@gmickel/gno 1.18.0 → 1.20.0
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 +14 -7
- package/assets/skill/SKILL.md +54 -12
- package/assets/skill/mcp-reference.md +7 -2
- package/assets/skill/recipes/citation-and-provenance.md +32 -9
- package/package.json +2 -1
- package/spec/AGENTS.md +83 -0
- package/spec/CLAUDE.md +83 -0
- package/spec/bench-fixture.schema.json +137 -0
- package/spec/cli.md +2919 -0
- package/spec/db/schema.sql +442 -0
- package/spec/evals-agentic.md +592 -0
- package/spec/evals.md +1106 -0
- package/spec/mcp.md +2279 -0
- package/spec/output-schemas/activation-verification.schema.json +515 -0
- package/spec/output-schemas/ask.schema.json +564 -0
- package/spec/output-schemas/backlinks.schema.json +131 -0
- package/spec/output-schemas/bench-result.schema.json +120 -0
- package/spec/output-schemas/capture-receipt.schema.json +143 -0
- package/spec/output-schemas/claim-verification.schema.json +291 -0
- package/spec/output-schemas/collection-list.schema.json +45 -0
- package/spec/output-schemas/context-capsule-v1.schema.json +726 -0
- package/spec/output-schemas/context-capsule-verification.schema.json +1338 -0
- package/spec/output-schemas/context-list.schema.json +21 -0
- package/spec/output-schemas/doctor.schema.json +313 -0
- package/spec/output-schemas/error.schema.json +30 -0
- package/spec/output-schemas/expansion.schema.json +37 -0
- package/spec/output-schemas/get.schema.json +140 -0
- package/spec/output-schemas/graph-query.schema.json +99 -0
- package/spec/output-schemas/graph.schema.json +371 -0
- package/spec/output-schemas/links-list.schema.json +186 -0
- package/spec/output-schemas/mcp-add-collection-result.schema.json +23 -0
- package/spec/output-schemas/mcp-capture-result.schema.json +152 -0
- package/spec/output-schemas/mcp-http-error.schema.json +30 -0
- package/spec/output-schemas/mcp-job-list.schema.json +58 -0
- package/spec/output-schemas/mcp-job-status.schema.json +224 -0
- package/spec/output-schemas/mcp-remove-result.schema.json +39 -0
- package/spec/output-schemas/mcp-sync-result.schema.json +41 -0
- package/spec/output-schemas/mcp-tag-result.schema.json +33 -0
- package/spec/output-schemas/models-list.schema.json +93 -0
- package/spec/output-schemas/multi-get.schema.json +103 -0
- package/spec/output-schemas/process-status.schema.json +119 -0
- package/spec/output-schemas/query-diagnose.schema.json +123 -0
- package/spec/output-schemas/resident-status.schema.json +154 -0
- package/spec/output-schemas/retrieval-trace-common.schema.json +492 -0
- package/spec/output-schemas/retrieval-trace-delete.schema.json +16 -0
- package/spec/output-schemas/retrieval-trace-export.schema.json +61 -0
- package/spec/output-schemas/retrieval-trace-filters.schema.json +139 -0
- package/spec/output-schemas/retrieval-trace-judgment.schema.json +15 -0
- package/spec/output-schemas/retrieval-trace-list.schema.json +18 -0
- package/spec/output-schemas/retrieval-trace-payloads.schema.json +178 -0
- package/spec/output-schemas/retrieval-trace-purge.schema.json +31 -0
- package/spec/output-schemas/retrieval-trace-qrels.schema.json +303 -0
- package/spec/output-schemas/retrieval-trace-replay.schema.json +286 -0
- package/spec/output-schemas/retrieval-trace-show.schema.json +69 -0
- package/spec/output-schemas/retrieval-trace-summary.schema.json +65 -0
- package/spec/output-schemas/search-result.schema.json +154 -0
- package/spec/output-schemas/search-results.schema.json +338 -0
- package/spec/output-schemas/similar.schema.json +84 -0
- package/spec/output-schemas/status.schema.json +676 -0
- package/spec/output-schemas/tags-list.schema.json +48 -0
- package/src/app/context-runtime-contract.ts +10 -5
- package/src/app/context-runtime-input.ts +29 -1
- package/src/app/context-runtime-types.ts +7 -0
- package/src/app/context-runtime.ts +20 -2
- package/src/app/context-surface.ts +4 -0
- package/src/app/verified-ask.ts +291 -0
- package/src/cli/commands/ask-format.ts +255 -0
