renor-engine 1.1.0 → 1.3.0
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 +27 -2
- package/THIRD-PARTY-NOTICES +2661 -0
- package/dist/BUILD.json +3 -3
- package/dist/cli.js +1656 -490
- package/package.json +16 -80
- package/dist/.tsbuildinfo +0 -1
- package/dist/api/apiAuth.js +0 -43
- package/dist/api/server.js +0 -174
- package/dist/buildIdentity.js +0 -149
- package/dist/callPhases.js +0 -82
- package/dist/cli-extractors.js +0 -46
- package/dist/config.js +0 -65
- package/dist/context/composer.js +0 -320
- package/dist/context/readHonesty.js +0 -478
- package/dist/context/rowStore.js +0 -238
- package/dist/conversation/formatter.js +0 -81
- package/dist/conversation/parser.js +0 -193
- package/dist/conversation/types.js +0 -2
- package/dist/conversation/writer.js +0 -42
- package/dist/db/client.js +0 -149
- package/dist/db/handoffTable.js +0 -31
- package/dist/db/indexSalvage.js +0 -101
- package/dist/db/migrations.js +0 -670
- package/dist/db/readCache.js +0 -109
- package/dist/db/schema.js +0 -5
- package/dist/db/sqlite.js +0 -396
- package/dist/db/statements.js +0 -89
- package/dist/digests/digestCache.js +0 -146
- package/dist/digests/digests.js +0 -597
- package/dist/digests/worker.js +0 -350
- package/dist/embeddings/caching.js +0 -337
- package/dist/embeddings/coalescing.js +0 -177
- package/dist/embeddings/interactive.js +0 -103
- package/dist/embeddings/local.js +0 -70
- package/dist/embeddings/openai.js +0 -106
- package/dist/embeddings/rebuild.js +0 -166
- package/dist/embeddings/simdScan.js +0 -260
- package/dist/embeddings/space.js +0 -21
- package/dist/embeddings/types.js +0 -14
- package/dist/embeddings/vectorIndex.js +0 -674
- package/dist/embeddings/vertex.js +0 -334
- package/dist/eval/artifacts.js +0 -42
- package/dist/eval/cacheReport.js +0 -146
- package/dist/eval/comparison.js +0 -182
- package/dist/eval/context700-candidate.js +0 -145
- package/dist/eval/evalCache.js +0 -568
- package/dist/eval/hashEmbedding.js +0 -35
- package/dist/eval/indexSnapshots.js +0 -548
- package/dist/eval/longmemeval.js +0 -56
- package/dist/eval/metrics.js +0 -181
- package/dist/eval/pageDump.js +0 -67
- package/dist/eval/renorbenchBaseline.js +0 -486
- package/dist/eval/renorbenchDocuments.js +0 -134
- package/dist/eval/renorbenchProduct.js +0 -182
- package/dist/eval/runLedger.js +0 -279
- package/dist/eval/runManifest.js +0 -485
- package/dist/eval/scoring.js +0 -197
- package/dist/eval/textTokens.js +0 -17
- package/dist/evidence/attach.js +0 -264
- package/dist/evidence/background.js +0 -147
- package/dist/evidence/evidenceCache.js +0 -182
- package/dist/evidence/gate.js +0 -37
- package/dist/evidence/keyPool.js +0 -85
- package/dist/evidence/ollamaChat.js +0 -84
- package/dist/evidence/optin.js +0 -165
- package/dist/evidence/options.js +0 -151
- package/dist/evidence/secondPass.js +0 -294
- package/dist/extraction/classifyCapture.js +0 -34
- package/dist/extraction/decision-model.js +0 -174
- package/dist/extraction/eventExtractor.js +0 -113
- package/dist/extraction/extractionClient.js +0 -122
- package/dist/extraction/factExtractor.js +0 -46
- package/dist/extraction/factsWritePath.js +0 -179
- package/dist/extraction/preferenceExtractor.js +0 -150
- package/dist/extraction/secretGuard.js +0 -176
