pi-smart-router 1.1.0 → 1.2.1
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/.pi/extensions/smart-router/delegate-stream.ts +7 -1
- package/.pi/extensions/smart-router/fleet-bootstrap.ts +2 -0
- package/.pi/extensions/smart-router/pi-model-scope.ts +18 -3
- package/.pi/extensions/smart-router/routing-context.ts +10 -0
- package/README.md +78 -0
- package/config/benchmark-profiles.json +2 -2
- package/config/operator-config.json.example +7 -0
- package/dist/config/defaults.d.ts.map +1 -1
- package/dist/config/defaults.js +3 -1
- package/dist/config/defaults.js.map +1 -1
- package/dist/config/pi-model-mapper.d.ts +12 -1
- package/dist/config/pi-model-mapper.d.ts.map +1 -1
- package/dist/config/pi-model-mapper.js +33 -8
- package/dist/config/pi-model-mapper.js.map +1 -1
- package/dist/domain/matching/embedding-provider.d.ts +63 -1
- package/dist/domain/matching/embedding-provider.d.ts.map +1 -1
- package/dist/domain/matching/embedding-provider.js +117 -0
- package/dist/domain/matching/embedding-provider.js.map +1 -1
- package/dist/domain/matching/encoder-gate.d.ts +73 -0
- package/dist/domain/matching/encoder-gate.d.ts.map +1 -0
- package/dist/domain/matching/encoder-gate.js +66 -0
- package/dist/domain/matching/encoder-gate.js.map +1 -0
- package/dist/domain/matching/hydra-matcher.d.ts +24 -3
- package/dist/domain/matching/hydra-matcher.d.ts.map +1 -1
- package/dist/domain/matching/hydra-matcher.js +21 -2
- package/dist/domain/matching/hydra-matcher.js.map +1 -1
- package/dist/domain/pipeline/router-pipeline.d.ts.map +1 -1
- package/dist/domain/pipeline/router-pipeline.js +12 -0
- package/dist/domain/pipeline/router-pipeline.js.map +1 -1
- package/dist/domain/types/entities.d.ts +15 -0
- package/dist/domain/types/entities.d.ts.map +1 -1
- package/dist/domain/types/schemas.d.ts +63 -7
- package/dist/domain/types/schemas.d.ts.map +1 -1
- package/dist/domain/types/schemas.js +46 -3
- package/dist/domain/types/schemas.js.map +1 -1
- package/dist/infrastructure/pricing/price-broker.d.ts +6 -1
- package/dist/infrastructure/pricing/price-broker.d.ts.map +1 -1
- package/dist/infrastructure/pricing/price-broker.js +16 -1
- package/dist/infrastructure/pricing/price-broker.js.map +1 -1
- package/package.json +3 -3
- package/src/config/defaults.ts +3 -0
- package/src/config/pi-model-mapper.ts +48 -9
- package/src/domain/matching/embedding-provider.ts +205 -1
- package/src/domain/matching/encoder-gate.ts +121 -0
- package/src/domain/matching/hydra-matcher.ts +53 -3
- package/src/domain/pipeline/router-pipeline.ts +13 -0
- package/src/domain/types/entities.ts +15 -0
- package/src/domain/types/schemas.ts +56 -7
- package/src/infrastructure/pricing/price-broker.ts +18 -2
|
@@ -303,6 +303,14 @@ export const PlanningDelegateObservabilitySchema = z
|
|
|
303
303
|
})
|
|
304
304
|
.strict();
|
|
305
305
|
|
|
306
|
+
/** HyDRA text encoder selection (SP-156, #80). Hoisted above the feature
|
|
307
|
+
* sidecar schema: cascade telemetry (SP-292) embeds EncoderSchema there. */
|
|
308
|
+
export const EncoderSchema = z.enum(['minilm', 'granite']);
|
|
309
|
+
|
|
310
|
+
export type Encoder = z.infer<typeof EncoderSchema>;
|
|
311
|
+
|
|
312
|
+
export const DEFAULT_ENCODER: Encoder = 'minilm';
|
|
313
|
+
|
|
306
314
|
/**
|
|
307
315
|
* Privacy-safe routing feature sidecar attached to live decisions (SP-057).
|
|
308
316
|
* Mirrors {@link RoutingFeatureSidecar} in entities.ts.
|
|
@@ -331,6 +339,12 @@ export const RoutingFeatureSidecarSchema = z
|
|
|
331
339
|
prewarm_attempted: z.boolean().nullable().optional(),
|
|
332
340
|
prewarm_accepted: z.boolean().nullable().optional(),
|
|
333
341
|
prewarm_disabled_reason: z.string().nullable().optional(),
|
|
342
|
+
// Encoder cascade decision telemetry (SP-292, #173); present only on
|
|
343
|
+
// cascade-enabled hydra_match decisions.
