@elizaos/plugin-local-inference 2.0.0-beta.1 → 2.0.11-beta.7
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/LICENSE +21 -0
- package/README.md +83 -0
- package/package.json +81 -15
- package/src/actions/generate-media.d.ts +59 -0
- package/src/actions/generate-media.d.ts.map +1 -0
- package/src/actions/generate-media.ts +647 -0
- package/src/actions/identify-speaker.d.ts +23 -0
- package/src/actions/identify-speaker.d.ts.map +1 -0
- package/src/actions/identify-speaker.ts +171 -0
- package/src/adapters/capacitor-llama/__tests__/compat-behavior.test.ts +218 -0
- package/src/adapters/capacitor-llama/__tests__/index.test.ts +68 -0
- package/src/adapters/capacitor-llama/__tests__/structured-output.test.ts +215 -0
- package/src/adapters/capacitor-llama/__tests__/text-streaming.test.ts +174 -0
- package/src/adapters/capacitor-llama/environment.ts +71 -0
- package/src/adapters/capacitor-llama/index.browser.ts +83 -0
- package/src/adapters/capacitor-llama/index.ts +807 -0
- package/src/adapters/capacitor-llama/loader.ts +109 -0
- package/src/adapters/capacitor-llama/structured-output.ts +165 -0
- package/src/adapters/capacitor-llama/text-streaming.ts +227 -0
- package/src/adapters/capacitor-llama/types.ts +374 -0
- package/src/backends/apple-foundation.ts +127 -0
- package/src/index.d.ts +7 -0
- package/src/index.d.ts.map +1 -0
- package/src/index.ts +54 -0
- package/src/local-inference-routes.d.ts +38 -0
- package/src/local-inference-routes.d.ts.map +1 -0
- package/src/local-inference-routes.test.ts +344 -0
- package/src/local-inference-routes.ts +1543 -0
- package/src/provider.d.ts +21 -0
- package/src/provider.d.ts.map +1 -0
- package/src/provider.ts +1171 -0
- package/src/routes/compat-helpers.d.ts +18 -0
- package/src/routes/compat-helpers.d.ts.map +1 -0
- package/src/routes/compat-helpers.ts +274 -0
- package/src/routes/family-member-route.d.ts +62 -0
- package/src/routes/family-member-route.d.ts.map +1 -0
- package/src/routes/family-member-route.ts +353 -0
- package/src/routes/index.d.ts +19 -0
- package/src/routes/index.d.ts.map +1 -0
- package/src/routes/index.ts +60 -0
- package/src/routes/live-diarization-route.d.ts +26 -0
- package/src/routes/live-diarization-route.d.ts.map +1 -0
- package/src/routes/live-diarization-route.test.ts +213 -0
- package/src/routes/live-diarization-route.ts +122 -0
- package/src/routes/local-inference-asr-route.d.ts +4 -0
- package/src/routes/local-inference-asr-route.d.ts.map +1 -0
- package/src/routes/local-inference-asr-route.test.ts +190 -0
- package/src/routes/local-inference-asr-route.ts +213 -0
- package/src/routes/local-inference-compat-routes.d.ts +16 -0
- package/src/routes/local-inference-compat-routes.d.ts.map +1 -0
- package/src/routes/local-inference-compat-routes.test.ts +423 -0
- package/src/routes/local-inference-compat-routes.ts +782 -0
- package/src/routes/local-inference-tts-route.d.ts +7 -0
- package/src/routes/local-inference-tts-route.d.ts.map +1 -0
- package/src/routes/local-inference-tts-route.test.ts +179 -0
- package/src/routes/local-inference-tts-route.ts +230 -0
- package/src/routes/voice-first-run-routes.d.ts +62 -0
- package/src/routes/voice-first-run-routes.d.ts.map +1 -0
- package/src/routes/voice-first-run-routes.ts +524 -0
- package/src/routes/voice-models-routes.d.ts +62 -0
- package/src/routes/voice-models-routes.d.ts.map +1 -0
- package/src/routes/voice-models-routes.ts +554 -0
- package/src/routes/voice-profile-plugin-routes.d.ts +19 -0
- package/src/routes/voice-profile-plugin-routes.d.ts.map +1 -0
- package/src/routes/voice-profile-plugin-routes.ts +138 -0
- package/src/routes/voice-profiles-management-routes.d.ts +52 -0
- package/src/routes/voice-profiles-management-routes.d.ts.map +1 -0
- package/src/routes/voice-profiles-management-routes.ts +476 -0
- package/src/routes/voice-speaker-profile-routes.d.ts +57 -0
- package/src/routes/voice-speaker-profile-routes.d.ts.map +1 -0
- package/src/routes/voice-speaker-profile-routes.ts +199 -0
- package/src/runtime/aosp-llama-loader-selection.test.ts +80 -0
- package/src/runtime/capacitor-llama.d.ts +25 -0
- package/src/runtime/embedding-manager-support.d.ts +77 -0
- package/src/runtime/embedding-manager-support.d.ts.map +1 -0
- package/src/runtime/embedding-manager-support.ts +497 -0
- package/src/runtime/embedding-presets.d.ts +16 -0
- package/src/runtime/embedding-presets.d.ts.map +1 -0
- package/src/runtime/embedding-presets.ts +81 -0
- package/src/runtime/embedding-warmup-policy.d.ts +14 -0
- package/src/runtime/embedding-warmup-policy.d.ts.map +1 -0
- package/src/runtime/embedding-warmup-policy.test.ts +53 -0
- package/src/runtime/embedding-warmup-policy.ts +48 -0
- package/src/runtime/ensure-local-inference-handler.d.ts +53 -0
- package/src/runtime/ensure-local-inference-handler.d.ts.map +1 -0
- package/src/runtime/ensure-local-inference-handler.test.ts +528 -0
- package/src/runtime/ensure-local-inference-handler.ts +1398 -0
- package/src/runtime/index.d.ts +14 -0
- package/src/runtime/index.d.ts.map +1 -0
- package/src/runtime/index.ts +27 -0
- package/src/runtime/mobile-local-inference-gate.d.ts +31 -0
- package/src/runtime/mobile-local-inference-gate.d.ts.map +1 -0
- package/src/runtime/mobile-local-inference-gate.test.ts +69 -0
- package/src/runtime/mobile-local-inference-gate.ts +44 -0
- package/src/runtime/voice-entity-binding.d.ts +103 -0
- package/src/runtime/voice-entity-binding.d.ts.map +1 -0
- package/src/runtime/voice-entity-binding.transcript.test.ts +69 -0
- package/src/runtime/voice-entity-binding.ts +328 -0
- package/src/services/README.md +71 -0
- package/src/services/__tests__/backend-selector.test.ts +101 -0
- package/src/services/__tests__/checkpoint-manager.test.ts +376 -0
- package/src/services/__tests__/gpu-autotune.test.ts +400 -0
- package/src/services/__tests__/llm-streaming-binding.test.ts +85 -0
- package/src/services/__tests__/planner-grammar.test.ts +372 -0
- package/src/services/__tests__/runtime-target.test.ts +176 -0
- package/src/services/active-model-switch-rollback.test.ts +183 -0
- package/src/services/active-model.d.ts +282 -0
- package/src/services/active-model.d.ts.map +1 -0
- package/src/services/active-model.ts +1213 -0
- package/src/services/asr/errors.d.ts +21 -0
- package/src/services/asr/errors.d.ts.map +1 -0
- package/src/services/asr/errors.ts +50 -0
- package/src/services/asr/hash.d.ts +28 -0
- package/src/services/asr/hash.d.ts.map +1 -0
- package/src/services/asr/hash.ts +49 -0
- package/src/services/asr/index.d.ts +76 -0
- package/src/services/asr/index.d.ts.map +1 -0
- package/src/services/asr/index.ts +178 -0
- package/src/services/asr/types.d.ts +91 -0
- package/src/services/asr/types.d.ts.map +1 -0
- package/src/services/asr/types.ts +95 -0
- package/src/services/assignments.d.ts +71 -0
- package/src/services/assignments.d.ts.map +1 -0
- package/src/services/assignments.test.ts +80 -0
- package/src/services/assignments.ts +230 -0
- package/src/services/backend-selector.ts +95 -0
- package/src/services/backend.d.ts +346 -0
- package/src/services/backend.d.ts.map +1 -0
- package/src/services/backend.ts +612 -0
- package/src/services/bundled-models.d.ts +34 -0
- package/src/services/bundled-models.d.ts.map +1 -0
- package/src/services/bundled-models.ts +129 -0
- package/src/services/cache-bridge.d.ts +206 -0
- package/src/services/cache-bridge.d.ts.map +1 -0
- package/src/services/cache-bridge.test.ts +516 -0
- package/src/services/cache-bridge.ts +423 -0
- package/src/services/catalog.d.ts +10 -0
- package/src/services/catalog.d.ts.map +1 -0
- package/src/services/catalog.test.ts +240 -0
- package/src/services/catalog.ts +27 -0