- package/dist/extraction/secretScanner.js +0 -665
- package/dist/extraction/typedFacts.js +0 -224
- package/dist/extraction/writePathAdjudication.js +0 -110
- package/dist/importing/apply.js +0 -672
- package/dist/importing/bundle-write.js +0 -80
- package/dist/importing/cli.js +0 -408
- package/dist/importing/existing.js +0 -51
- package/dist/importing/facts.js +0 -121
- package/dist/importing/forgetImported.js +0 -79
- package/dist/importing/hosted.js +0 -467
- package/dist/importing/journalLog.js +0 -96
- package/dist/importing/markdown.js +0 -98
- package/dist/importing/plan.js +0 -201
- package/dist/importing/secretScreen.js +0 -113
- package/dist/importing/serialize.js +0 -188
- package/dist/importing/source.js +0 -463
- package/dist/importing/types.js +0 -6
- package/dist/indexer/chunker.js +0 -319
- package/dist/indexer/dailyEntries.js +0 -63
- package/dist/indexer/hasher.js +0 -10
- package/dist/indexer/laneRuns.js +0 -48
- package/dist/indexer/o200k.js +0 -481
- package/dist/indexer/outsideRoot.js +0 -77
- package/dist/indexer/parser.js +0 -170
- package/dist/indexer/pipeline.js +0 -988
- package/dist/indexer/staleness.js +0 -201
- package/dist/indexer/tokenizer.js +0 -77
- package/dist/indexer/walker.js +0 -170
- package/dist/mcp/hive.js +0 -88
- package/dist/mcp/requestErrors.js +0 -47
- package/dist/mcp/server.js +0 -426
- package/dist/memory/agentLease.js +0 -43
- package/dist/memory/confinedWrite.js +0 -315
- package/dist/memory/controls.js +0 -73
- package/dist/memory/dailyNote.js +0 -265
- package/dist/memory/durableAppend.js +0 -57
- package/dist/memory/factConflicts.js +0 -203
- package/dist/memory/factTerms.js +0 -321
- package/dist/memory/handoff.js +0 -363
- package/dist/memory/hiveNameHeader.cjs +0 -10
- package/dist/memory/inventory.js +0 -55
- package/dist/memory/invisibleText.js +0 -20
- package/dist/memory/lifecycle.js +0 -1991
- package/dist/memory/limits.js +0 -13
- package/dist/memory/requests.js +0 -183
- package/dist/memory/sourceName.js +0 -276
- package/dist/memory/statusFacts.js +0 -61
- package/dist/memory/whatsNew.js +0 -158
- package/dist/memory/writer.js +0 -175
- package/dist/memory/writerConflicts.js +0 -85
- package/dist/memorymap/memoryMap.js +0 -1594
- package/dist/memorymap/replacedJournal.js +0 -58
- package/dist/memorymap/replacements.js +0 -410
- package/dist/mvp/activity.js +0 -86
- package/dist/mvp/agentMemory.js +0 -1800
- package/dist/mvp/answerCheck.js +0 -211
- package/dist/mvp/auth.js +0 -531
- package/dist/mvp/cli.js +0 -75
- package/dist/mvp/client-store.js +0 -462
- package/dist/mvp/compileCache.js +0 -242
- package/dist/mvp/compileSeed.js +0 -33
- package/dist/mvp/config.js +0 -288
- package/dist/mvp/connectionStatus.js +0 -80
- package/dist/mvp/contextDistinct.js +0 -53
- package/dist/mvp/contextProfile.js +0 -87
- package/dist/mvp/contextRelevance.js +0 -413
- package/dist/mvp/directorySurface.js +0 -68
- package/dist/mvp/engine.js +0 -1078
- package/dist/mvp/exportAll.js +0 -374
- package/dist/mvp/factLanePack.js +0 -77
- package/dist/mvp/factMarkers.js +0 -65
- package/dist/mvp/factsLane.js +0 -203
- package/dist/mvp/hiddenSpans.js +0 -585
- package/dist/mvp/history.js +0 -414
- package/dist/mvp/hitDates.js +0 -31
- package/dist/mvp/hitFloor.js +0 -36