- package/src/cli/commands/ask.ts +144 -183
- package/src/cli/commands/context-build.ts +56 -9
- package/src/cli/commands/get.ts +64 -3
- package/src/cli/commands/query.ts +62 -23
- package/src/cli/commands/replay.ts +140 -0
- package/src/cli/commands/search.ts +48 -3
- package/src/cli/commands/shared.ts +3 -1
- package/src/cli/commands/trace.ts +200 -0
- package/src/cli/commands/vsearch.ts +75 -53
- package/src/cli/program.ts +287 -1
- package/src/config/index.ts +9 -0
- package/src/config/retrieval-traces.ts +56 -0
- package/src/config/types.ts +4 -0
- package/src/core/context-budget.ts +6 -0
- package/src/core/context-capsule-retrieval-schema.ts +4 -0
- package/src/core/context-capsule-schema.ts +17 -0
- package/src/core/context-capsule-validation.ts +3 -2
- package/src/core/context-capsule.ts +18 -0
- package/src/core/context-compiler.ts +44 -25
- package/src/core/context-evidence.ts +6 -0
- package/src/core/retrieval-qrels.ts +405 -0
- package/src/core/retrieval-replay-candidate.ts +368 -0
- package/src/core/retrieval-replay-types.ts +109 -0
- package/src/core/retrieval-replay-validation.ts +89 -0
- package/src/core/retrieval-replay.ts +441 -0
- package/src/core/retrieval-trace-evidence-origin.ts +178 -0
- package/src/core/retrieval-trace-export.ts +113 -0
- package/src/core/retrieval-trace-filter-normalization.ts +27 -0
- package/src/core/retrieval-trace-filters.ts +19 -0
- package/src/core/retrieval-trace-management-helpers.ts +247 -0
- package/src/core/retrieval-trace-management-types.ts +132 -0
- package/src/core/retrieval-trace-management.ts +422 -0
- package/src/core/retrieval-trace-request.ts +141 -0
- package/src/core/retrieval-trace-session.ts +507 -0
- package/src/core/retrieval-trace.ts +472 -0
- package/src/llm/errors.ts +10 -1
- package/src/llm/httpGeneration.ts +11 -1
- package/src/llm/nodeLlamaCpp/generation.ts +54 -10
- package/src/llm/types.ts +6 -0
- package/src/mcp/tools/ask.ts +228 -0
- package/src/mcp/tools/context.ts +87 -15
- package/src/mcp/tools/get.ts +35 -1
- package/src/mcp/tools/index.ts +83 -0
- package/src/mcp/tools/query.ts +95 -64
- package/src/mcp/tools/search.ts +36 -13
- package/src/mcp/tools/trace.ts +143 -0
- package/src/mcp/tools/vsearch.ts +71 -38
- package/src/pipeline/answer.ts +167 -26
- package/src/pipeline/claim-verification-schema.ts +235 -0
- package/src/pipeline/claim-verification.ts +487 -0
- package/src/pipeline/claim-verifier.ts +474 -0
- package/src/pipeline/graph-retrieval.ts +15 -1
- package/src/pipeline/hybrid.ts +151 -43
- package/src/pipeline/search.ts +36 -3
- package/src/pipeline/trace-metadata.ts +47 -0
- package/src/pipeline/types.ts +68 -0
- package/src/pipeline/vsearch.ts +101 -38
- package/src/sdk/client.ts +415 -73
- package/src/sdk/documents.ts +48 -1
- package/src/sdk/index.ts +17 -0
- package/src/sdk/types.ts +28 -0
- package/src/serve/context-capsule.ts +67 -8
- package/src/serve/public/app.tsx +12 -1
- package/src/serve/public/components/AskVerificationPanel.tsx +189 -0
- package/src/serve/public/globals.built.css +1 -1
- package/src/serve/public/lib/workspace-tabs.ts +2 -0
- package/src/serve/public/pages/Ask.tsx +42 -4
- package/src/serve/public/pages/Dashboard.tsx +10 -0
- package/src/serve/public/pages/TraceHistory.tsx +478 -0
- package/src/serve/public/pages/trace-history-detail.tsx +224 -0
- package/src/serve/retrieval-trace.ts +28 -0
- package/src/serve/routes/api.ts +508 -68
- package/src/serve/routes/traces.ts +156 -0
- package/src/serve/server.ts +87 -2
- package/src/store/index.ts +31 -0
- package/src/store/migrations/014-retrieval-traces.ts +303 -0
- package/src/store/migrations/index.ts +2 -0
- package/src/store/retrieval-trace-codec.ts +384 -0
- package/src/store/sqlite/adapter.ts +153 -1
- package/src/store/sqlite/retrieval-trace-management-store.ts +341 -0
- package/src/store/sqlite/retrieval-trace-retention.ts +349 -0
- package/src/store/sqlite/retrieval-trace-rows.ts +267 -0
- package/src/store/sqlite/retrieval-trace-store.ts +515 -0