- package/src/services/checkpoint-client.d.ts +109 -0
- package/src/services/checkpoint-client.d.ts.map +1 -0
- package/src/services/checkpoint-client.ts +258 -0
- package/src/services/checkpoint-manager.ts +474 -0
- package/src/services/cloud-fallback.d.ts +102 -0
- package/src/services/cloud-fallback.d.ts.map +1 -0
- package/src/services/cloud-fallback.ts +230 -0
- package/src/services/conversation-registry.d.ts +142 -0
- package/src/services/conversation-registry.d.ts.map +1 -0
- package/src/services/conversation-registry.test.ts +235 -0
- package/src/services/conversation-registry.ts +264 -0
- package/src/services/desktop-fused-ffi-backend-runtime.d.ts +92 -0
- package/src/services/desktop-fused-ffi-backend-runtime.d.ts.map +1 -0
- package/src/services/desktop-fused-ffi-backend-runtime.ts +333 -0
- package/src/services/device-bridge.d.ts +188 -0
- package/src/services/device-bridge.d.ts.map +1 -0
- package/src/services/device-bridge.ts +1237 -0
- package/src/services/device-resource-metrics.d.ts +149 -0
- package/src/services/device-resource-metrics.d.ts.map +1 -0
- package/src/services/device-resource-metrics.test.ts +98 -0
- package/src/services/device-resource-metrics.ts +346 -0
- package/src/services/device-tier.d.ts +115 -0
- package/src/services/device-tier.d.ts.map +1 -0
- package/src/services/device-tier.test.ts +371 -0
- package/src/services/device-tier.ts +410 -0
- package/src/services/downloader.d.ts +82 -0
- package/src/services/downloader.d.ts.map +1 -0
- package/src/services/downloader.test.ts +724 -0
- package/src/services/downloader.ts +899 -0
- package/src/services/engine-direct-bundle.test.ts +58 -0
- package/src/services/engine-streaming.test.ts +80 -0
- package/src/services/engine.d.ts +534 -0
- package/src/services/engine.d.ts.map +1 -0
- package/src/services/engine.ts +1891 -0
- package/src/services/ensure-local-artifacts.integration.test.ts +273 -0
- package/src/services/ensure-local-artifacts.test.ts +368 -0
- package/src/services/ensure-local-artifacts.ts +351 -0
- package/src/services/external-scanner.d.ts +17 -0
- package/src/services/external-scanner.d.ts.map +1 -0
- package/src/services/external-scanner.ts +312 -0
- package/src/services/ffi-llm-mock.ts +354 -0
- package/src/services/ffi-llm-streaming-abi.ts +442 -0
- package/src/services/ffi-streaming-backend.d.ts +180 -0
- package/src/services/ffi-streaming-backend.d.ts.map +1 -0
- package/src/services/ffi-streaming-backend.ts +382 -0
- package/src/services/ffi-streaming-runner.d.ts +122 -0
- package/src/services/ffi-streaming-runner.d.ts.map +1 -0
- package/src/services/ffi-streaming-runner.test.ts +60 -0
- package/src/services/ffi-streaming-runner.ts +354 -0
- package/src/services/ffi-unload-ordering.test.ts +162 -0
- package/src/services/gpu-autotune.ts +534 -0
- package/src/services/gpu-detect.ts +139 -0
- package/src/services/handler-registry.d.ts +72 -0
- package/src/services/handler-registry.d.ts.map +1 -0
- package/src/services/handler-registry.ts +240 -0
- package/src/services/hardware.d.ts +63 -0
- package/src/services/hardware.d.ts.map +1 -0
- package/src/services/hardware.test.ts +183 -0
- package/src/services/hardware.ts +404 -0
- package/src/services/hf-search.d.ts +26 -0
- package/src/services/hf-search.d.ts.map +1 -0
- package/src/services/hf-search.test.ts +69 -0
- package/src/services/hf-search.ts +420 -0
- package/src/services/image-description-runtime.d.ts +14 -0
- package/src/services/image-description-runtime.d.ts.map +1 -0
- package/src/services/image-description-runtime.test.ts +61 -0
- package/src/services/image-description-runtime.ts +118 -0
- package/src/services/imagegen/aosp-unavailable.d.ts +134 -0
- package/src/services/imagegen/aosp-unavailable.d.ts.map +1 -0
- package/src/services/imagegen/aosp-unavailable.ts +229 -0
- package/src/services/imagegen/backend-selector.d.ts +118 -0
- package/src/services/imagegen/backend-selector.d.ts.map +1 -0
- package/src/services/imagegen/backend-selector.ts +281 -0
- package/src/services/imagegen/coreml-unavailable.d.ts +105 -0
- package/src/services/imagegen/coreml-unavailable.d.ts.map +1 -0
- package/src/services/imagegen/coreml-unavailable.ts +237 -0
- package/src/services/imagegen/errors.d.ts +16 -0
- package/src/services/imagegen/errors.d.ts.map +1 -0
- package/src/services/imagegen/errors.ts +40 -0
- package/src/services/imagegen/index.d.ts +58 -0
- package/src/services/imagegen/index.d.ts.map +1 -0
- package/src/services/imagegen/index.ts +144 -0
- package/src/services/imagegen/mflux.d.ts +74 -0
- package/src/services/imagegen/mflux.d.ts.map +1 -0
- package/src/services/imagegen/mflux.ts +313 -0
- package/src/services/imagegen/sd-cpp.d.ts +180 -0
- package/src/services/imagegen/sd-cpp.d.ts.map +1 -0
- package/src/services/imagegen/sd-cpp.ts +718 -0
- package/src/services/imagegen/tensorrt-unavailable.d.ts +83 -0
- package/src/services/imagegen/tensorrt-unavailable.d.ts.map +1 -0
- package/src/services/imagegen/tensorrt-unavailable.ts +295 -0
- package/src/services/imagegen/types.d.ts +181 -0
- package/src/services/imagegen/types.d.ts.map +1 -0
- package/src/services/imagegen/types.ts +193 -0
- package/src/services/index.d.ts +30 -0
- package/src/services/index.d.ts.map +1 -0
- package/src/services/index.ts +225 -0
- package/src/services/inference-capabilities.d.ts +132 -0
- package/src/services/inference-capabilities.d.ts.map +1 -0
- package/src/services/inference-capabilities.test.ts +75 -0
- package/src/services/inference-capabilities.ts +204 -0
- package/src/services/inference-telemetry.d.ts +59 -0
- package/src/services/inference-telemetry.d.ts.map +1 -0
- package/src/services/inference-telemetry.ts +143 -0
- package/src/services/ios-llama-streaming.ts +248 -0
- package/src/services/kv-spill.d.ts +189 -0
- package/src/services/kv-spill.d.ts.map +1 -0
- package/src/services/kv-spill.test.ts +222 -0
- package/src/services/kv-spill.ts +356 -0
- package/src/services/latency-trace.d.ts +346 -0
- package/src/services/latency-trace.d.ts.map +1 -0
- package/src/services/latency-trace.test.ts +266 -0
- package/src/services/latency-trace.ts +844 -0
- package/src/services/llama-server-metrics.ts +304 -0
- package/src/services/llm-streaming-binding.d.ts +96 -0
- package/src/services/llm-streaming-binding.d.ts.map +1 -0
- package/src/services/llm-streaming-binding.ts +136 -0
- package/src/services/load-args.d.ts +82 -0
- package/src/services/load-args.d.ts.map +1 -0
- package/src/services/load-args.ts +81 -0
- package/src/services/manifest/eliza-1.manifest.v1.json +708 -0
- package/src/services/manifest/index.d.ts +4 -0
- package/src/services/manifest/index.d.ts.map +1 -0
- package/src/services/manifest/index.ts +66 -0
- package/src/services/manifest/manifest.test.ts +693 -0
- package/src/services/manifest/schema.d.ts +715 -0
- package/src/services/manifest/schema.d.ts.map +1 -0
- package/src/services/manifest/schema.ts +655 -0
- package/src/services/manifest/types.d.ts +30 -0
- package/src/services/manifest/types.d.ts.map +1 -0
- package/src/services/manifest/types.ts +55 -0
- package/src/services/manifest/validator.d.ts +66 -0
- package/src/services/manifest/validator.d.ts.map +1 -0
- package/src/services/manifest/validator.ts +569 -0
- package/src/services/memory-arbiter.d.ts +343 -0
- package/src/services/memory-arbiter.d.ts.map +1 -0
- package/src/services/memory-arbiter.test.ts +419 -0
- package/src/services/memory-arbiter.ts +1000 -0
- package/src/services/memory-monitor.d.ts +119 -0
- package/src/services/memory-monitor.d.ts.map +1 -0
- package/src/services/memory-monitor.test.ts +208 -0
- package/src/services/memory-monitor.ts +296 -0
- package/src/services/memory-pressure.d.ts +127 -0