- package/dist/mvp/identity.js +0 -18
- package/dist/mvp/idle.js +0 -113
- package/dist/mvp/log.js +0 -118
- package/dist/mvp/memoryDir.js +0 -98
- package/dist/mvp/posthogMcp.js +0 -210
- package/dist/mvp/readScope.js +0 -104
- package/dist/mvp/readUsefulness.js +0 -222
- package/dist/mvp/sanitize.js +0 -48
- package/dist/mvp/schemaRejections.js +0 -91
- package/dist/mvp/scope.js +0 -127
- package/dist/mvp/server.js +0 -2604
- package/dist/mvp/serverCard.js +0 -38
- package/dist/mvp/storedSize.js +0 -139
- package/dist/mvp/telemetry.js +0 -474
- package/dist/net/poolFetch.js +0 -174
- package/dist/recall/actionRecall.js +0 -282
- package/dist/recall/autoRecall.js +0 -447
- package/dist/recall/searchRecall.js +0 -189
- package/dist/retrieval/adviceLines.js +0 -195
- package/dist/retrieval/adviceQuestion.js +0 -44
- package/dist/retrieval/conversationShape.js +0 -44
- package/dist/retrieval/dailyEntryTime.js +0 -55
- package/dist/retrieval/entityLookup.js +0 -162
- package/dist/retrieval/factsFirst.js +0 -188
- package/dist/retrieval/keyword.js +0 -84
- package/dist/retrieval/ledgerEventTime.js +0 -51
- package/dist/retrieval/ledgerLiveness.js +0 -83
- package/dist/retrieval/ledgerPresentation.js +0 -161
- package/dist/retrieval/noteIdentity.js +0 -57
- package/dist/retrieval/orchestrator.js +0 -1171
- package/dist/retrieval/plainWording.js +0 -236
- package/dist/retrieval/queryHelpers.js +0 -142
- package/dist/retrieval/questionSpan.js +0 -475
- package/dist/retrieval/readerEvents.js +0 -690
- package/dist/retrieval/readerLayout.js +0 -422
- package/dist/retrieval/relativeTimeHints.js +0 -113
- package/dist/retrieval/semantic.js +0 -67
- package/dist/router/routerCache.js +0 -106
- package/dist/router/sessionRouter.js +0 -726
- package/dist/toolRefusal.js +0 -120
- package/dist/watcher/watcher.js +0 -77
- /package/dist/{indexer/o200k.json → o200k.json} +0 -0
package/dist/eval/comparison.js
DELETED
|
@@ -1,182 +0,0 @@
|
|
|
1
|
-
"use strict";
|
|
2
|
-
/**
|
|
3
|
-
* Pure comparison policy between two eval runs. Extracted from the CLI
|
|
4
|
-
* wrapper so the promotion logic itself lives in the frozen evaluator
|
|
5
|
-
* trust domain (its hash is recorded in run manifests) and is unit tested.
|
|
6
|
-
*/
|
|
7
|
-
Object.defineProperty(exports, "__esModule", { value: true });
|
|
8
|
-
exports.QUALITY_METRICS = void 0;
|
|
9
|
-
exports.compareRuns = compareRuns;
|
|
10
|
-
exports.QUALITY_METRICS = [
|
|
11
|
-
{ key: 'hitAt1', higherBetter: true },
|
|
12
|
-
{ key: 'hitAt5', higherBetter: true },
|
|
13
|
-
{ key: 'recallAt5', higherBetter: true },
|
|
14
|
-
{ key: 'recallAllAt10', higherBetter: true },
|
|
15
|
-
{ key: 'mrr', higherBetter: true },
|
|
16
|
-
{ key: 'ndcgAt5', higherBetter: true },
|
|
17
|
-
{ key: 'contradictionFirstRate', higherBetter: false },
|
|
18
|
-
{ key: 'forbiddenRetrievalRate', higherBetter: false },
|
|
19
|
-
{ key: 'abstentionEmptyRate', higherBetter: true },
|
|
20
|
-
];
|
|
21
|
-
const EFFICIENCY_METRICS = ['meanLatencyMs', 'p50LatencyMs', 'p95LatencyMs', 'meanContextTokens'];
|
|
22
|
-
const MATERIAL_EFFICIENCY_METRICS = ['p95LatencyMs', 'meanContextTokens'];
|
|
23
|
-
function same(a, b) {
|
|
24
|
-
return JSON.stringify(a) === JSON.stringify(b);