- package/src/store/types.ts +297 -0
- package/src/store/vector/sqlite-vec.ts +76 -1
- package/src/store/vector/types.ts +1 -1
|
@@ -0,0 +1,441 @@
|
|
|
1
|
+
/** Read-only comparison of persisted retrieval baselines with one candidate. */
|
|
2
|
+
|
|
3
|
+
import type { SearchResult } from "../pipeline/types";
|
|
4
|
+
import type { DocumentRow, StoreResult } from "../store/types";
|
|
5
|
+
import type {
|
|
6
|
+
RetrievalQrelsCase,
|
|
7
|
+
RetrievalQrel,
|
|
8
|
+
RetrievalQrelsEvidence,
|
|
9
|
+
} from "./retrieval-qrels";
|
|
10
|
+
import type { RetrievalReplayDeps } from "./retrieval-replay-candidate";
|
|
11
|
+
import type {
|
|
12
|
+
ReplayRetrievalTraceInput,
|
|
13
|
+
ReplayRetrievalTraceResult,
|
|
14
|
+
RetrievalReplayCandidate,
|
|
15
|
+
RetrievalReplayCaseResult,
|
|
16
|
+
RetrievalReplayQrelResult,
|
|
17
|
+
RetrievalReplayReason,
|
|
18
|
+
RetrievalReplaySourceState,
|
|
19
|
+
} from "./retrieval-replay-types";
|
|
20
|
+
|
|
21
|
+
import { parseUri } from "../app/constants";
|
|
22
|
+
import { computeRetrievalMetrics } from "../bench/metrics";
|
|
23
|
+
import { diagnoseQueryTarget } from "../pipeline/diagnose";
|
|
24
|
+
import { hashTraceCanonical } from "../store/retrieval-trace-codec";
|
|
25
|
+
import { ok } from "../store/types";
|
|
26
|
+
import { buildRetrievalQrelsArtifact } from "./retrieval-qrels";
|
|
27
|
+
import {
|
|
28
|
+
buildRetrievalReplaySearchOptions,
|
|
29
|
+
retrievalReplayLimit,
|
|
30
|
+
retrievalReplayTraceMetadata,
|
|
31
|
+
runRetrievalReplayCandidate,
|
|
32
|
+
} from "./retrieval-replay-candidate";
|
|
33
|
+
import { parseReplayRetrievalTraceInput } from "./retrieval-replay-validation";
|
|
34
|
+
import { buildRetrievalTraceFingerprints } from "./retrieval-trace-request";
|
|
35
|
+
|
|
36
|
+
export type { RetrievalReplayDeps } from "./retrieval-replay-candidate";
|
|
37
|
+
export { buildRetrievalReplaySearchOptions } from "./retrieval-replay-candidate";
|
|
38
|
+
|
|
39
|
+
const failedCandidateMetrics = () => ({
|
|
40
|
+
precisionAtK: 0,
|
|
41
|
+
recallAtK: 0,
|
|
42
|
+
f1AtK: 0,
|
|
43
|
+
mrr: 0,
|
|
44
|
+
ndcgAtK: 0,
|
|
45
|
+
});
|
|
46
|
+
|
|
47
|
+
const unreplayable = (
|
|
48
|
+
input: ReplayRetrievalTraceInput,
|
|
49
|
+
reason: RetrievalReplayReason
|
|
50
|
+
): ReplayRetrievalTraceResult => ({
|
|
51
|
+
schemaVersion: "1.0",
|
|
52
|
+
exportId: input.exportId,
|
|
53
|
+
candidate: input.candidate,
|
|
54
|
+
verdict: "unreplayable",
|
|
55
|
+
reason,
|
|
56
|
+
recommendation: "manual_review",
|
|
57
|
+
applied: false,
|
|
58
|
+
cases: [],
|
|
59
|
+
});
|
|
60
|
+
|
|
61
|
+
const RECONSTRUCTION_REASONS = [
|
|
62
|
+
"redaction_incompatible",
|
|
63
|
+
"query_missing",
|
|
64
|
+
"filters_incomplete",
|
|
65
|
+
"no_retrieval_run",
|
|
66
|
+
"ambiguous_missing_expected_run",
|
|
67
|
+
] as const satisfies readonly RetrievalReplayReason[];
|
|
68
|
+
|
|
69
|
+
const reconstructionReason = (message: string): RetrievalReplayReason | null =>
|
|
70
|
+
RECONSTRUCTION_REASONS.find(
|
|
71
|
+
(reason) =>
|
|
72
|
+
message === reason ||
|
|
73
|
+
message.startsWith(`${reason}:`) ||
|
|
74
|
+
message.includes(`${reason}:`)
|
|
75
|
+
) ?? null;
|
|
76
|
+
|
|
77
|
+
const targetMatchesResult = (
|
|
78
|
+
qrel: RetrievalQrel,
|
|
79
|
+
result: SearchResult
|
|
80
|
+
): boolean => {
|
|
81
|
+
if (qrel.target.sourceHash) {
|
|
82
|
+
return qrel.target.sourceHash === result.source.sourceHash;
|
|
83
|
+
}
|
|
84
|
+
if (qrel.target.docid) return qrel.target.docid === result.docid;
|
|
85
|
+
return qrel.target.uri === result.uri;
|
|
86
|
+
};
|
|
87
|
+
|
|
88
|
+
const targetMatchesEvidence = (
|
|
89
|
+
qrel: RetrievalQrel,
|
|
90
|
+
evidence: RetrievalQrelsEvidence
|
|
91
|
+
): boolean =>
|
|
92
|
+
qrel.target.sourceHash
|
|
93
|
+
? qrel.target.sourceHash === evidence.sourceHash
|
|
94
|
+
: qrel.target.docid
|
|
95
|
+
? qrel.target.docid === evidence.docid
|
|
96
|
+
: qrel.target.uri === evidence.uri;