- package/src/services/memory-pressure.d.ts.map +1 -0
- package/src/services/memory-pressure.ts +413 -0
- package/src/services/mtp-doctor.d.ts +13 -0
- package/src/services/mtp-doctor.d.ts.map +1 -0
- package/src/services/mtp-doctor.ts +78 -0
- package/src/services/network-policy.d.ts +127 -0
- package/src/services/network-policy.d.ts.map +1 -0
- package/src/services/network-policy.ts +346 -0
- package/src/services/paths.d.ts +6 -0
- package/src/services/paths.d.ts.map +1 -0
- package/src/services/paths.ts +25 -0
- package/src/services/planner-skeleton.d.ts +124 -0
- package/src/services/planner-skeleton.d.ts.map +1 -0
- package/src/services/planner-skeleton.ts +175 -0
- package/src/services/providers.d.ts +38 -0
- package/src/services/providers.d.ts.map +1 -0
- package/src/services/providers.ts +507 -0
- package/src/services/ram-budget-cache.test.ts +163 -0
- package/src/services/ram-budget.d.ts +110 -0
- package/src/services/ram-budget.d.ts.map +1 -0
- package/src/services/ram-budget.ts +0 -0
- package/src/services/readiness.d.ts +9 -0
- package/src/services/readiness.d.ts.map +1 -0
- package/src/services/readiness.test.ts +87 -0
- package/src/services/readiness.ts +238 -0
- package/src/services/recommendation.d.ts +111 -0
- package/src/services/recommendation.d.ts.map +1 -0
- package/src/services/recommendation.ts +672 -0
- package/src/services/registry.d.ts +35 -0
- package/src/services/registry.d.ts.map +1 -0
- package/src/services/registry.ts +151 -0
- package/src/services/router-handler.d.ts +92 -0
- package/src/services/router-handler.d.ts.map +1 -0
- package/src/services/router-handler.test.ts +45 -0
- package/src/services/router-handler.ts +376 -0
- package/src/services/routing-policy.d.ts +55 -0
- package/src/services/routing-policy.d.ts.map +1 -0
- package/src/services/routing-policy.ts +228 -0
- package/src/services/routing-preferences.d.ts +8 -0
- package/src/services/routing-preferences.d.ts.map +1 -0
- package/src/services/routing-preferences.ts +15 -0
- package/src/services/runtime-target.d.ts +98 -0
- package/src/services/runtime-target.d.ts.map +1 -0
- package/src/services/runtime-target.ts +154 -0
- package/src/services/service.d.ts +128 -0
- package/src/services/service.d.ts.map +1 -0
- package/src/services/service.test.ts +223 -0
- package/src/services/service.ts +735 -0
- package/src/services/session-pool.d.ts +72 -0
- package/src/services/session-pool.d.ts.map +1 -0
- package/src/services/session-pool.ts +153 -0
- package/src/services/structured-output/deterministic-repair.d.ts +23 -0
- package/src/services/structured-output/deterministic-repair.d.ts.map +1 -0
- package/src/services/structured-output/deterministic-repair.test.ts +169 -0
- package/src/services/structured-output/deterministic-repair.ts +443 -0
- package/src/services/structured-output/index.ts +4 -0
- package/src/services/structured-output.d.ts +311 -0
- package/src/services/structured-output.d.ts.map +1 -0
- package/src/services/structured-output.test.ts +483 -0
- package/src/services/structured-output.ts +712 -0
- package/src/services/transcription-priority.test.ts +211 -0
- package/src/services/tts/errors.ts +46 -0
- package/src/services/tts/index.ts +214 -0
- package/src/services/tts/tts-audio-cache.ts +235 -0
- package/src/services/tts/types.ts +157 -0
- package/src/services/types.d.ts +19 -0
- package/src/services/types.d.ts.map +1 -0
- package/src/services/types.ts +55 -0
- package/src/services/verify-on-device.d.ts +34 -0
- package/src/services/verify-on-device.d.ts.map +1 -0
- package/src/services/verify-on-device.test.ts +87 -0
- package/src/services/verify-on-device.ts +127 -0
- package/src/services/verify.d.ts +8 -0
- package/src/services/verify.d.ts.map +1 -0
- package/src/services/verify.ts +13 -0
- package/src/services/vision/aosp-unavailable.d.ts +115 -0
- package/src/services/vision/aosp-unavailable.d.ts.map +1 -0
- package/src/services/vision/aosp-unavailable.ts +163 -0
- package/src/services/vision/capacitor-llama.d.ts +99 -0
- package/src/services/vision/capacitor-llama.d.ts.map +1 -0
- package/src/services/vision/capacitor-llama.ts +255 -0
- package/src/services/vision/cloud-fallback.d.ts +47 -0
- package/src/services/vision/cloud-fallback.d.ts.map +1 -0
- package/src/services/vision/cloud-fallback.test.ts +243 -0
- package/src/services/vision/cloud-fallback.ts +268 -0
- package/src/services/vision/fallback-chain.test.ts +86 -0
- package/src/services/vision/hash.d.ts +71 -0
- package/src/services/vision/hash.d.ts.map +1 -0
- package/src/services/vision/hash.ts +157 -0
- package/src/services/vision/index.d.ts +95 -0
- package/src/services/vision/index.d.ts.map +1 -0
- package/src/services/vision/index.ts +251 -0
- package/src/services/vision/llama-server.d.ts +73 -0
- package/src/services/vision/llama-server.d.ts.map +1 -0
- package/src/services/vision/llama-server.ts +177 -0
- package/src/services/vision/types.d.ts +153 -0
- package/src/services/vision/types.d.ts.map +1 -0
- package/src/services/vision/types.ts +154 -0
- package/src/services/vision/vast-fallback.d.ts +18 -0
- package/src/services/vision/vast-fallback.d.ts.map +1 -0
- package/src/services/vision/vast-fallback.ts +127 -0
- package/src/services/vision-embedding-cache.d.ts +98 -0
- package/src/services/vision-embedding-cache.d.ts.map +1 -0
- package/src/services/vision-embedding-cache.ts +189 -0
- package/src/services/voice/VOICE_WORKBENCH.md +88 -0
- package/src/services/voice/__test-helpers__/fake-ffi.ts +92 -0
- package/src/services/voice/__test-helpers__/synthetic-speech.ts +124 -0
- package/src/services/voice/__tests__/checkpoint-manager.test.ts +241 -0
- package/src/services/voice/__tests__/checkpoint-policy.test.ts +270 -0
- package/src/services/voice/__tests__/eager-context-builder.test.ts +257 -0
- package/src/services/voice/__tests__/eliza1-eot-scorer.test.ts +288 -0
- package/src/services/voice/__tests__/eot-classifier.test.ts +431 -0
- package/src/services/voice/__tests__/optimistic-rollback.test.ts +312 -0
- package/src/services/voice/__tests__/prefill-client.test.ts +266 -0
- package/src/services/voice/__tests__/prefix-preserving-queue.test.ts +208 -0
- package/src/services/voice/__tests__/streaming-asr.test.ts +450 -0
- package/src/services/voice/__tests__/streaming-transcriber.test.ts +339 -0
- package/src/services/voice/__tests__/turn-detector-resolver.test.ts +197 -0
- package/src/services/voice/__tests__/voice-state-machine-prefill.test.ts +275 -0
- package/src/services/voice/__tests__/voice-state-machine.test.ts +354 -0
- package/src/services/voice/audio-frame-consumer.d.ts +212 -0
- package/src/services/voice/audio-frame-consumer.d.ts.map +1 -0
- package/src/services/voice/audio-frame-consumer.test.ts +343 -0
- package/src/services/voice/audio-frame-consumer.ts +491 -0
- package/src/services/voice/barge-in.d.ts +112 -0
- package/src/services/voice/barge-in.d.ts.map +1 -0
- package/src/services/voice/barge-in.test.ts +244 -0
- package/src/services/voice/barge-in.ts +336 -0
- package/src/services/voice/cancellation-coordinator.d.ts +127 -0
- package/src/services/voice/cancellation-coordinator.d.ts.map +1 -0
- package/src/services/voice/cancellation-coordinator.test.ts +196 -0
- package/src/services/voice/cancellation-coordinator.ts +269 -0
- package/src/services/voice/checkpoint-manager.d.ts +199 -0
- package/src/services/voice/checkpoint-manager.d.ts.map +1 -0
- package/src/services/voice/checkpoint-manager.ts +401 -0
- package/src/services/voice/checkpoint-policy.ts +336 -0
- package/src/services/voice/composite-eot-classifier.test.ts +59 -0
- package/src/services/voice/e2e-harness.test.ts +182 -0
- package/src/services/voice/e2e-harness.ts +743 -0