|
|
25
|
-
}
|
|
26
|
-
function num(map, key) {
|
|
27
|
-
const value = map[key];
|
|
28
|
-
return typeof value === 'number' && Number.isFinite(value) ? value : null;
|
|
29
|
-
}
|
|
30
|
-
function delta(candidate, baseline) {
|
|
31
|
-
return candidate === null || baseline === null ? null : Math.round((candidate - baseline) * 10000) / 10000;
|
|
32
|
-
}
|
|
33
|
-
/** Relative improvement of candidate over baseline; null when not computable. */
|
|
34
|
-
function relativeGain(baseline, candidate) {
|
|
35
|
-
if (baseline === null || candidate === null || baseline <= 0)
|
|
36
|
-
return null;
|
|
37
|
-
return (baseline - candidate) / baseline;
|
|
38
|
-
}
|
|
39
|
-
function compareRuns(baseline, candidate, options = {}) {
|
|
40
|
-
const tolerance = options.maxQualityRegression ?? 0;
|
|
41
|
-
const metricList = options.metrics ?? exports.QUALITY_METRICS;
|
|
42
|
-
const minEfficiencyGain = options.minEfficiencyGain ?? 0.05;
|
|
43
|
-
if (!Number.isFinite(tolerance) || tolerance < 0 || tolerance > 1) {
|
|
44
|
-
throw new Error('maxQualityRegression must be between 0 and 1');
|
|
45
|
-
}
|
|
46
|
-
if (!Number.isFinite(minEfficiencyGain) || minEfficiencyGain < 0 || minEfficiencyGain > 1) {
|
|
47
|
-
throw new Error('minEfficiencyGain must be between 0 and 1');
|
|
48
|
-
}
|
|
49
|
-
const invariantChecks = {
|
|
50
|
-
benchmark: same(baseline.manifest.benchmark, candidate.manifest.benchmark),
|
|
51
|
-
cases: same(baseline.manifest.casesSha256, candidate.manifest.casesSha256),
|
|
52
|
-
fixture: same(baseline.manifest.fixtureSha256, candidate.manifest.fixtureSha256),
|
|
53
|
-
topK: same(baseline.manifest.topK, candidate.manifest.topK),
|
|
54
|
-
provider: same(baseline.manifest.provider, candidate.manifest.provider),
|
|
55
|
-
chunking: same(baseline.manifest.chunking, candidate.manifest.chunking),
|
|
56
|
-
evaluator: same(baseline.manifest.evaluatorSha256, candidate.manifest.evaluatorSha256),
|
|
57
|
-
cleanTree: baseline.manifest.gitDirty === false && candidate.manifest.gitDirty === false,
|
|
58
|
-
qualityEligible: baseline.manifest.qualityEligible === true && candidate.manifest.qualityEligible === true,
|
|
59
|
-
};
|
|
60
|
-
const reasons = [];
|
|
61
|
-
const datasetCompatible = invariantChecks.benchmark && invariantChecks.cases && invariantChecks.fixture;
|
|
62
|
-
if (!datasetCompatible) {
|
|
63
|
-
throw new Error(`Cannot compare runs from different benchmark data: ${JSON.stringify(invariantChecks)}`);
|
|
64
|
-
}
|
|
65
|
-
if (!invariantChecks.evaluator) {
|
|
66
|
-
if (!options.allowEvaluatorChange) {
|
|
67
|
-
throw new Error('Evaluator changed between baseline and candidate (evaluatorSha256 mismatch). ' +
|
|
68
|
-
'The candidate cannot grade its own exam: re-run the candidate against the frozen evaluator, ' +
|
|
69
|
-
'or pass --allow-evaluator-change to accept a DESCRIPTIVE_ONLY comparison during evaluator development.');
|
|
70
|
-
}
|
|
71
|
-
reasons.push('Evaluator code changed between runs; measurement meaning differs.');
|
|
72
|
-
}
|
|
73
|
-
if (!invariantChecks.cleanTree) {
|
|
74
|
-
if (!options.allowDirtyTree) {
|
|
75
|
-
reasons.push('Working tree was dirty in at least one run; executed code does not match the recorded git SHA.');
|
|