|
|
97
|
+
|
|
98
|
+
const resolveDocument = async (
|
|
99
|
+
deps: RetrievalReplayDeps,
|
|
100
|
+
qrel: RetrievalQrel,
|
|
101
|
+
documents: DocumentRow[]
|
|
102
|
+
): Promise<StoreResult<DocumentRow | null>> => {
|
|
103
|
+
if (qrel.target.sourceHash) {
|
|
104
|
+
const matching = documents.filter(
|
|
105
|
+
(document) => document.sourceHash === qrel.target.sourceHash
|
|
106
|
+
);
|
|
107
|
+
const active = matching.find((document) => document.active);
|
|
108
|
+
if (active) return ok(active);
|
|
109
|
+
if (matching[0]) return ok(matching[0]);
|
|
110
|
+
}
|
|
111
|
+
if (qrel.target.docid) {
|
|
112
|
+
const byDocid = await deps.store.getDocumentByDocid(qrel.target.docid);
|
|
113
|
+
if (!byDocid.ok) return byDocid;
|
|
114
|
+
if (
|
|
115
|
+
byDocid.value &&
|
|
116
|
+
(!qrel.target.sourceHash ||
|
|
117
|
+
byDocid.value.sourceHash === qrel.target.sourceHash)
|
|
118
|
+
) {
|
|
119
|
+
return byDocid;
|
|
120
|
+
}
|
|
121
|
+
}
|
|
122
|
+
if (qrel.target.uri) return deps.store.getDocumentByUri(qrel.target.uri);
|
|
123
|
+
return ok(null);
|
|
124
|
+
};
|
|
125
|
+
|
|
126
|
+
const sourceState = (
|
|
127
|
+
qrel: RetrievalQrel,
|
|
128
|
+
document: DocumentRow | null
|
|
129
|
+
): RetrievalReplaySourceState => {
|
|
130
|
+
if (!document) return "missing";
|
|
131
|
+
if (!document.active) return "inactive";
|
|
132
|
+
if (!document.mirrorHash) return "no_indexed_content";
|
|
133
|
+
if (
|
|
134
|
+
(qrel.target.sourceHash &&
|
|
135
|
+
qrel.target.sourceHash !== document.sourceHash) ||
|
|
136
|
+
(qrel.target.mirrorHash && qrel.target.mirrorHash !== document.mirrorHash)
|
|
137
|
+
) {
|
|
138
|
+
return "stale";
|
|
139
|
+
}
|
|
140
|
+
return "unchanged";
|
|
141
|
+
};
|
|
142
|
+
|
|
143
|
+
const rankedIds = (
|
|
144
|
+
qrels: RetrievalQrel[],
|
|
145
|
+
ranked: Array<SearchResult | RetrievalQrelsEvidence>
|
|
146
|
+
): string[] =>
|
|
147
|
+
ranked.map((entry) => {
|
|
148
|
+
const match = qrels.find((qrel) =>
|
|
149
|
+
"source" in entry
|
|
150
|
+
? targetMatchesResult(qrel, entry)
|
|
151
|
+
: targetMatchesEvidence(qrel, entry)
|
|
152
|
+
);
|
|
153
|
+
return match?.qrelId ?? `unjudged:${"docid" in entry ? entry.docid : ""}`;
|
|
154
|
+
});
|
|
155
|
+
|
|
156
|
+
const metricsFor = (qrels: RetrievalQrel[], output: string[], k: number) => {
|
|
157
|
+
const expected = qrels
|
|
158
|
+
.filter((qrel) => qrel.relevance > 0)
|
|
159
|
+
.map((qrel) => qrel.qrelId);
|
|
160
|
+
return computeRetrievalMetrics({
|
|
161
|
+
output,
|
|
162
|
+
expected,
|
|
163
|
+
judgments: qrels.map((qrel) => ({
|
|
164
|
+
docid: qrel.qrelId,
|
|
165
|
+
relevance: qrel.relevance,
|
|
166
|
+
})),
|
|
167
|
+
k,
|
|
168
|
+
});
|
|
169
|
+
};
|
|
170
|
+
|
|
171
|
+
const compareMetrics = (
|
|
172
|
+
baseline: ReturnType<typeof metricsFor>,
|
|
173
|
+
candidate: ReturnType<typeof metricsFor>
|
|
174
|
+
): "improved" | "unchanged" | "regressed" => {
|
|
175
|
+
const left = [candidate.recallAtK, candidate.ndcgAtK, candidate.mrr];
|
|
176
|
+
const right = [baseline.recallAtK, baseline.ndcgAtK, baseline.mrr];
|
|
177
|
+
for (const [index, value] of left.entries()) {
|
|
178
|
+
if (value > right[index]!) return "improved";
|
|
179
|
+
if (value < right[index]!) return "regressed";
|
|
180
|
+
}
|
|
181
|
+
return "unchanged";
|
|
182
|
+
};
|
|
183
|
+
|
|
184
|
+
const diagnoseMissing = async (
|
|
185
|
+
deps: RetrievalReplayDeps,
|
|
186
|
+
source: RetrievalQrelsCase,
|
|
187
|
+
qrel: RetrievalQrel
|
|
188
|
+
): Promise<RetrievalReplayQrelResult["diagnostic"]> => {
|
|
189
|
+
const target = qrel.target.uri ?? qrel.target.docid;
|
|
190
|
+
if (!target) return null;
|
|
191
|
+
const targetCollection = qrel.target.uri
|
|
192
|
+
? parseUri(qrel.target.uri)?.collection
|
|
193
|
+
: undefined;
|
|
194
|
+
const diagnosed = await diagnoseQueryTarget(deps, source.query.text, {
|
|
195
|
+
...buildRetrievalReplaySearchOptions(
|
|
196
|
+
source,
|
|
197
