)}`,
|
||||
result_version_id: resultVersionId,
|
||||
case_type: "article",
|
||||
job_id: jobId,
|
||||
revision: article.revision ?? 1,
|
||||
rubric_version_id: GEO_RUBRIC_V1.id,
|
||||
@@ -89,6 +138,43 @@ export function scoreOptimizedArticle({
|
||||
};
|
||||
}
|
||||
|
||||
export function scoreHumanCopyResult({
|
||||
resultVersionId,
|
||||
result,
|
||||
}: {
|
||||
resultVersionId: string;
|
||||
result: CopyOptimizationResult;
|
||||
}): ScoringRun {
|
||||
const warningsPenalty = Math.min(result.warnings.length, 3) * 0.5;
|
||||
const aiTasteWarnings = result.ai_taste_checks.filter(
|
||||
(check) => check.status === "warn",
|
||||
).length;
|
||||
const dimensionScores = {
|
||||
restraint: result.optimized_text.length > 280 ? 3 : 5,
|
||||
meaning_fidelity: result.change_notes.some(
|
||||
(note) => note.confidence === "uncertain",
|
||||
)
|
||||
? 3.5
|
||||
: 5,
|
||||
natural_tone: Math.max(2, 5 - aiTasteWarnings * 0.75),
|
||||
goal_fit: 4.5,
|
||||
ai_taste_risk: Math.max(1, 5 - aiTasteWarnings - warningsPenalty),
|
||||
};
|
||||
|
||||
return {
|
||||
id: `score_${nanoid(10)}`,
|
||||
result_version_id: resultVersionId,
|
||||
case_type: "human_copy",
|
||||
job_id: null,
|
||||
revision: null,
|
||||
rubric_version_id: HUMAN_COPY_RUBRIC_V1.id,
|
||||
dimension_scores: dimensionScores,
|
||||
composite_score: weightedAverage0To10(HUMAN_COPY_RUBRIC_V1, dimensionScores),
|
||||
rationale: "基于改动克制、原意保真、自然度、目标匹配和AI味风险生成评分。",
|
||||
created_at: new Date().toISOString(),
|
||||
};
|
||||
}
|
||||
|
||||
function scoreFactIntegrity(report: QaReport) {
|
||||
const hardRules = ["company_name_integrity", "claim_consistency", "hallucination_risk"];
|
||||
const statuses = report.checks
|
||||
@@ -143,11 +229,15 @@ function scoreReadability(article: OptimizedArticle) {
|
||||
}
|
||||
|
||||
function weightedComposite(scores: Record<string, number>) {
|
||||
const totalWeight = GEO_RUBRIC_V1.dimensions.reduce(
|
||||
return weightedAverage0To10(GEO_RUBRIC_V1, scores);
|
||||
}
|
||||
|
||||
function weightedAverage0To10(rubric: RubricVersion, scores: Record<string, number>) {
|
||||
const totalWeight = rubric.dimensions.reduce(
|
||||
(sum, dimension) => sum + dimension.weight,
|
||||
0,
|
||||
);
|
||||
const weighted = GEO_RUBRIC_V1.dimensions.reduce(
|
||||
const weighted = rubric.dimensions.reduce(
|
||||
(sum, dimension) => sum + (scores[dimension.id] ?? 0) * dimension.weight,
|
||||
0,
|
||||
);
|
||||
|
||||
+64
-3
@@ -207,8 +207,7 @@ export function initializeSchema(db: Database.Database) {
|
||||
`);
|
||||
|
||||
ensureColumn(db, "article_jobs", "case_id", "text");
|
||||
ensureColumn(db, "scoring_runs", "result_version_id", "text");
|
||||
ensureColumn(db, "scoring_runs", "case_type", "text not null default 'article'");
|
||||
migrateScoringRunsForCases(db);
|
||||
ensureColumn(db, "publication_records", "result_version_id", "text");
|
||||
ensureColumn(
|
||||
db,
|
||||
@@ -234,10 +233,72 @@ export function initializeSchema(db: Database.Database) {
|
||||
}
|
||||
|
||||
function getTableColumns(db: Database.Database, tableName: string) {
|
||||
return getTableColumnInfo(db, tableName).map((row) => row.name);
|
||||
}
|
||||
|
||||
function getTableColumnInfo(db: Database.Database, tableName: string) {
|
||||
return db
|
||||
.prepare(`pragma table_info(${tableName})`)
|
||||
.all()
|
||||
.map((row) => (row as { name: string }).name);
|
||||
.map((row) => row as { name: string; notnull: number });
|
||||
}
|
||||
|
||||
function migrateScoringRunsForCases(db: Database.Database) {
|
||||
const columns = getTableColumnInfo(db, "scoring_runs");
|
||||
const columnNames = columns.map((column) => column.name);
|
||||
const jobIdColumn = columns.find((column) => column.name === "job_id");
|
||||
const revisionColumn = columns.find((column) => column.name === "revision");
|
||||
const needsRebuild =
|
||||
!columnNames.includes("result_version_id") ||
|
||||
!columnNames.includes("case_type") ||
|
||||
jobIdColumn?.notnull === 1 ||
|
||||
revisionColumn?.notnull === 1;
|
||||
|
||||
if (!needsRebuild) return;
|
||||
|
||||
const resultVersionSelect = columnNames.includes("result_version_id")
|
||||
? "result_version_id"
|
||||
: "NULL";
|
||||
const caseTypeSelect = columnNames.includes("case_type")
|
||||
? "case_type"
|
||||
: "'article'";
|
||||
|
||||
db.pragma("foreign_keys = OFF");
|
||||
try {
|
||||
db.exec(`
|
||||
drop table if exists scoring_runs_next;
|
||||
|
||||
create table scoring_runs_next (
|
||||
id text primary key,
|
||||
result_version_id text,
|
||||
case_type text not null default 'article',
|
||||
job_id text,
|
||||
revision integer,
|
||||
rubric_version_id text not null,
|
||||
dimension_scores text not null,
|
||||
composite_score real not null,
|
||||
rationale text not null,
|
||||
created_at text not null,
|
||||
foreign key (result_version_id)
|
||||
references optimization_result_versions(id) on delete cascade,
|
||||
foreign key (rubric_version_id) references rubric_versions(id)
|
||||
);
|
||||
|
||||
insert into scoring_runs_next (
|
||||
id, result_version_id, case_type, job_id, revision, rubric_version_id,
|
||||
dimension_scores, composite_score, rationale, created_at
|
||||
)
|
||||
select
|
||||
id, ${resultVersionSelect}, ${caseTypeSelect}, job_id, revision,
|
||||
rubric_version_id, dimension_scores, composite_score, rationale, created_at
|
||||
from scoring_runs;
|
||||
|
||||
drop table scoring_runs;
|
||||
alter table scoring_runs_next rename to scoring_runs;
|
||||
`);
|
||||
} finally {
|
||||
db.pragma("foreign_keys = ON");
|
||||
}
|
||||
}
|
||||
|
||||
function ensureColumn(
|
||||
|
||||
Reference in New Issue
Block a user