76
|
-
}
|
|
77
|
-
else {
|
|
78
|
-
reasons.push('Dirty working tree tolerated (--allow-dirty-tree).');
|
|
79
|
-
}
|
|
80
|
-
}
|
|
81
|
-
if (!invariantChecks.qualityEligible) {
|
|
82
|
-
reasons.push('At least one run used a non-quality provider (e.g. hash embeddings); infrastructure evidence only.');
|
|
83
|
-
}
|
|
84
|
-
const configCompatible = invariantChecks.topK && invariantChecks.provider && invariantChecks.chunking;
|
|
85
|
-
if (!configCompatible) {
|
|
86
|
-
reasons.push('Provider, topK, or chunking changed between runs.');
|
|
87
|
-
}
|
|
88
|
-
const regressionGate = options.regressionGate === true;
|
|
89
|
-
const experimentCompatible = Boolean(invariantChecks.evaluator
|
|
90
|
-
&& (regressionGate || (invariantChecks.cleanTree && invariantChecks.qualityEligible))
|
|
91
|
-
&& (configCompatible || (!regressionGate && options.allowConfigChange)));
|
|
92
|
-
const quality = Object.fromEntries(metricList.map(metric => {
|
|
93
|
-
const baselineValue = num(baseline.summary.overall, metric.key);
|
|
94
|
-
const candidateValue = num(candidate.summary.overall, metric.key);
|
|
95
|
-
return [metric.key, { baseline: baselineValue, candidate: candidateValue, delta: delta(candidateValue, baselineValue) }];
|
|
96
|
-
}));
|
|
97
|
-
const efficiency = Object.fromEntries(EFFICIENCY_METRICS.map(key => {
|
|
98
|
-
const baselineValue = num(baseline.summary.overall, key);
|
|
99
|
-
const candidateValue = num(candidate.summary.overall, key);
|
|
100
|
-
return [key, { baseline: baselineValue, candidate: candidateValue, delta: delta(candidateValue, baselineValue) }];
|
|
101
|
-
}));
|
|
102
|
-
let verdict;
|
|
103
|
-
if (!experimentCompatible) {
|
|
104
|
-
verdict = 'DESCRIPTIVE_ONLY';
|
|
105
|
-
}
|
|
106
|
-
else {
|
|
107
|
-
const hasRegression = metricList.some(metric => {
|
|
108
|
-
const { delta: d } = quality[metric.key];
|
|
109
|
-
if (typeof d !== 'number')
|
|
110
|
-
return false;
|
|
111
|
-
return metric.higherBetter ? d < -tolerance : d > tolerance;
|
|
112
|
-
});
|
|
113
|
-
const hasQualityGain = metricList.some(metric => {
|
|
114
|
-
const { delta: d } = quality[metric.key];
|
|
115
|
-
if (typeof d !== 'number')
|
|
116
|
-
return false;
|
|
117
|
-
return metric.higherBetter ? d > tolerance : d < -tolerance;
|
|
118
|
-
});
|
|
119
|
-
const hasMaterialEfficiencyGain = MATERIAL_EFFICIENCY_METRICS.some(key => {
|
|
120
|
-
const gain = relativeGain(efficiency[key].baseline, efficiency[key].candidate);
|
|
121
|
-
return gain !== null && gain >= minEfficiencyGain;
|
|
122
|
-
});
|
|
123
|
-
if (hasRegression)
|
|
124
|
-
verdict = 'REGRESSED';
|
|
125
|
-
else if (hasQualityGain || hasMaterialEfficiencyGain)
|
|
126
|
-
verdict = 'IMPROVED';
|
|
127
|
-
else
|
|
128
|
-
verdict = 'NEUTRAL';
|
|
129
|
-
}
|
|
130
|
-
// Category-collapse guard: an aggregate improvement must not hide one
|
|
131
|
-
// category regressing beyond tolerance (independent-review finding).
|
|
132
|
-
const categoryRegressions = [];
|
|
133
|
-
const baselineCategories = baseline.summary.byCategory ?? {};
|
|
134
|
-
const candidateCategories = candidate.summary.byCategory ?? {};
|
|
135
|
-
// Both sides' categories: a category the baseline has and the candidate lost is a regression too.