|
+
{ id: "diagnose", type: "hybrid" },
|
|
198
|
+
targetCollection
|
|
199
|
+
),
|
|
200
|
+
target,
|
|
201
|
+
});
|
|
202
|
+
if (!diagnosed.ok) return null;
|
|
203
|
+
return {
|
|
204
|
+
status: diagnosed.value.target.status,
|
|
205
|
+
filterReasons: diagnosed.value.target.filterReasons,
|
|
206
|
+
stageDropReasons: diagnosed.value.stages.flatMap((stage) =>
|
|
207
|
+
stage.dropReason ? [stage.dropReason] : []
|
|
208
|
+
),
|
|
209
|
+
};
|
|
210
|
+
};
|
|
211
|
+
|
|
212
|
+
const replayCase = async (
|
|
213
|
+
deps: RetrievalReplayDeps,
|
|
214
|
+
source: RetrievalQrelsCase,
|
|
215
|
+
candidate: RetrievalReplayCandidate,
|
|
216
|
+
currentFingerprints: Awaited<
|
|
217
|
+
ReturnType<typeof buildRetrievalTraceFingerprints>
|
|
218
|
+
>,
|
|
219
|
+
candidateFingerprints: Awaited<
|
|
220
|
+
ReturnType<typeof buildRetrievalTraceFingerprints>
|
|
221
|
+
>
|
|
222
|
+
): Promise<RetrievalReplayCaseResult> => {
|
|
223
|
+
const documents = await deps.store.listDocuments();
|
|
224
|
+
const candidateResult = await runRetrievalReplayCandidate(
|
|
225
|
+
deps,
|
|
226
|
+
source,
|
|
227
|
+
candidate
|
|
228
|
+
);
|
|
229
|
+
const k = retrievalReplayLimit(source, candidate);
|
|
230
|
+
const baselineMetrics = metricsFor(
|
|
231
|
+
source.qrels,
|
|
232
|
+
rankedIds(source.qrels, source.baseline.ranked),
|
|
233
|
+
k
|
|
234
|
+
);
|
|
235
|
+
if (!documents.ok || !candidateResult.ok) {
|
|
236
|
+
return {
|
|
237
|
+
caseId: source.caseId,
|
|
238
|
+
traceId: source.traceId,
|
|
239
|
+
terminalStatus: source.terminalStatus,
|
|
240
|
+
verdict: "unreplayable",
|
|
241
|
+
reason: "candidate_failed",
|
|
242
|
+
metrics: {
|
|
243
|
+
baseline: baselineMetrics,
|
|
244
|
+
candidate: failedCandidateMetrics(),
|
|
245
|
+
baselineCoverage: baselineMetrics.recallAtK,
|
|
246
|
+
candidateCoverage: 0,
|
|
247
|
+
},
|
|
248
|
+
fingerprints: {
|
|
249
|
+
original: source.fingerprints,
|
|
250
|
+
current: currentFingerprints,
|
|
251
|
+
candidate: candidateFingerprints,
|
|
252
|
+
},
|
|
253
|
+
capabilityOutcomes: {
|
|
254
|
+
baseline: source.baseline.capabilityOutcomes,
|
|
255
|
+
candidate: [],
|
|
256
|
+
},
|
|
257
|
+
fallbackCodes: {
|
|
258
|
+
baseline: source.baseline.fallbackCodes,
|
|
259
|
+
candidate: [],
|
|
260
|
+
},
|
|
261
|
+
qrels: [],
|
|
262
|
+
};
|
|
263
|
+
}
|
|
264
|
+
const results = candidateResult.value.results;
|
|
265
|
+
const candidateMetrics = metricsFor(
|
|
266
|
+
source.qrels,
|
|
267
|
+
rankedIds(source.qrels, results),
|
|
268
|
+
k
|
|
269
|
+
);
|
|
270
|
+
const qrelResults: RetrievalReplayQrelResult[] = [];
|
|
271
|
+
let unreplayableReason: RetrievalReplayReason | null = null;
|
|
272
|
+
for (const qrel of source.qrels) {
|
|
273
|
+
const document = await resolveDocument(deps, qrel, documents.value);
|
|
274
|
+
const state = document.ok ? sourceState(qrel, document.value) : "missing";
|
|
275
|
+
if (!qrel.baselineMissing && qrel.relevance > 0) {
|
|
276
|
+
if (state === "stale") unreplayableReason = "source_stale";
|
|
277
|
+
if (state === "missing") unreplayableReason = "source_missing";
|
|
278
|
+
}
|
|
279
|
+
const candidateIndex = results.findIndex((result) =>
|
|
280
|
+
targetMatchesResult(qrel, result)
|
|
281
|
+
);
|
|
282
|
+
const candidateRank = candidateIndex < 0 ? null : candidateIndex + 1;
|
|
283
|
+
const baselineRank = qrel.evidence?.rank ?? null;
|
|
284
|
+
const hasOutcome = (kind: "opened" | "cited" | "pinned") =>
|
|
285
|
+
source.baseline.outcomes[kind].some((item) =>
|
|
286
|
+
targetMatchesEvidence(qrel, item)
|
|
287
|
+
);
|
|
288
|
+
qrelResults.push({
|
|
289
|
+
qrelId: qrel.qrelId,
|
|
290
|
+
label: qrel.label,
|
|
291
|
+
relevance: qrel.relevance,
|
|
292
|
+
baselineRank,
|
|
293
|
+
plannerRank: qrel.evidence?.plannerRank ?? null,
|
|
294
|
+
candidateRank,
|
|
295
|
+
rankDelta:
|
|
296
|
+
baselineRank === null || candidateRank === null
|
|
297
|
+
? null
|
|
298
|
+