- package/src/services/voice/eager-context-builder.d.ts +170 -0
- package/src/services/voice/eager-context-builder.d.ts.map +1 -0
- package/src/services/voice/eager-context-builder.ts +262 -0
- package/src/services/voice/eliza1-eot-scorer.d.ts +124 -0
- package/src/services/voice/eliza1-eot-scorer.d.ts.map +1 -0
- package/src/services/voice/eliza1-eot-scorer.ts +242 -0
- package/src/services/voice/embedding-server.ts +200 -0
- package/src/services/voice/embedding.d.ts +133 -0
- package/src/services/voice/embedding.d.ts.map +1 -0
- package/src/services/voice/embedding.test.ts +148 -0
- package/src/services/voice/embedding.ts +244 -0
- package/src/services/voice/emotion-attribution.d.ts +68 -0
- package/src/services/voice/emotion-attribution.d.ts.map +1 -0
- package/src/services/voice/emotion-attribution.test.ts +129 -0
- package/src/services/voice/emotion-attribution.ts +361 -0
- package/src/services/voice/engine-bridge-cancellation.test.ts +422 -0
- package/src/services/voice/engine-bridge.d.ts +746 -0
- package/src/services/voice/engine-bridge.d.ts.map +1 -0
- package/src/services/voice/engine-bridge.test.ts +384 -0
- package/src/services/voice/engine-bridge.ts +2226 -0
- package/src/services/voice/eot-classifier-ggml.d.ts +179 -0
- package/src/services/voice/eot-classifier-ggml.d.ts.map +1 -0
- package/src/services/voice/eot-classifier-ggml.ts +566 -0
- package/src/services/voice/eot-classifier.d.ts +214 -0
- package/src/services/voice/eot-classifier.d.ts.map +1 -0
- package/src/services/voice/eot-classifier.ts +533 -0
- package/src/services/voice/errors.d.ts +20 -0
- package/src/services/voice/errors.d.ts.map +1 -0
- package/src/services/voice/errors.ts +32 -0
- package/src/services/voice/expressive-tags.d.ts +158 -0
- package/src/services/voice/expressive-tags.d.ts.map +1 -0
- package/src/services/voice/expressive-tags.ts +405 -0
- package/src/services/voice/ffi-bindings.d.ts +636 -0
- package/src/services/voice/ffi-bindings.d.ts.map +1 -0
- package/src/services/voice/ffi-bindings.test.ts +671 -0
- package/src/services/voice/ffi-bindings.ts +3050 -0
- package/src/services/voice/first-line-cache.d.ts +181 -0
- package/src/services/voice/first-line-cache.d.ts.map +1 -0
- package/src/services/voice/first-line-cache.ts +725 -0
- package/src/services/voice/fused-eot-scorer.d.ts +51 -0
- package/src/services/voice/fused-eot-scorer.d.ts.map +1 -0
- package/src/services/voice/fused-eot-scorer.ts +135 -0
- package/src/services/voice/index.d.ts +91 -0
- package/src/services/voice/index.d.ts.map +1 -0
- package/src/services/voice/index.ts +481 -0
- package/src/services/voice/kokoro/__tests__/kokoro-backend.test.ts +151 -0
- package/src/services/voice/kokoro/__tests__/kokoro-engine-bridge.real.test.ts +151 -0
- package/src/services/voice/kokoro/__tests__/kokoro-engine-bridge.test.ts +60 -0
- package/src/services/voice/kokoro/__tests__/kokoro-engine-discovery.test.ts +277 -0
- package/src/services/voice/kokoro/__tests__/kokoro-ffi-runtime.test.ts +235 -0
- package/src/services/voice/kokoro/__tests__/kokoro-runtime.test.ts +95 -0
- package/src/services/voice/kokoro/__tests__/phonemizer.test.ts +53 -0
- package/src/services/voice/kokoro/__tests__/runtime-selection.test.ts +231 -0
- package/src/services/voice/kokoro/__tests__/voices.test.ts +57 -0
- package/src/services/voice/kokoro/index.ts +79 -0
- package/src/services/voice/kokoro/kokoro-backend.d.ts +72 -0
- package/src/services/voice/kokoro/kokoro-backend.d.ts.map +1 -0
- package/src/services/voice/kokoro/kokoro-backend.ts +207 -0
- package/src/services/voice/kokoro/kokoro-engine-discovery.d.ts +58 -0
- package/src/services/voice/kokoro/kokoro-engine-discovery.d.ts.map +1 -0
- package/src/services/voice/kokoro/kokoro-engine-discovery.ts +177 -0
- package/src/services/voice/kokoro/kokoro-ffi-runtime.d.ts +75 -0
- package/src/services/voice/kokoro/kokoro-ffi-runtime.d.ts.map +1 -0
- package/src/services/voice/kokoro/kokoro-ffi-runtime.ts +233 -0
- package/src/services/voice/kokoro/kokoro-runtime.d.ts +100 -0
- package/src/services/voice/kokoro/kokoro-runtime.d.ts.map +1 -0
- package/src/services/voice/kokoro/kokoro-runtime.ts +170 -0
- package/src/services/voice/kokoro/phoneme-stream.ts +123 -0
- package/src/services/voice/kokoro/phonemizer.d.ts +50 -0
- package/src/services/voice/kokoro/phonemizer.d.ts.map +1 -0
- package/src/services/voice/kokoro/phonemizer.ts +344 -0
- package/src/services/voice/kokoro/pick-runtime.d.ts +61 -0
- package/src/services/voice/kokoro/pick-runtime.d.ts.map +1 -0
- package/src/services/voice/kokoro/pick-runtime.test.ts +91 -0
- package/src/services/voice/kokoro/pick-runtime.ts +130 -0
- package/src/services/voice/kokoro/runtime-selection.d.ts +92 -0
- package/src/services/voice/kokoro/runtime-selection.d.ts.map +1 -0
- package/src/services/voice/kokoro/runtime-selection.ts +237 -0
- package/src/services/voice/kokoro/types.d.ts +82 -0
- package/src/services/voice/kokoro/types.d.ts.map +1 -0
- package/src/services/voice/kokoro/types.ts +95 -0
- package/src/services/voice/kokoro/voice-presets.d.ts +23 -0
- package/src/services/voice/kokoro/voice-presets.d.ts.map +1 -0
- package/src/services/voice/kokoro/voice-presets.ts +129 -0
- package/src/services/voice/kokoro/voices.d.ts +30 -0
- package/src/services/voice/kokoro/voices.d.ts.map +1 -0
- package/src/services/voice/kokoro/voices.ts +64 -0
- package/src/services/voice/lifecycle.d.ts +135 -0
- package/src/services/voice/lifecycle.d.ts.map +1 -0
- package/src/services/voice/lifecycle.test.ts +315 -0
- package/src/services/voice/lifecycle.ts +301 -0
- package/src/services/voice/live-diarization-session.d.ts +96 -0
- package/src/services/voice/live-diarization-session.d.ts.map +1 -0
- package/src/services/voice/live-diarization-session.ts +289 -0
- package/src/services/voice/mic-source.d.ts +136 -0
- package/src/services/voice/mic-source.d.ts.map +1 -0
- package/src/services/voice/mic-source.test.ts +210 -0
- package/src/services/voice/mic-source.ts +503 -0
- package/src/services/voice/optimistic-policy.d.ts +109 -0
- package/src/services/voice/optimistic-policy.d.ts.map +1 -0
- package/src/services/voice/optimistic-policy.test.ts +101 -0
- package/src/services/voice/optimistic-policy.ts +192 -0
- package/src/services/voice/optimistic-rollback.ts +343 -0
- package/src/services/voice/partial-stabilizer.d.ts +73 -0
- package/src/services/voice/partial-stabilizer.d.ts.map +1 -0
- package/src/services/voice/partial-stabilizer.test.ts +68 -0
- package/src/services/voice/partial-stabilizer.ts +140 -0
- package/src/services/voice/phoneme-tokenizer.d.ts +49 -0
- package/src/services/voice/phoneme-tokenizer.d.ts.map +1 -0
- package/src/services/voice/phoneme-tokenizer.ts +158 -0
- package/src/services/voice/phrase-cache.d.ts +76 -0
- package/src/services/voice/phrase-cache.d.ts.map +1 -0
- package/src/services/voice/phrase-cache.test.ts +242 -0
- package/src/services/voice/phrase-cache.ts +186 -0
- package/src/services/voice/phrase-chunker.d.ts +62 -0
- package/src/services/voice/phrase-chunker.d.ts.map +1 -0
- package/src/services/voice/phrase-chunker.test.ts +239 -0
- package/src/services/voice/phrase-chunker.ts +281 -0
- package/src/services/voice/pipeline-impls.d.ts +151 -0
- package/src/services/voice/pipeline-impls.d.ts.map +1 -0
- package/src/services/voice/pipeline-impls.l6.test.ts +110 -0
- package/src/services/voice/pipeline-impls.test.ts +292 -0
- package/src/services/voice/pipeline-impls.ts +315 -0
- package/src/services/voice/pipeline.d.ts +216 -0
- package/src/services/voice/pipeline.d.ts.map +1 -0
- package/src/services/voice/pipeline.ts +505 -0
- package/src/services/voice/prefill-client.d.ts +123 -0