|
|
136
|
-
for (const name of new Set([...Object.keys(baselineCategories), ...Object.keys(candidateCategories)])) {
|
|
137
|
-
const baselineMetrics = baselineCategories[name];
|
|
138
|
-
const metrics = candidateCategories[name];
|
|
139
|
-
if (!baselineMetrics)
|
|
140
|
-
continue; // new category: nothing to regress against
|
|
141
|
-
if (!metrics) {
|
|
142
|
-
categoryRegressions.push(`${name}.missing`);
|
|
143
|
-
continue;
|
|
144
|
-
}
|
|
145
|
-
for (const metric of metricList) {
|
|
146
|
-
const b = num(baselineMetrics, metric.key);
|
|
147
|
-
const c = num(metrics, metric.key);
|
|
148
|
-
if (b === null || c === null)
|
|
149
|
-
continue;
|
|
150
|
-
const d = delta(c, b);
|
|
151
|
-
if (typeof d === 'number' && (metric.higherBetter ? d < -tolerance : d > tolerance)) {
|
|
152
|
-
categoryRegressions.push(`${name}.${metric.key}`);
|
|
153
|
-
}
|
|
154
|
-
}
|
|
155
|
-
}
|
|
156
|
-
for (const name of categoryRegressions) {
|
|
157
|
-
reasons.push(`Category regression beyond tolerance: ${name}`);
|
|
158
|
-
}
|
|
159
|
-
if (categoryRegressions.length > 0) {
|
|
160
|
-
// A category regression beyond tolerance is a rejection regardless of the
|
|
161
|
-
// aggregate verdict: an overall gain must not hide a collapsing category.
|
|
162
|
-
verdict = 'REGRESSED';
|
|
163
|
-
}
|
|
164
|
-
const note = regressionGate
|
|
165
|
-
? 'Regression gate against a committed baseline: this is a no-regression check, never a clean ablation or a promotion.'
|
|
166
|
-
: experimentCompatible
|
|
167
|
-
? 'IMPROVED requires no quality regression beyond tolerance and a quality gain or a material (>= threshold) efficiency gain.'
|
|
168
|
-
: 'Not a clean ablation: this comparison is descriptive only and cannot support promotion.';
|
|
169
|
-
return {
|
|
170
|
-
verdict,
|
|
171
|
-
cleanAblation: experimentCompatible && !regressionGate,
|
|
172
|
-
regressionGate,
|
|
173
|
-
invariantChecks,
|
|
174
|
-
tolerance,
|
|
175
|
-
minEfficiencyGain,
|
|
176
|
-
quality,
|
|
177
|
-
efficiency,
|
|
178
|
-
categoryRegressions,
|
|
179
|
-
reasons,
|
|
180
|
-
note,
|
|
181
|
-
};
|
|
182
|
-
}
|
|
@@ -1,145 +0,0 @@
|
|
|
1
|
-
"use strict";
|
|
2
|
-
Object.defineProperty(exports, "__esModule", { value: true });
|
|
3
|
-
exports.canonicalSegment = exports.COMPILER_VERSION = exports.isHeaderOnly = void 0;
|
|
4
|
-
exports.compileContext = compileContext;
|
|
5
|
-
/**
|
|
6
|
-
* Context700 campaign — MUTABLE context compiler.
|
|
7
|
-
*
|
|
8
|
-
* Campaign contract (docs/agents/autoresearch/context700/program.md):
|
|
9
|
-
* - Target: <= 700 mean Renor memory-evidence tokens/query on LongMemEval-S full-500.
|
|
10
|
-
* - Frozen quality floor: >= 428/500, no category regression > 2.0 pp vs N22 (430/500).
|
|
11
|
-
* - ONLY this file is mutable after harness freeze.
|
|
12
|
-
*
|
|
13
|
-
* The compiler is gold-blind: it may use the question text/date, retrieval hits,
|
|
14
|
-
* and structured extraction tables. It may NOT use gold answers, benchmark
|
|
15
|
-
* question IDs (as decision keys), judge outputs, or correctness labels.
|
|
16
|
-
*
|
|
17
|
-
* v3 (structured-first + date-preserving compaction + recency raw fill):
|
|
18
|
-
* Lesson from v1 (31/118 screen KILL): raw-first 55% budget + aggressive
|
|
19
|
-
* compaction starved date-bearing stubs and chronicle breadth. v3:
|
|
20
|
-
* 1. Query-gated structured rows FIRST (facts cap 12, chronicle cap 10, prefs cap 6).
|
|
21
|
-
* 2. Raw excerpts: compact turn markers, PRESERVE Timestamp/date lines and
|
|
22
|
-
* header stubs (they carry session dates), recency-ordered like N22.
|
|
23
|
-
* 3. Fill remaining budget by rank; no minimum raw count (tables win ties).