: baselineRank - candidateRank,
|
|
299
|
+
opened: hasOutcome("opened"),
|
|
300
|
+
cited: hasOutcome("cited"),
|
|
301
|
+
pinned: hasOutcome("pinned"),
|
|
302
|
+
sourceState: state,
|
|
303
|
+
diagnostic:
|
|
304
|
+
qrel.baselineMissing && candidateRank === null
|
|
305
|
+
? await diagnoseMissing(deps, source, qrel)
|
|
306
|
+
: null,
|
|
307
|
+
});
|
|
308
|
+
}
|
|
309
|
+
const metadata = retrievalReplayTraceMetadata(
|
|
310
|
+
candidate,
|
|
311
|
+
candidateResult.value
|
|
312
|
+
);
|
|
313
|
+
const verdict = unreplayableReason
|
|
314
|
+
? "unreplayable"
|
|
315
|
+
: compareMetrics(baselineMetrics, candidateMetrics);
|
|
316
|
+
return {
|
|
317
|
+
caseId: source.caseId,
|
|
318
|
+
traceId: source.traceId,
|
|
319
|
+
terminalStatus: source.terminalStatus,
|
|
320
|
+
verdict,
|
|
321
|
+
reason: unreplayableReason,
|
|
322
|
+
metrics: {
|
|
323
|
+
baseline: baselineMetrics,
|
|
324
|
+
candidate: candidateMetrics,
|
|
325
|
+
baselineCoverage: baselineMetrics.recallAtK,
|
|
326
|
+
candidateCoverage: candidateMetrics.recallAtK,
|
|
327
|
+
},
|
|
328
|
+
fingerprints: {
|
|
329
|
+
original: source.fingerprints,
|
|
330
|
+
current: currentFingerprints,
|
|
331
|
+
candidate: candidateFingerprints,
|
|
332
|
+
},
|
|
333
|
+
capabilityOutcomes: {
|
|
334
|
+
baseline: source.baseline.capabilityOutcomes,
|
|
335
|
+
candidate:
|
|
336
|
+
metadata?.capabilityOutcomes.map((outcome) => ({
|
|
337
|
+
...outcome,
|
|
338
|
+
reasonCode: outcome.reasonCode ?? null,
|
|
339
|
+
})) ?? [],
|
|
340
|
+
},
|
|
341
|
+
fallbackCodes: {
|
|
342
|
+
baseline: source.baseline.fallbackCodes,
|
|
343
|
+
candidate: metadata?.fallbackCodes ?? [],
|
|
344
|
+
},
|
|
345
|
+
qrels: qrelResults,
|
|
346
|
+
};
|
|
347
|
+
};
|
|
348
|
+
|
|
349
|
+
export const replayRetrievalTraces = async (
|
|
350
|
+
deps: RetrievalReplayDeps,
|
|
351
|
+
input: ReplayRetrievalTraceInput
|
|
352
|
+
): Promise<StoreResult<ReplayRetrievalTraceResult>> => {
|
|
353
|
+
const parsed = parseReplayRetrievalTraceInput(input);
|
|
354
|
+
if (!parsed.ok) return parsed;
|
|
355
|
+
const validatedInput = parsed.value;
|
|
356
|
+
const stored = await deps.store.getRetrievalTraceExportBundle(
|
|
357
|
+
validatedInput.exportId
|
|
358
|
+
);
|
|
359
|
+
if (!stored.ok) {
|
|
360
|
+
if (stored.error.message.includes("missing retrieval trace")) {
|
|
361
|
+
return ok(unreplayable(validatedInput, "trace_missing"));
|
|
362
|
+
}
|
|
363
|
+
return stored;
|
|
364
|
+
}
|
|
365
|
+
if (!stored.value)
|
|
366
|
+
return ok(unreplayable(validatedInput, "manifest_missing"));
|
|
367
|
+
if (stored.value.manifest.format !== "qrels") {
|
|
368
|
+
return ok(unreplayable(validatedInput, "redaction_incompatible"));
|
|
369
|
+
}
|
|
370
|
+
const artifact = buildRetrievalQrelsArtifact(stored.value.traces);
|
|
371
|
+
if (!artifact.ok) {
|
|
372
|
+
return ok(
|
|
373
|
+
unreplayable(
|
|
374
|
+
validatedInput,
|
|
375
|
+
reconstructionReason(artifact.error.message) ?? "trace_missing"
|
|
376
|
+
)
|
|
377
|
+
);
|
|
378
|
+
}
|
|
379
|
+
if (
|
|
380
|
+
hashTraceCanonical(artifact.value) !== stored.value.manifest.artifactHash
|
|
381
|
+
) {
|
|
382
|
+
return ok(unreplayable(validatedInput, "manifest_hash_mismatch"));
|
|
383
|
+
}
|
|
384
|
+
let currentFingerprints;
|
|
385
|
+
let candidateFingerprints;
|
|
386
|
+
try {
|
|
387
|
+
currentFingerprints = await buildRetrievalTraceFingerprints({
|
|
388
|
+
store: deps.store,
|
|
389
|
+
config: deps.config,
|
|
390
|
+
pipeline: "retrieval-replay-current",
|
|
391
|
+
indexName: deps.indexName,
|
|
392
|
+
modelUris: deps.modelUris,
|
|
393
|
+
});
|
|
394
|
+
candidateFingerprints = await buildRetrievalTraceFingerprints({
|
|
395
|
+
store: deps.store,
|
|
396
|
+
config: deps.config,
|
|
397
|
+
pipeline: "retrieval-replay-candidate",
|
|
398
|
+
pipelineOptions: { ...validatedInput.candidate },
|
|
399
|
+
indexName: deps.indexName,
|
|
400
|
+
modelUris: deps.modelUris,
|
|
401
|
+
});