- package/src/services/voice/prefill-client.d.ts.map +1 -0
- package/src/services/voice/prefill-client.ts +316 -0
- package/src/services/voice/prefix-preserving-queue.d.ts +113 -0
- package/src/services/voice/prefix-preserving-queue.d.ts.map +1 -0
- package/src/services/voice/prefix-preserving-queue.ts +162 -0
- package/src/services/voice/profile-store.d.ts +248 -0
- package/src/services/voice/profile-store.d.ts.map +1 -0
- package/src/services/voice/profile-store.ts +887 -0
- package/src/services/voice/ring-buffer.d.ts +40 -0
- package/src/services/voice/ring-buffer.d.ts.map +1 -0
- package/src/services/voice/ring-buffer.ts +105 -0
- package/src/services/voice/rollback-queue.d.ts +24 -0
- package/src/services/voice/rollback-queue.d.ts.map +1 -0
- package/src/services/voice/rollback-queue.ts +74 -0
- package/src/services/voice/samantha-preset-placeholder.d.ts +67 -0
- package/src/services/voice/samantha-preset-placeholder.d.ts.map +1 -0
- package/src/services/voice/samantha-preset-placeholder.test.ts +97 -0
- package/src/services/voice/samantha-preset-placeholder.ts +148 -0
- package/src/services/voice/samantha-preset-regenerator.d.ts +87 -0
- package/src/services/voice/samantha-preset-regenerator.d.ts.map +1 -0
- package/src/services/voice/samantha-preset-regenerator.ts +393 -0
- package/src/services/voice/scheduler.d.ts +146 -0
- package/src/services/voice/scheduler.d.ts.map +1 -0
- package/src/services/voice/scheduler.t2.test.ts +141 -0
- package/src/services/voice/scheduler.ts +927 -0
- package/src/services/voice/shared-resources.d.ts +190 -0
- package/src/services/voice/shared-resources.d.ts.map +1 -0
- package/src/services/voice/shared-resources.ts +320 -0
- package/src/services/voice/speaker/attribution-pipeline.d.ts +74 -0
- package/src/services/voice/speaker/attribution-pipeline.d.ts.map +1 -0
- package/src/services/voice/speaker/attribution-pipeline.ts +386 -0
- package/src/services/voice/speaker/diarizer-fused.d.ts +59 -0
- package/src/services/voice/speaker/diarizer-fused.d.ts.map +1 -0
- package/src/services/voice/speaker/diarizer-fused.real.test.ts +100 -0
- package/src/services/voice/speaker/diarizer-fused.ts +154 -0
- package/src/services/voice/speaker/diarizer.d.ts +75 -0
- package/src/services/voice/speaker/diarizer.d.ts.map +1 -0
- package/src/services/voice/speaker/diarizer.ts +218 -0
- package/src/services/voice/speaker/encoder-fused.d.ts +60 -0
- package/src/services/voice/speaker/encoder-fused.d.ts.map +1 -0
- package/src/services/voice/speaker/encoder-fused.real.test.ts +113 -0
- package/src/services/voice/speaker/encoder-fused.ts +138 -0
- package/src/services/voice/speaker/encoder-ggml.d.ts +33 -0
- package/src/services/voice/speaker/encoder-ggml.d.ts.map +1 -0
- package/src/services/voice/speaker/encoder-ggml.ts +79 -0
- package/src/services/voice/speaker/encoder.d.ts +37 -0
- package/src/services/voice/speaker/encoder.d.ts.map +1 -0
- package/src/services/voice/speaker/encoder.ts +105 -0
- package/src/services/voice/speaker-imprint.d.ts +83 -0
- package/src/services/voice/speaker-imprint.d.ts.map +1 -0
- package/src/services/voice/speaker-imprint.test.ts +185 -0
- package/src/services/voice/speaker-imprint.ts +312 -0
- package/src/services/voice/speaker-preset-cache.d.ts +77 -0
- package/src/services/voice/speaker-preset-cache.d.ts.map +1 -0
- package/src/services/voice/speaker-preset-cache.test.ts +154 -0
- package/src/services/voice/speaker-preset-cache.ts +195 -0
- package/src/services/voice/streaming-asr/streaming-pipeline-adapter.ts +292 -0
- package/src/services/voice/system-audio-sink.d.ts +73 -0
- package/src/services/voice/system-audio-sink.d.ts.map +1 -0
- package/src/services/voice/system-audio-sink.test.ts +29 -0
- package/src/services/voice/system-audio-sink.ts +366 -0
- package/src/services/voice/transcriber.d.ts +244 -0
- package/src/services/voice/transcriber.d.ts.map +1 -0
- package/src/services/voice/transcriber.test.ts +392 -0
- package/src/services/voice/transcriber.ts +704 -0
- package/src/services/voice/turn-controller.d.ts +183 -0
- package/src/services/voice/turn-controller.d.ts.map +1 -0
- package/src/services/voice/turn-controller.test.ts +575 -0
- package/src/services/voice/turn-controller.ts +596 -0
- package/src/services/voice/types.d.ts +643 -0
- package/src/services/voice/types.d.ts.map +1 -0
- package/src/services/voice/types.ts +699 -0
- package/src/services/voice/vad.d.ts +282 -0
- package/src/services/voice/vad.d.ts.map +1 -0
- package/src/services/voice/vad.test.ts +480 -0
- package/src/services/voice/vad.ts +827 -0
- package/src/services/voice/vad.v1-v4.test.ts +222 -0
- package/src/services/voice/voice-budget.d.ts +241 -0
- package/src/services/voice/voice-budget.d.ts.map +1 -0
- package/src/services/voice/voice-budget.test.ts +420 -0
- package/src/services/voice/voice-budget.ts +656 -0
- package/src/services/voice/voice-duet.test.ts +375 -0
- package/src/services/voice/voice-emotion-classifier.d.ts +95 -0
- package/src/services/voice/voice-emotion-classifier.d.ts.map +1 -0
- package/src/services/voice/voice-emotion-classifier.test.ts +210 -0
- package/src/services/voice/voice-emotion-classifier.ts +273 -0
- package/src/services/voice/voice-preset-format.d.ts +158 -0
- package/src/services/voice/voice-preset-format.d.ts.map +1 -0
- package/src/services/voice/voice-preset-format.ts +700 -0
- package/src/services/voice/voice-preset-generator.test.ts +89 -0
- package/src/services/voice/voice-profile-artifact.d.ts +116 -0
- package/src/services/voice/voice-profile-artifact.d.ts.map +1 -0
- package/src/services/voice/voice-profile-artifact.test.ts +138 -0
- package/src/services/voice/voice-profile-artifact.ts +518 -0
- package/src/services/voice/voice-profile-routes.d.ts +83 -0
- package/src/services/voice/voice-profile-routes.d.ts.map +1 -0
- package/src/services/voice/voice-profile-routes.test.ts +429 -0
- package/src/services/voice/voice-profile-routes.ts +425 -0
- package/src/services/voice/voice-scenario.ts +154 -0
- package/src/services/voice/voice-settings.d.ts +82 -0
- package/src/services/voice/voice-settings.d.ts.map +1 -0
- package/src/services/voice/voice-settings.ts +172 -0
- package/src/services/voice/voice-state-machine.d.ts +364 -0
- package/src/services/voice/voice-state-machine.d.ts.map +1 -0
- package/src/services/voice/voice-state-machine.ts +727 -0
- package/src/services/voice/voice-workbench-report.test.ts +168 -0
- package/src/services/voice/voice-workbench-report.ts +326 -0
- package/src/services/voice/voice-workbench.test.ts +158 -0
- package/src/services/voice/voice.test.ts +1070 -0
- package/src/services/voice/wake-word-ggml.d.ts +101 -0
- package/src/services/voice/wake-word-ggml.d.ts.map +1 -0
- package/src/services/voice/wake-word-ggml.ts +320 -0
- package/src/services/voice/wake-word.d.ts +255 -0
- package/src/services/voice/wake-word.d.ts.map +1 -0
- package/src/services/voice/wake-word.test.ts +298 -0
- package/src/services/voice/wake-word.ts +554 -0
- package/src/services/voice/wrap-with-first-line-cache.d.ts +70 -0
- package/src/services/voice/wrap-with-first-line-cache.d.ts.map +1 -0
- package/src/services/voice/wrap-with-first-line-cache.ts +267 -0
- package/src/services/voice-model-updater.d.ts +240 -0
- package/src/services/voice-model-updater.d.ts.map +1 -0
- package/src/services/voice-model-updater.ts +724 -0
- package/src/services/voice-prewarm.d.ts +3 -0
- package/src/services/voice-prewarm.d.ts.map +1 -0
- package/src/services/voice-prewarm.ts +51 -0
- package/dist/index.d.ts +0 -37
- package/dist/index.js +0 -1098
|
@@ -0,0 +1,72 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Session pool for the in-process node-llama-cpp engine.