|
|
24
|
-
*/
|
|
25
|
-
const textTokens_1 = require("./textTokens");
|
|
26
|
-
const canonicalSegment = (notePath) => notePath.replace(/^Sessions\//, '').replace(/-occ\d+\.md$/, '.md').replace(/\.md$/, '');
|
|
27
|
-
exports.canonicalSegment = canonicalSegment;
|
|
28
|
-
const isHeaderOnly = (text) => !text.includes('### Turn') && text.trim().length < 250;
|
|
29
|
-
exports.isHeaderOnly = isHeaderOnly;
|
|
30
|
-
exports.COMPILER_VERSION = 'v3-struct-first-date-keeping';
|
|
31
|
-
const CONTENT_STOP = new Set('the a an of to for and or in on my i with is are was be that this it at as by from have has had do does did not no you your me we our they their he she his her about into over under then than so if but'.split(' '));
|
|
32
|
-
const contentWords = (s) => {
|
|
33
|
-
const out = new Set();
|
|
34
|
-
for (const w of s.toLowerCase().matchAll(/[a-z0-9]{3,}/g)) {
|
|
35
|
-
if (!CONTENT_STOP.has(w[0]))
|
|
36
|
-
out.add(w[0]);
|
|
37
|
-
}
|
|
38
|
-
return out;
|
|
39
|
-
};
|
|
40
|
-
/** Compact date: "2023/05/30 (Tue) 23:40" -> "2023/05/30". */
|
|
41
|
-
const compactDate = (d) => d ? d.replace(/\s*\((?:Mon|Tue|Wed|Thu|Fri|Sat|Sun)\).*/, '') : '';
|
|
42
|
-
/**
|
|
43
|
-
* v3 compaction: shorten turn markers to `[N-role]`, drop long session headers,
|
|
44
|
-
* but PRESERVE Timestamp lines (dates are load-bearing for temporal/multi-session).
|
|
45
|
-
*/
|
|
46
|
-
const compactText = (t) => t
|
|
47
|
-
.replace(/^### Turn (\d+) — (User|Assistant)\s*\|?\s*/gim, '[$1-$2] ')
|
|
48
|
-
.replace(/^Timestamp: (.*)$/gim, '@$1')
|
|
49
|
-
.replace(/^# Session .*\n?/gim, '')
|
|
50
|
-
.replace(/\n{3,}/g, '\n\n')
|
|
51
|
-
.trim();
|
|
52
|
-
const gateRows = (rows, qWords, cap) => {
|
|
53
|
-
if (!rows || rows.length === 0)
|
|
54
|
-
return [];
|
|
55
|
-
const scored = [];
|
|
56
|
-
for (const row of rows) {
|
|
57
|
-
const text = Object.values(row).map(v => (typeof v === 'string' ? v : JSON.stringify(v) ?? '')).join(' ');
|
|
58
|
-
let overlap = 0;
|
|
59
|
-
for (const w of contentWords(text))
|
|
60
|
-
if (qWords.has(w))
|
|
61
|
-
overlap++;
|
|
62
|
-
if (overlap > 0)
|
|
63
|
-
scored.push({ score: overlap, row });
|
|
64
|
-
}
|
|
65
|
-
scored.sort((a, b) => b.score - a.score);
|
|
66
|
-
return scored.slice(0, cap);
|
|
67
|
-
};
|
|
68
|
-
function compileContext(args) {
|
|
69
|
-
const { question, hits, tables, dateMap, maxContextTokens } = args;
|
|
70
|
-
const qWords = contentWords(question.question);
|
|
71
|
-
const parts = { rawTokens: 0, factsTokens: 0, chronicleTokens: 0, prefsTokens: 0 };
|
|
72
|
-
const excerpts = [];
|
|
73
|
-
// --- Structured rows first (gated) ---
|
|
74
|
-
let tokens = 0;
|
|
75
|
-
let remaining = maxContextTokens;
|
|
76
|
-
const factScored = gateRows(tables.facts ?? [], qWords, 12);
|
|
77
|
-
if (factScored.length) {
|
|
78
|
-
const lines = factScored.map(({ row: f }) => {
|
|
79
|
-
const co = f.coexist ? ` (also: ${f.coexist})` : '';
|
|
80
|
-
return `- ${f.value} ${f.unit ?? ''} ${f.object ?? ''}${f.event ? ` — ${f.event}` : ''}${co}`.replace(/\s+/g, ' ').trim();
|
|
81
|
-
});
|
|
82
|
-