|
|
344
|
+
encoder_selected: EncoderSchema.nullable().optional(),
|
|
345
|
+
token_estimate: z.number().nullable().optional(),
|
|
346
|
+
cascade_threshold: z.number().nullable().optional(),
|
|
347
|
+
cascade_fallback_reason: z.string().nullable().optional(),
|
|
334
348
|
})
|
|
335
349
|
.strict();
|
|
336
350
|
|
|
@@ -437,13 +451,6 @@ export const DEFAULT_LOCAL_ZERO_CONFIG: Readonly<LocalZeroConfig> = {
|
|
|
437
451
|
max_tool_use_requirement: 0.25,
|
|
438
452
|
} as const;
|
|
439
453
|
|
|
440
|
-
/** HyDRA text encoder selection (SP-156, #80). */
|
|
441
|
-
export const EncoderSchema = z.enum(['minilm', 'granite']);
|
|
442
|
-
|
|
443
|
-
export type Encoder = z.infer<typeof EncoderSchema>;
|
|
444
|
-
|
|
445
|
-
export const DEFAULT_ENCODER: Encoder = 'minilm';
|
|
446
|
-
|
|
447
454
|
/** HyDRA requirement head mode (SP-158, #81). */
|
|
448
455
|
export const HydraHeadsSchema = z.enum(['learned_projection', 'modernbert_k4']);
|
|
449
456
|
|
|
@@ -451,10 +458,52 @@ export type HydraHeads = z.infer<typeof HydraHeadsSchema>;
|
|
|
451
458
|
|
|
452
459
|
export const DEFAULT_HYDRA_HEADS: HydraHeads = 'learned_projection';
|
|
453
460
|
|
|
461
|
+
/**
|
|
462
|
+
* Opt-in encoder cascade knobs (SP-291, #173 part 1).
|
|
463
|
+
*
|
|
464
|
+
* When enabled, prompts whose estimated token count reaches
|
|
465
|
+
* `token_threshold` are routed to `long_context_encoder` (Granite) instead of
|
|
466
|
+
* the primary `hydra.encoder` (MiniLM), avoiding 512-token truncation on long
|
|
467
|
+
* prompts. Default OFF: existing single-encoder installs are unaffected and
|
|
468
|
+
* pay zero added cost (the gate is pre-embedding arithmetic).
|
|
469
|
+
*
|
|
470
|
+
* Hard constraint (#173): embeddings from different encoders live in different
|
|
471
|
+
* vector spaces — the cascade must never mix per-encoder artifacts (centroids,
|
|
472
|
+
* learned projection). Per-encoder artifact sets + cascading embedder land in
|
|
473
|
+
* SP-292.
|
|
474
|
+
*/
|
|
475
|
+
export const EncoderCascadeConfigSchema = z.object({
|
|
476
|
+
/** When false (default), every prompt uses the primary `hydra.encoder`. */
|
|
477
|
+
enabled: z.boolean().default(false),
|
|
478
|
+
/** Long-context encoder selected at/over the threshold. Granite 97M (SP-156). */
|
|
479
|
+
long_context_encoder: EncoderSchema.default('granite'),
|
|
480
|
+
/**
|
|
481
|
+
* Estimated-token boundary: estimates >= threshold route to the long-context
|
|
482
|
+
* encoder. Default 512 matches the MiniLM context window.
|
|
483
|
+
*/
|
|
484
|
+
token_threshold: z.number().int().min(1).default(512),
|
|
485
|
+
});
|
|
486
|
+
|
|
487
|
+
export type EncoderCascadeConfig = z.infer<typeof EncoderCascadeConfigSchema>;
|
|
488
|
+
|
|
489
|
+
/** Encoder cascade defaults per #173: opt-in, default off, 512-token threshold. */
|
|
490
|
+
export const DEFAULT_ENCODER_CASCADE_CONFIG: Readonly<EncoderCascadeConfig> = {
|
|
491
|
+
enabled: false,
|
|
492
|
+
long_context_encoder: 'granite',
|
|
493
|
+
token_threshold: 512,
|
|
494
|
+
} as const;
|
|
495
|
+
|
|
454
496
|
export const HydraConfigSchema = z.object({
|
|
455
497
|
artifact_cache_path: z.string(),
|
|
456
498
|
/** ONNX encoder: MiniLM (default) or Granite 97M 384-dim long-context trial. */
|
|
457
499
|
encoder: EncoderSchema.default(DEFAULT_ENCODER),
|
|
500
|
+
/**
|
|
501
|
+
* Opt-in per-prompt encoder cascade (SP-291, #173). Default off; when
|
|
502
|
+
* enabled, over-threshold prompts use the long-context encoder.