|
|
3
|
+
*
|
|
4
|
+
* node-llama-cpp's `LlamaChatSession` keeps an internal KV cache. The
|
|
5
|
+
* stock engine code resets that cache between every turn, which is
|
|
6
|
+
* correct for stateless per-call generation but actively defeats prefix
|
|
7
|
+
* reuse for callers that pass a `promptCacheKey` (the runtime's
|
|
8
|
+
* cloud-style cache hint).
|
|
9
|
+
*
|
|
10
|
+
* This pool keeps one session per `promptCacheKey`, LRU-evicted, so
|
|
11
|
+
* sequential calls with the same key reuse the on-GPU/on-CPU KV cache.
|
|
12
|
+
* Calls without a cache key share the synthetic `_default` slot, which
|
|
13
|
+
* preserves the previous "stateless per-call" semantics by resetting
|
|
14
|
+
* before each turn.
|
|
15
|
+
*
|
|
16
|
+
* The pool owns nothing at module scope; the engine constructs a pool
|
|
17
|
+
* tied to the loaded model and disposes it on unload.
|
|
18
|
+
*/
|
|
19
|
+
export interface PoolSession {
|
|
20
|
+
/** Reset accumulated chat history. Called for the default slot only. */
|
|
21
|
+
resetChatHistory?(): void | Promise<void>;
|
|
22
|
+
/** Dispose underlying KV state. Called on eviction + on pool close. */
|
|
23
|
+
dispose?(): void | Promise<void>;
|
|
24
|
+
}
|
|
25
|
+
export type SessionFactory<TSession extends PoolSession> = (key: string) => Promise<TSession>;
|
|
26
|
+
/**
|
|
27
|
+
* Synthetic key used for callers that didn't supply a `promptCacheKey`.
|
|
28
|
+
* These callers want the old "history-free" behaviour, so the engine
|
|
29
|
+
* resets chat history each turn for this slot only.
|
|
30
|
+
*/
|
|
31
|
+
export declare const DEFAULT_SESSION_KEY = "_default";
|
|
32
|
+
export declare class SessionPool<TSession extends PoolSession> {
|
|
33
|
+
private readonly maxSize;
|
|
34
|
+
private readonly factory;
|
|
35
|
+
/**
|
|
36
|
+
* Insertion order = LRU order. We re-key on each access so the most
|
|
37
|
+
* recently used entry is always last in iteration order.
|
|
38
|
+
*/
|
|
39
|
+
private readonly entries;
|
|
40
|
+
constructor(args: {
|
|
41
|
+
maxSize: number;
|
|
42
|
+
factory: SessionFactory<TSession>;
|
|
43
|
+
});
|
|
44
|
+
/**
|
|
45
|
+
* Get-or-create the session for `key`. Promotes the entry to MRU.
|
|
46
|
+
* On eviction, the oldest entry's `dispose()` is awaited before the
|
|
47
|
+
* new entry is returned so the caller never holds two live sessions
|
|
48
|
+
* over the same KV memory.
|
|
49
|
+
*/
|
|
50
|
+
acquire(key: string): Promise<TSession>;
|
|
51
|
+
/** Number of live sessions, for diagnostics. */
|
|
52
|
+
size(): number;
|
|
53
|
+
/** Snapshot of live keys ordered LRU → MRU. */
|
|
54
|
+
keys(): string[];
|
|
55
|
+
/**
|
|
56
|
+
* Drop a single session by key. Used when the caller knows the prefix
|
|
57
|
+
* has gone stale (e.g. system prompt changed) and the cached KV is no
|
|
58
|
+
* longer valid.
|
|
59
|
+
*/
|
|
60
|
+
drop(key: string): Promise<void>;
|
|
61
|
+
/**
|
|
62
|
+
* Tear down every cached session. Called by the engine on model
|
|
63
|
+
* unload. After `close()` the pool is empty but reusable.
|
|
64
|
+
*/
|
|
65
|
+
close(): Promise<void>;
|
|
66
|
+
private disposeQuietly;
|
|
67
|
+
}
|
|
68
|
+
/**
|
|
69
|
+
* Resolve the pool size from env, with a sane default. Bound by 1..64.
|
|
70
|
+
*/
|
|
71
|
+
export declare function resolveDefaultPoolSize(envValue?: string | null): number;
|
|
72
|
+
//# sourceMappingURL=session-pool.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"session-pool.d.ts","sourceRoot":"","sources":["session-pool.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAEH,MAAM,WAAW,WAAW;IAC3B,wEAAwE;IACxE,gBAAgB,CAAC,IAAI,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IAC1C,uEAAuE;IACvE,OAAO,CAAC,IAAI,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;CACjC;AAED,MAAM,MAAM,cAAc,CAAC,QAAQ,SAAS,WAAW,IAAI,CAC1D,GAAG,EAAE,MAAM,KACP,OAAO,CAAC,QAAQ,CAAC,CAAC;AASvB;;;;GAIG;AACH,eAAO,MAAM,mBAAmB,aAAa,CAAC;AAE9C,qBAAa,WAAW,CAAC,QAAQ,SAAS,WAAW;IACpD,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAS;IACjC,OAAO,CAAC,QAAQ,CAAC,OAAO,CAA2B;IACnD;;;OAGG;IACH,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAsC;gBAElD,IAAI,EAAE;QACjB,OAAO,EAAE,MAAM,CAAC;QAChB,OAAO,EAAE,cAAc,CAAC,QAAQ,CAAC,CAAC;KAClC;IAUD;;;;;OAKG;IACG,OAAO,CAAC,GAAG,EAAE,MAAM,GAAG,OAAO,CAAC,QAAQ,CAAC;IA0B7C,gDAAgD;IAChD,IAAI,IAAI,MAAM;IAId,+CAA+C;IAC/C,IAAI,IAAI,MAAM,EAAE;IAIhB;;;;OAIG;IACG,IAAI,CAAC,GAAG,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC;IAOtC;;;OAGG;IACG,KAAK,IAAI,OAAO,CAAC,IAAI,CAAC;YAQd,cAAc;CAS5B;AAED;;GAEG;AACH,wBAAgB,sBAAsB,CAAC,QAAQ,CAAC,EAAE,MAAM,GAAG,IAAI,GAAG,MAAM,CAMvE"}
|
|
@@ -0,0 +1,153 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Session pool for the in-process node-llama-cpp engine.
|
|
3
|
+
*
|
|
4
|
+
* node-llama-cpp's `LlamaChatSession` keeps an internal KV cache. The
|
|
5
|
+
* stock engine code resets that cache between every turn, which is
|
|
6
|
+
* correct for stateless per-call generation but actively defeats prefix
|
|
7
|
+
* reuse for callers that pass a `promptCacheKey` (the runtime's
|
|
8
|
+
* cloud-style cache hint).
|
|
9
|
+
*
|
|
10
|
+
* This pool keeps one session per `promptCacheKey`, LRU-evicted, so
|
|
11
|
+
* sequential calls with the same key reuse the on-GPU/on-CPU KV cache.
|
|
12
|
+
* Calls without a cache key share the synthetic `_default` slot, which
|
|
13
|
+
* preserves the previous "stateless per-call" semantics by resetting
|
|
14
|
+
* before each turn.
|
|
15
|
+
*
|
|
16
|
+
* The pool owns nothing at module scope; the engine constructs a pool
|
|
17
|
+
* tied to the loaded model and disposes it on unload.
|
|
18
|
+
*/
|
|
19
|
+
|
|
20
|
+
export interface PoolSession {
|
|
21
|
+
/** Reset accumulated chat history. Called for the default slot only. */
|
|
22
|
+
resetChatHistory?(): void | Promise<void>;
|
|
23
|
+
/** Dispose underlying KV state. Called on eviction + on pool close. */
|
|
24
|
+
dispose?(): void | Promise<void>;
|
|
25
|
+
}
|
|
26
|
+
|
|
27
|
+
export type SessionFactory<TSession extends PoolSession> = (
|
|
28
|
+
key: string,
|
|
29
|
+
) => Promise<TSession>;
|
|
30
|
+
|
|
31
|
+
interface Entry<TSession extends PoolSession> {
|
|
32
|
+
key: string;
|
|
33
|
+
session: TSession;
|
|
34
|
+
/** Wall-clock ms of last access. Used purely to break LRU ties. */
|
|
35
|
+
lastUsedMs: number;
|
|
36
|
+
}
|
|
37
|
+
|
|
38
|
+
/**
|
|
39
|
+
* Synthetic key used for callers that didn't supply a `promptCacheKey`.
|
|
40
|
+
* These callers want the old "history-free" behaviour, so the engine
|
|
41
|
+
* resets chat history each turn for this slot only.
|
|
42
|
+
*/
|
|
43
|
+
export const DEFAULT_SESSION_KEY = "_default";
|
|
44
|
+
|
|
45
|
+
export class SessionPool<TSession extends PoolSession> {
|
|
46
|
+
private readonly maxSize: number;
|
|
47
|
+
private readonly factory: SessionFactory<TSession>;
|
|
48
|
+
/**
|
|
49
|
+
* Insertion order = LRU order. We re-key on each access so the most
|
|
50
|
+
* recently used entry is always last in iteration order.