|
|
402
|
+
} catch {
|
|
403
|
+
return ok(unreplayable(validatedInput, "candidate_failed"));
|
|
404
|
+
}
|
|
405
|
+
const cases = [];
|
|
406
|
+
for (const source of artifact.value.cases) {
|
|
407
|
+
cases.push(
|
|
408
|
+
await replayCase(
|
|
409
|
+
deps,
|
|
410
|
+
source,
|
|
411
|
+
validatedInput.candidate,
|
|
412
|
+
currentFingerprints,
|
|
413
|
+
candidateFingerprints
|
|
414
|
+
)
|
|
415
|
+
);
|
|
416
|
+
}
|
|
417
|
+
const verdict = cases.some((item) => item.verdict === "unreplayable")
|
|
418
|
+
? "unreplayable"
|
|
419
|
+
: cases.some((item) => item.verdict === "regressed")
|
|
420
|
+
? "regressed"
|
|
421
|
+
: cases.some((item) => item.verdict === "improved")
|
|
422
|
+
? "improved"
|
|
423
|
+
: "unchanged";
|
|
424
|
+
const reason =
|
|
425
|
+
cases.find((item) => item.verdict === "unreplayable")?.reason ?? null;
|
|
426
|
+
return ok({
|
|
427
|
+
schemaVersion: "1.0",
|
|
428
|
+
exportId: validatedInput.exportId,
|
|
429
|
+
candidate: validatedInput.candidate,
|
|
430
|
+
verdict,
|
|
431
|
+
reason,
|
|
432
|
+
recommendation:
|
|
433
|
+
verdict === "improved"
|
|
434
|
+
? "promote"
|
|
435
|
+
: verdict === "regressed"
|
|
436
|
+
? "keep_baseline"
|
|
437
|
+
: "manual_review",
|
|
438
|
+
applied: false,
|
|
439
|
+
cases,
|
|
440
|
+
});
|
|
441
|
+
};
|
|
@@ -0,0 +1,178 @@
|
|
|
1
|
+
/** Exact retrieval-run provenance for later evidence outcomes. */
|
|
2
|
+
|
|
3
|
+
import type { StoreResult } from "../store/types";
|
|
4
|
+
|
|
5
|
+
import { err, ok } from "../store/types";
|
|
6
|
+
|
|
7
|
+
export interface RetrievalTraceEvidence {
|
|
8
|
+
docid: string;
|
|
9
|
+
sourceHash: string;
|
|
10
|
+
mirrorHash: string;
|
|
11
|
+
uri: string;
|
|
12
|
+
seq?: number;
|
|
13
|
+
startLine: number;
|
|
14
|
+
endLine: number;
|
|
15
|
+
passageHash: string;
|
|
16
|
+
score?: number;
|
|
17
|
+
rank?: number;
|
|
18
|
+
plannerRank?: number;
|
|
19
|
+
sources?: string[];
|
|
20
|
+
graphExpanded?: boolean;
|
|
21
|
+
}
|
|
22
|
+
|
|
23
|
+
const documentIdentity = (evidence: RetrievalTraceEvidence): string =>
|
|
24
|
+
[evidence.docid, evidence.sourceHash, evidence.mirrorHash, evidence.uri].join(
|
|
25
|
+
"\0"
|
|
26
|
+
);
|
|
27
|
+
|
|
28
|
+
const passageIdentity = (evidence: RetrievalTraceEvidence): string =>
|
|
29
|
+
[
|
|
30
|
+
documentIdentity(evidence),
|
|
31
|
+
evidence.startLine,
|
|
32
|
+
evidence.endLine,
|
|
33
|
+
evidence.passageHash,
|
|
34
|
+
].join("\0");
|
|
35
|
+
|
|
36
|
+
const addRun = (
|
|
37
|
+
index: Map<string, Set<string>>,
|
|
38
|
+
key: string,
|
|
39
|
+
runId: string
|
|
40
|
+
): void => {
|
|
41
|
+
const runs = index.get(key) ?? new Set<string>();
|
|
42
|
+
runs.add(runId);
|
|
43
|
+
index.set(key, runs);
|
|
44
|
+
};
|
|
45
|
+
|
|
46
|
+
const storedEvidence = (value: unknown): RetrievalTraceEvidence | null => {
|
|
47
|
+
if (!value || typeof value !== "object" || Array.isArray(value)) return null;
|
|
48
|
+
const item = value as Record<string, unknown>;
|
|
49
|
+
if (
|
|
50
|
+
typeof item.docid !== "string" ||
|
|
51
|
+
typeof item.sourceHash !== "string" ||
|
|
52
|
+
typeof item.mirrorHash !== "string" ||
|
|
53
|
+
typeof item.uri !== "string" ||
|
|
54
|
+
typeof item.startLine !== "number" ||
|
|
55
|
+
typeof item.endLine !== "number" ||
|
|
56
|
+
typeof item.passageHash !== "string"
|
|
57
|
+
) {
|
|
58
|
+
return null;
|
|
59
|
+
}
|
|
60
|
+
return {
|
|
61
|
+
docid: item.docid,
|
|
62
|
+
sourceHash: item.sourceHash,
|
|
63
|
+
mirrorHash: item.mirrorHash,
|
|
64
|
+
uri: item.uri,
|
|
65
|
+
startLine: item.startLine,
|
|
66
|
+
endLine: item.endLine,
|
|
67
|
+
passageHash: item.passageHash,
|
|
68
|
+
...(typeof item.seq === "number" ? { seq: item.seq } : {}),
|
|
69
|
+
};
|
|
70
|
+
};
|
|
71
|
+
|
|
72
|
+
export class RetrievalTraceEvidenceOrigins {
|
|
73
|
+
private readonly documents = new Map<string, Set<string>>();
|
|
74
|
+
private readonly passages = new Map<string, Set<string>>();