|
|
503
|
+
*/
|
|
504
|
+
encoder_cascade: EncoderCascadeConfigSchema.default(
|
|
505
|
+
DEFAULT_ENCODER_CASCADE_CONFIG,
|
|
506
|
+
),
|
|
458
507
|
/**
|
|
459
508
|
* Requirement extraction mode:
|
|
460
509
|
* - `learned_projection` — SP-115 384×3 linear projection (default)
|
|
@@ -11,10 +11,14 @@
|
|
|
11
11
|
* to finish the domain→infra inversion; this module re-exports them for
|
|
12
12
|
* import-path stability. What remains here is limit resolution and
|
|
13
13
|
* fleet-mapping, which depend on `config/pi-model-mapper` tier defaults.
|
|
14
|
+
*
|
|
15
|
+
* SP-299 (#177): limit resolution is a four-level cascade — YAML profile
|
|
16
|
+
* override → LiteLLM registry snapshot → pi registry (`contextWindow` /
|
|
17
|
+
* `maxTokens` carried on the mapped profile) → tier default.
|
|
14
18
|
*/
|
|
15
19
|
|
|
16
20
|
import type { ModelProfile, ModelLimits, PriceCatalog } from '../../domain/types/index.js';
|
|
17
|
-
import { getDefaultLimitsForTier } from '../../config/pi-model-mapper.js';
|
|
21
|
+
import { getDefaultLimitsForTier, type MappedModelProfile } from '../../config/pi-model-mapper.js';
|
|
18
22
|
|
|
19
23
|
export {
|
|
20
24
|
resolvePrice,
|
|
@@ -44,10 +48,19 @@ function lookupRegistryLimits(
|
|
|
44
48
|
return snapshot[registryKey] ?? snapshot[model.id];
|
|
45
49
|
}
|
|
46
50
|
|
|
51
|
+
/**
|
|
52
|
+
* Pi registry limits attached by `mapPiModelToProfile` (SP-299 / #177).
|
|
53
|
+
* Profiles built from fleet YAML never carry this field.
|
|
54
|
+
*/
|
|
55
|
+
function lookupPiRegistryLimits(model: ModelProfile): ModelLimits | undefined {
|
|
56
|
+
return (model as MappedModelProfile).pi_registry_limits;
|
|
57
|
+
}
|
|
58
|
+
|
|
47
59
|
/**
|
|
48
60
|
* Resolve effective context limits for a model.
|
|
49
61
|
*
|
|
50
|
-
* Priority: YAML/profile override → LiteLLM registry snapshot →
|
|
62
|
+
* Priority: YAML/profile override → LiteLLM registry snapshot → pi registry
|
|
63
|
+
* (`contextWindow`/`maxTokens`) → tier default (SP-299 / #177).
|
|
51
64
|
*/
|
|
52
65
|
export function resolveLimits(
|
|
53
66
|
model: ModelProfile,
|
|
@@ -55,6 +68,7 @@ export function resolveLimits(
|
|
|
55
68
|
): ResolvedLimits {
|
|
56
69
|
const yamlLimits = model.limits;
|
|
57
70
|
const registryLimits = catalog ? lookupRegistryLimits(model, catalog) : undefined;
|
|
71
|
+
const piRegistryLimits = lookupPiRegistryLimits(model);
|
|
58
72
|
const defaults = getDefaultLimitsForTier(model.tier);
|
|
59
73
|
|
|
60
74
|
return {
|
|
@@ -63,10 +77,12 @@ export function resolveLimits(
|
|
|
63
77
|
max_input_tokens:
|
|
64
78
|
yamlLimits?.max_input_tokens ??
|
|
65
79
|
registryLimits?.max_input_tokens ??
|
|
80
|
+
piRegistryLimits?.max_input_tokens ??
|
|
66
81
|
defaults.max_input_tokens,
|
|
67
82
|
max_output_tokens:
|
|
68
83
|
yamlLimits?.max_output_tokens ??
|
|
69
84
|
registryLimits?.max_output_tokens ??
|
|
85
|
+
piRegistryLimits?.max_output_tokens ??
|
|
70
86
|
defaults.max_output_tokens,
|
|
71
87
|
},
|
|
72
88
|
};
|