const block = `Numeric facts:\n${lines.join('\n')}\nCite only values present here; otherwise use excerpts.`;
|
|
83
|
-
const blockTokens = (0, textTokens_1.countTextTokens)(block);
|
|
84
|
-
if (blockTokens <= remaining) {
|
|
85
|
-
excerpts.push({ text: block, date: null });
|
|
86
|
-
tokens += blockTokens;
|
|
87
|
-
parts.factsTokens = blockTokens;
|
|
88
|
-
remaining -= blockTokens;
|
|
89
|
-
}
|
|
90
|
-
}
|
|
91
|
-
const chronScored = gateRows(tables.chronicle ?? [], qWords, 10);
|
|
92
|
-
if (chronScored.length) {
|
|
93
|
-
const lines = chronScored.map(({ row: e }) => `- ${e.when ?? ''}: ${e.event ?? ''}`.replace(/\s+/g, ' ').trim());
|
|
94
|
-
const block = `Events:\n${lines.join('\n')}\nCite only events present here; otherwise use excerpts.`;
|
|
95
|
-
const blockTokens = (0, textTokens_1.countTextTokens)(block);
|
|
96
|
-
if (blockTokens <= remaining) {
|
|
97
|
-
excerpts.push({ text: block, date: null });
|
|
98
|
-
tokens += blockTokens;
|
|
99
|
-
parts.chronicleTokens = blockTokens;
|
|
100
|
-
remaining -= blockTokens;
|
|
101
|
-
}
|
|
102
|
-
}
|
|
103
|
-
const prefScored = gateRows(tables.prefs ?? [], qWords, 6);
|
|
104
|
-
if (prefScored.length) {
|
|
105
|
-
const lines = prefScored.map(({ row: p }) => `- ${p.statement ?? ''}`.replace(/\s+/g, ' ').trim());
|
|
106
|
-
const block = `Standing preferences:\n${lines.join('\n')}\nApply matching entries; otherwise use excerpts.`;
|
|
107
|
-
const blockTokens = (0, textTokens_1.countTextTokens)(block);
|
|
108
|
-
if (blockTokens <= remaining) {
|
|
109
|
-
excerpts.push({ text: block, date: null });
|
|
110
|
-
tokens += blockTokens;
|
|
111
|
-
parts.prefsTokens = blockTokens;
|
|
112
|
-
remaining -= blockTokens;
|
|
113
|
-
}
|
|
114
|
-
}
|
|
115
|
-
const raws = [];
|
|
116
|
-
const seen = new Set();
|
|
117
|
-
for (const hit of args.hits) {
|
|
118
|
-
const text = compactText((hit.text ?? '').trim());
|
|
119
|
-
if (!text || seen.has(hit.text ?? ''))
|
|
120
|
-
continue;
|
|
121
|
-
const sessionDate = dateMap?.get(canonicalSegment(hit.notePath)) ?? null;
|
|
122
|
-
seen.add(hit.text ?? '');
|
|
123
|
-
raws.push({ text, date: sessionDate });
|
|
124
|
-
}
|
|
125
|
-
// N22 recency ordering (newest-first) before selection.
|
|
126
|
-
if (args.recencyOrder) {
|
|
127
|
-
raws.sort((a, b) => (b.date ?? '').localeCompare(a.date ?? ''));
|
|
128
|
-
}
|
|
129
|
-
for (const raw of raws) {
|
|
130
|
-
const dateMark = raw.date ? `(${compactDate(raw.date)}) ` : '';
|
|
131
|
-
const candidate = `${dateMark}${raw.text}`;
|
|
132
|
-
const candidateTokens = (0, textTokens_1.countTextTokens)(candidate);
|
|
133
|
-
if (tokens + candidateTokens > maxContextTokens)
|
|
134
|
-
break;
|
|
135
|
-
tokens += candidateTokens;
|
|
136
|
-
excerpts.push({ text: candidate, date: raw.date });
|
|
137
|
-
}
|
|
138
|
-
parts.rawTokens = tokens - parts.factsTokens - parts.chronicleTokens - parts.prefsTokens;
|
|
139
|
-
return {
|
|
140
|
-
context: excerpts.map(excerpt => excerpt.text).join('\n\n'),
|
|
141
|
-
tokens,
|
|
142
|
-
parts,
|
|
143
|
-
diagnostics: { compilerVersion: args.compilerVersion },
|
|
144
|
-
};
|
|
145
|
-
}
|