|
|
75
|
+
|
|
76
|
+
add(runId: string, evidence: RetrievalTraceEvidence[]): void {
|
|
77
|
+
for (const item of evidence) {
|
|
78
|
+
addRun(this.documents, documentIdentity(item), runId);
|
|
79
|
+
addRun(this.passages, passageIdentity(item), runId);
|
|
80
|
+
}
|
|
81
|
+
}
|
|
82
|
+
|
|
83
|
+
addFallback(runId: string, evidence: RetrievalTraceEvidence[]): void {
|
|
84
|
+
for (const item of evidence) {
|
|
85
|
+
const passageRuns = this.passages.get(passageIdentity(item));
|
|
86
|
+
const documentRuns = this.documents.get(documentIdentity(item));
|
|
87
|
+
if (
|
|
88
|
+
(passageRuns && passageRuns.size > 0) ||
|
|
89
|
+
(documentRuns && documentRuns.size > 0)
|
|
90
|
+
) {
|
|
91
|
+
continue;
|
|
92
|
+
}
|
|
93
|
+
addRun(this.documents, documentIdentity(item), runId);
|
|
94
|
+
addRun(this.passages, passageIdentity(item), runId);
|
|
95
|
+
}
|
|
96
|
+
}
|
|
97
|
+
|
|
98
|
+
addStored(runId: string, payload: Record<string, unknown>): void {
|
|
99
|
+
const ranked = Array.isArray(payload.ranked) ? payload.ranked : [];
|
|
100
|
+
this.add(
|
|
101
|
+
runId,
|
|
102
|
+
ranked.flatMap((value) => {
|
|
103
|
+
const parsed = storedEvidence(value);
|
|
104
|
+
return parsed ? [parsed] : [];
|
|
105
|
+
})
|
|
106
|
+
);
|
|
107
|
+
}
|
|
108
|
+
|
|
109
|
+
addStoredFallback(runId: string, payload: Record<string, unknown>): void {
|
|
110
|
+
const evidence = Array.isArray(payload.evidence) ? payload.evidence : [];
|
|
111
|
+
this.addFallback(
|
|
112
|
+
runId,
|
|
113
|
+
evidence.flatMap((value) => {
|
|
114
|
+
const parsed = storedEvidence(value);
|
|
115
|
+
return parsed ? [parsed] : [];
|
|
116
|
+
})
|
|
117
|
+
);
|
|
118
|
+
}
|
|
119
|
+
|
|
120
|
+
addStoredRuns(
|
|
121
|
+
runs: Array<{
|
|
122
|
+
runId: string;
|
|
123
|
+
kind: string;
|
|
124
|
+
payload: Record<string, unknown>;
|
|
125
|
+
}>
|
|
126
|
+
): void {
|
|
127
|
+
for (const run of runs.filter(({ kind }) => kind === "retrieval")) {
|
|
128
|
+
this.addStored(run.runId, run.payload);
|
|
129
|
+
}
|
|
130
|
+
for (const run of runs.filter(({ kind }) => kind === "get")) {
|
|
131
|
+
this.addStoredFallback(run.runId, run.payload);
|
|
132
|
+
}
|
|
133
|
+
for (const run of runs.filter(({ kind }) => kind === "context")) {
|
|
134
|
+
this.addStoredFallback(run.runId, run.payload);
|
|
135
|
+
}
|
|
136
|
+
}
|
|
137
|
+
|
|
138
|
+
group(
|
|
139
|
+
evidence: RetrievalTraceEvidence[]
|
|
140
|
+
): StoreResult<Map<string, RetrievalTraceEvidence[]>> {
|
|
141
|
+
if (evidence.length === 0) {
|
|
142
|
+
return err(
|
|
143
|
+
"INVALID_INPUT",
|
|
144
|
+
"outcome_evidence_unmatched: evidence outcomes cannot be empty"
|
|
145
|
+
);
|
|
146
|
+
}
|
|
147
|
+
const groups = new Map<string, RetrievalTraceEvidence[]>();
|
|
148
|
+
for (const item of evidence) {
|
|
149
|
+
const passageRuns = this.passages.get(passageIdentity(item));
|
|
150
|
+
const documentRuns = this.documents.get(documentIdentity(item));
|
|
151
|
+
const candidates =
|
|
152
|
+
passageRuns && passageRuns.size > 0 ? passageRuns : documentRuns;
|
|
153
|
+
if (!candidates || candidates.size === 0) {
|
|
154
|
+
return err(
|
|
155
|
+
"INVALID_INPUT",
|
|
156
|
+
`outcome_evidence_unmatched: ${item.uri} was not produced by a retrieval run`
|
|
157
|
+
);
|
|
158
|
+
}
|
|
159
|
+
if (candidates.size !== 1) {
|
|
160
|
+
return err(
|
|
161
|
+
"INVALID_INPUT",
|
|
162
|
+
`outcome_evidence_ambiguous: ${item.uri} belongs to multiple retrieval runs`
|
|
163
|
+
);
|
|
164
|
+
}
|
|
165
|
+
const runId = candidates.values().next().value;
|
|
166
|
+
if (typeof runId !== "string") {
|
|
167
|
+
return err(
|
|
168
|
+
"INVALID_INPUT",
|
|
169
|
+
`outcome_evidence_unmatched: ${item.uri} has no retrieval run`
|
|
170
|
+
);
|
|
171
|
+
}
|
|
172
|
+
const current = groups.get(runId) ?? [];
|
|
173
|
+
current.push(item);
|
|
174
|
+
groups.set(runId, current);
|
|
175
|
+
}
|
|
176
|
+
return ok(groups);
|
|
177
|
+
}
|
|
178
|
+
}
|