|
|
51
|
+
*/
|
|
52
|
+
private readonly entries = new Map<string, Entry<TSession>>();
|
|
53
|
+
|
|
54
|
+
constructor(args: {
|
|
55
|
+
maxSize: number;
|
|
56
|
+
factory: SessionFactory<TSession>;
|
|
57
|
+
}) {
|
|
58
|
+
if (!Number.isFinite(args.maxSize) || args.maxSize < 1) {
|
|
59
|
+
throw new Error(
|
|
60
|
+
`[session-pool] maxSize must be >= 1, got ${args.maxSize}`,
|
|
61
|
+
);
|
|
62
|
+
}
|
|
63
|
+
this.maxSize = Math.floor(args.maxSize);
|
|
64
|
+
this.factory = args.factory;
|
|
65
|
+
}
|
|
66
|
+
|
|
67
|
+
/**
|
|
68
|
+
* Get-or-create the session for `key`. Promotes the entry to MRU.
|
|
69
|
+
* On eviction, the oldest entry's `dispose()` is awaited before the
|
|
70
|
+
* new entry is returned so the caller never holds two live sessions
|
|
71
|
+
* over the same KV memory.
|
|
72
|
+
*/
|
|
73
|
+
async acquire(key: string): Promise<TSession> {
|
|
74
|
+
const existing = this.entries.get(key);
|
|
75
|
+
if (existing) {
|
|
76
|
+
this.entries.delete(key);
|
|
77
|
+
existing.lastUsedMs = Date.now();
|
|
78
|
+
this.entries.set(key, existing);
|
|
79
|
+
return existing.session;
|
|
80
|
+
}
|
|
81
|
+
|
|
82
|
+
while (this.entries.size >= this.maxSize) {
|
|
83
|
+
const oldestKey = this.entries.keys().next().value as string | undefined;
|
|
84
|
+
if (oldestKey === undefined) break;
|
|
85
|
+
const oldest = this.entries.get(oldestKey);
|
|
86
|
+
this.entries.delete(oldestKey);
|
|
87
|
+
if (oldest) await this.disposeQuietly(oldest.session);
|
|
88
|
+
}
|
|
89
|
+
|
|
90
|
+
const session = await this.factory(key);
|
|
91
|
+
this.entries.set(key, {
|
|
92
|
+
key,
|
|
93
|
+
session,
|
|
94
|
+
lastUsedMs: Date.now(),
|
|
95
|
+
});
|
|
96
|
+
return session;
|
|
97
|
+
}
|
|
98
|
+
|
|
99
|
+
/** Number of live sessions, for diagnostics. */
|
|
100
|
+
size(): number {
|
|
101
|
+
return this.entries.size;
|
|
102
|
+
}
|
|
103
|
+
|
|
104
|
+
/** Snapshot of live keys ordered LRU → MRU. */
|
|
105
|
+
keys(): string[] {
|
|
106
|
+
return [...this.entries.keys()];
|
|
107
|
+
}
|
|
108
|
+
|
|
109
|
+
/**
|
|
110
|
+
* Drop a single session by key. Used when the caller knows the prefix
|
|
111
|
+
* has gone stale (e.g. system prompt changed) and the cached KV is no
|
|
112
|
+
* longer valid.
|
|
113
|
+
*/
|
|
114
|
+
async drop(key: string): Promise<void> {
|
|
115
|
+
const entry = this.entries.get(key);
|
|
116
|
+
if (!entry) return;
|
|
117
|
+
this.entries.delete(key);
|
|
118
|
+
await this.disposeQuietly(entry.session);
|
|
119
|
+
}
|
|
120
|
+
|
|
121
|
+
/**
|
|
122
|
+
* Tear down every cached session. Called by the engine on model
|
|
123
|
+
* unload. After `close()` the pool is empty but reusable.
|
|
124
|
+
*/
|
|
125
|
+
async close(): Promise<void> {
|
|
126
|
+
const entries = [...this.entries.values()];
|
|
127
|
+
this.entries.clear();
|
|
128
|
+
for (const entry of entries) {
|
|
129
|
+
await this.disposeQuietly(entry.session);
|
|
130
|
+
}
|
|
131
|
+
}
|
|
132
|
+
|
|
133
|
+
private async disposeQuietly(session: TSession): Promise<void> {
|
|
134
|
+
if (typeof session.dispose !== "function") return;
|
|
135
|
+
try {
|
|
136
|
+
await session.dispose();
|
|
137
|
+
} catch {
|
|
138
|
+
// Eviction is best-effort: if the underlying binding throws, we still
|
|
139
|
+
// need the slot freed in the pool.
|
|
140
|
+
}
|
|
141
|
+
}
|
|
142
|
+
}
|
|
143
|
+
|
|
144
|
+
/**
|
|
145
|
+
* Resolve the pool size from env, with a sane default. Bound by 1..64.
|
|
146
|
+
*/
|
|
147
|
+
export function resolveDefaultPoolSize(envValue?: string | null): number {
|
|
148
|
+
const raw = (envValue ?? "").trim();
|
|
149
|
+
if (!raw) return 8;
|
|
150
|
+
const parsed = Number.parseInt(raw, 10);
|
|
151
|
+
if (!Number.isFinite(parsed) || parsed <= 0) return 8;
|
|
152
|
+
return Math.min(64, Math.max(1, parsed));
|
|
153
|
+
}
|
|
@@ -0,0 +1,23 @@
|
|
|
1
|
+
import type { JSONSchema, ResponseSkeleton } from "@elizaos/core";
|
|
2
|
+
export type StructuredOutputRepairStatus = "unchanged" | "repaired" | "ambiguous" | "invalid";
|
|
3
|
+
export interface StructuredOutputRepairOptions {
|
|
4
|
+
skeleton?: ResponseSkeleton;
|
|
5
|
+
jsonSchema?: JSONSchema;
|
|
6
|
+
jsonSchemasByKey?: Readonly<Record<string, JSONSchema | undefined>>;
|
|
7
|
+
}
|
|
8
|
+
export interface StructuredOutputRepairResult {
|
|
9
|
+
text: string;
|
|
10
|
+
status: StructuredOutputRepairStatus;
|
|
11
|
+
reason?: string;
|
|
12
|
+
}
|
|
13
|
+
export declare function repairStructuredOutput(text: string, options: StructuredOutputRepairOptions): StructuredOutputRepairResult;
|
|
14
|
+
export declare class StructuredOutputRepairStream {
|
|
15
|
+
private readonly options;
|
|
16
|
+
private text;
|
|
17
|
+
private syntheticTail;
|
|
18
|
+
constructor(options: StructuredOutputRepairOptions);
|
|
19
|
+
push(chunk: string): string;
|
|
20
|
+
flush(): string;
|
|
21
|
+
currentText(): string;
|
|
22
|
+
}
|
|
23
|
+
//# sourceMappingURL=deterministic-repair.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"deterministic-repair.d.ts","sourceRoot":"","sources":["deterministic-repair.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EACX,UAAU,EACV,gBAAgB,EAEhB,MAAM,eAAe,CAAC;AAEvB,MAAM,MAAM,4BAA4B,GACrC,WAAW,GACX,UAAU,GACV,WAAW,GACX,SAAS,CAAC;AAEb,MAAM,WAAW,6BAA6B;IAC7C,QAAQ,CAAC,EAAE,gBAAgB,CAAC;IAC5B,UAAU,CAAC,EAAE,UAAU,CAAC;IACxB,gBAAgB,CAAC,EAAE,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,UAAU,GAAG,SAAS,CAAC,CAAC,CAAC;CACpE;AAED,MAAM,WAAW,4BAA4B;IAC5C,IAAI,EAAE,MAAM,CAAC;IACb,MAAM,EAAE,4BAA4B,CAAC;IACrC,MAAM,CAAC,EAAE,MAAM,CAAC;CAChB;AAuWD,wBAAgB,sBAAsB,CACrC,IAAI,EAAE,MAAM,EACZ,OAAO,EAAE,6BAA6B,GACpC,4BAA4B,CAa9B;AASD,qBAAa,4BAA4B;IAI5B,OAAO,CAAC,QAAQ,CAAC,OAAO;IAHpC,OAAO,CAAC,IAAI,CAAM;IAClB,OAAO,CAAC,aAAa,CAAM;gBAEE,OAAO,EAAE,6BAA6B;IAEnE,IAAI,CAAC,KAAK,EAAE,MAAM,GAAG,MAAM;IAkB3B,KAAK,IAAI,MAAM;IASf,WAAW,IAAI,MAAM;CAGrB"}
|
|
@@ -0,0 +1,169 @@
|
|
|
1
|
+
import type { JSONSchema, ResponseSkeleton } from "@elizaos/core";
|
|
2
|
+
import { describe, expect, it } from "vitest";
|
|
3
|
+
import {
|
|
4
|
+
repairStructuredOutput,
|
|
5
|
+
StructuredOutputRepairStream,
|
|
6
|
+
} from "./deterministic-repair";
|
|
7
|
+
|
|
8
|
+
const actionSkeleton: ResponseSkeleton = {
|
|
9
|
+
spans: [
|
|
10
|
+
{ kind: "literal", value: '{"action":' },
|
|
11
|
+
{ kind: "enum", key: "action", enumValues: ["BLOCK", "BRIEF"] },
|
|
12
|
+
{ kind: "literal", value: ',"parameters":' },
|
|
13
|
+
{ kind: "free-json", key: "parameters" },
|
|
14
|
+
{ kind: "literal", value: ',"thought":' },
|
|
15
|
+
{ kind: "free-string", key: "thought" },
|
|
16
|
+
{ kind: "literal", value: "}" },
|
|
17
|
+
],
|
|
18
|
+
};
|
|
19
|
+
|
|
20
|
+
const responseHandlerSkeleton: ResponseSkeleton = {
|
|
21
|
+
spans: [
|
|
22
|
+
{ kind: "literal", value: '{"shouldRespond":"' },
|
|
23
|
+
{
|
|
24
|
+
kind: "enum",
|
|
25
|
+
key: "shouldRespond",
|
|
26
|
+
enumValues: ["RESPOND", "IGNORE", "STOP"],
|
|
27
|
+
},
|
|
28
|
+
{ kind: "literal", value: '","replyText":' },
|
|
29
|
+
{ kind: "free-string", key: "replyText" },
|
|
30
|
+
{ kind: "literal", value: "}" },
|
|
31
|
+
],
|
|
32
|
+
};
|
|
33
|
+
|
|
34
|
+
describe("deterministic structured-output repair", () => {
|
|
35
|
+
it("completes response-handler RE/IG enum prefixes and advances to replyText", () => {
|
|
36
|
+
expect(
|
|
37
|
+
repairStructuredOutput('{"shouldRespond":"RE', {
|
|
38
|
+
skeleton: responseHandlerSkeleton,
|
|
39
|
+
}),
|
|
40
|
+
).toEqual({
|
|
41
|
+
text: '{"shouldRespond":"RESPOND","replyText":',
|
|
42
|
+
status: "repaired",
|
|
43
|
+
reason: "free-span-incomplete",
|
|
44
|
+
});
|
|
45
|
+
|
|
46
|
+
expect(
|
|
47
|
+
repairStructuredOutput('{"shouldRespond":"IG', {
|
|
48
|
+
skeleton: responseHandlerSkeleton,
|
|
49
|
+
}),
|
|
50
|
+
).toEqual({
|
|
51
|
+
text: '{"shouldRespond":"IGNORE","replyText":',
|
|
52
|
+
status: "repaired",
|
|
53
|
+
reason: "free-span-incomplete",
|
|
54
|
+
});
|
|
55
|
+
});
|
|
56
|
+
|
|
57
|
+
it("streams deterministic response-handler enum completions without waiting for full enum tokens", () => {
|
|
58
|
+
const stream = new StructuredOutputRepairStream({
|
|
59
|
+
skeleton: responseHandlerSkeleton,
|
|
60
|
+
});
|
|
61
|
+
expect(stream.push('{"shouldRespond":"IG')).toBe(
|
|
62
|
+
'{"shouldRespond":"IGNORE","replyText":',
|
|
63
|
+
);
|
|
64
|
+
expect(stream.push('NORE","replyText":"')).toBe('"');
|
|
65
|
+
});
|
|
66
|
+
|
|
67
|
+
it("completes an unambiguous enum and advances to the next JSON field", () => {
|
|
68
|
+
expect(
|
|
69
|
+
repairStructuredOutput('{"action":"BLO', {
|
|
70
|
+
skeleton: actionSkeleton,
|
|
71
|
+
}),
|
|
72
|
+
).toEqual({
|
|
73
|
+
text: '{"action":"BLOCK","parameters":',
|
|
74
|
+
status: "repaired",
|
|
75
|
+
reason: "free-span-incomplete",
|
|
76
|
+
});
|
|
77
|
+
});
|
|
78
|
+
|
|
79
|
+
it("fails safely when an enum prefix is ambiguous", () => {
|
|
80
|
+
const repaired = repairStructuredOutput('{"action":"B', {
|
|
81
|
+
skeleton: actionSkeleton,
|
|
82
|
+
});
|
|
83
|
+
expect(repaired.text).toBe('{"action":"B');
|
|
84
|
+
expect(repaired.status).toBe("ambiguous");
|
|
85
|
+
});
|
|
86
|
+
|
|
87
|
+
it("fails safely when an enum value is hallucinated", () => {
|
|
88
|
+
const repaired = repairStructuredOutput('{"action":"DELETE"', {
|
|
89
|
+
skeleton: actionSkeleton,
|
|
90
|
+
});
|
|
91
|
+
expect(repaired.text).toBe('{"action":"DELETE"');
|
|
92
|
+
expect(repaired.status).toBe("invalid");
|
|
93
|
+
});
|
|
94
|
+
|
|
95
|
+
it("repairs a missing required-parameter comma and closes the object", () => {
|
|
96
|
+
const schema: JSONSchema = {
|
|
97
|
+
type: "object",
|
|
98
|
+
properties: {
|
|
99
|
+
url: { type: "string" },
|
|
100
|
+
selector: { type: "string" },
|
|
101
|
+
},
|
|
102
|
+
required: ["url", "selector"],
|
|
103
|
+
};
|
|
104
|
+
expect(
|
|
105
|
+
repairStructuredOutput('{"url":"https://example.test" "selector":"#go"', {
|
|
106
|
+
jsonSchema: schema,
|
|
107
|
+
}),
|
|
108
|
+
).toEqual({
|
|
109
|
+
text: '{"url":"https://example.test","selector":"#go"}',
|
|
110
|
+
status: "repaired",
|
|
111
|
+
});
|
|
112
|
+
});
|
|
113
|
+
|
|
114
|
+
it("repairs a semicolon separator when the following key is schema-known", () => {
|
|
115
|
+
const schema: JSONSchema = {
|
|
116
|
+
type: "object",
|
|
117
|
+
properties: {
|
|
118
|
+
url: { type: "string" },
|
|
119
|
+
selector: { type: "string" },
|
|
120
|
+
},
|
|
121
|
+
required: ["url", "selector"],
|
|
122
|
+
};
|
|
123
|
+
expect(
|
|
124
|
+
repairStructuredOutput('{"url":"https://example.test";"selector":"#go"', {
|
|
125
|
+
jsonSchema: schema,
|
|
126
|
+
}).text,
|
|
127
|
+
).toBe('{"url":"https://example.test","selector":"#go"}');
|
|
128
|
+
});
|
|
129
|
+
|
|
130
|
+
it("fills a required deterministic enum parameter", () => {
|
|
131
|
+
const schema: JSONSchema = {
|
|
132
|
+
type: "object",
|
|
133
|
+
properties: {
|
|
134
|
+
op: { type: "string", enum: ["open"] },
|
|
135
|
+
},
|
|
136
|
+
required: ["op"],
|
|
137
|
+
};
|
|
138
|
+
expect(repairStructuredOutput("{", { jsonSchema: schema })).toEqual({
|
|
139
|
+
text: '{"op":"open"}',
|
|
140
|
+
status: "repaired",
|
|
141
|
+
});
|
|
142
|
+
});
|
|
143
|
+
|
|
144
|
+
it("does not invent an ambiguous required parameter value", () => {
|
|
145
|
+
const schema: JSONSchema = {
|
|
146
|
+
type: "object",
|
|
147
|
+
properties: {
|
|
148
|
+
url: { type: "string" },
|
|
149
|
+
},
|
|
150
|
+
required: ["url"],
|
|
151
|
+
};
|
|
152
|
+
expect(repairStructuredOutput("{", { jsonSchema: schema })).toEqual({
|
|
153
|
+
text: "{",
|
|
154
|
+
status: "unchanged",
|
|
155
|
+
reason: "missing-required-value-ambiguous",
|
|
156
|
+
});
|
|
157
|
+
});
|
|
158
|
+
|
|
159
|
+
it("deduplicates deterministic scaffold if the stream later emits it", () => {
|
|
160
|
+
const stream = new StructuredOutputRepairStream({
|
|
161
|
+
skeleton: actionSkeleton,
|
|
162
|
+
});
|
|
163
|
+
expect(stream.push('{"action":"BLO')).toBe(
|
|
164
|
+
'{"action":"BLOCK","parameters":',
|
|
165
|
+
);
|
|
166
|
+
expect(stream.push('CK","parameters":{}')).toBe('{},"thought":');
|
|
167
|
+
expect(stream.flush()).toBe("");
|
|
168
|
+
});
|
|
169
|
+
});
|