@snoai/memory 1.2.13 → 1.5.0

package/dist/{rem-trigger-B1P5_RqV.js → rem-trigger-DL0MlLeo.js}

@@ -1,14 +1,14 @@
-import { $ as FIXED_IDX_MEMORY_ACTIVE_TASK_CURRENT_REVISION, A as modelCallDestination, At as FIXED_MEMORY_PROFILE_RECOVERY_ENTRIES, B as FIXED_EXTERNAL_VALUE_49, C as getSnoStationMemStateDir, Cn as CHUNK_OVERLAP_TOKENS, Cr as CachedEmbeddingProvider, Ct as FIXED_MEMORY_CHUNKS_MEM_FACET_IDX, D as settingsToPluginConfig, Dn as DEFAULT_IMPORTANCE, Dr as LocalEmbedProvider, Dt as FIXED_MEMORY_MEMORIES_FACT_ID_INSERT_GUARD, En as DEFAULT_BM25_WEIGHT, Et as FIXED_MEMORY_MEMORIES, Fn as DEFAULT_VECTOR_WEIGHT, G as FIXED_EXTERNAL_VALUE_55, Gn as IMPORTANCE_WEIGHT_BASE, H as FIXED_EXTERNAL_VALUE_52, Hn as HEAD_EXTRACT_MIN_CONTENT_TOKENS, In as EMBEDDER_DIM, Ir as REM_SIDECAR_TOKEN_HEADER, It as FIXED_MEMORY_SNO_EXTRACT_CHAT, J as FIXED_EXTERNAL_VALUE_58, Jn as LIGHTWEIGHT_COSINE_WEIGHT, K as FIXED_EXTERNAL_VALUE_56, Kt as FIXED_PROTOCOL_VALUE_69, Ln as EMBEDDER_MODEL_DEFAULT, Mn as DEFAULT_RERANK_MODEL, Mt as FIXED_MEMORY_PROVIDER_PROJECT_AGENTS, Nn as DEFAULT_RERANK_TIMEOUT_MS, Nt as FIXED_MEMORY_PROVIDER_PROJECT_MAPPINGS, O as MODEL_CALLS, On as DEFAULT_RECALL_LIFECYCLE, Or as ensureModelDownloaded, Ot as FIXED_MEMORY_MEMORIES_FACT_ID_UPDATE_GUARD, P as observeAgentId, Q as FIXED_IDX_MEMORY, Rn as EMBEDDER_PROVIDER_DEFAULT, Rt as FIXED_MEMORY_TASK_LIFECYCLE_COMMANDS, Sn as CHUNK_MIN_TOKENS, St as FIXED_MEMORY_CHUNKS_MEMORY_ID_IDX, T as readSettings, Tn as CHUNK_TOKENIZER_MODE, Tr as LOCAL_EMBEDDING_MODEL, Tt as FIXED_MEMORY_EXTRACTION_TIMESTAMPS, U as FIXED_EXTERNAL_VALUE_53, Un as HEAD_EXTRACT_OVERLAP_DROP_RATIO, V as FIXED_EXTERNAL_VALUE_51, W as FIXED_EXTERNAL_VALUE_54, Wn as HEAD_EXTRACT_TOKEN_BUDGET, X as FIXED_EXTERNAL_VALUE_8, Xn as LIGHTWEIGHT_RERANK_PENALTY, Y as FIXED_EXTERNAL_VALUE_62, Yn as LIGHTWEIGHT_FUSION_WEIGHT, Z as FIXED_EXTERNAL_VALUE_9, _t as FIXED_MEMORY_CHUNKS, a as REM_MODEL_OUTPUT_TOKEN_CAP, an as PERSISTED_PROVIDER_SYSTEM, ar as MMR_LAMBDA, at as FIXED_IDX_MEMORY_MEMORIES_LANE_PROJECT, bn as CHUNKING_VERSION$1, br as experimentCandidatePoolOverride, bt as FIXED_MEMORY_CHUNKS_AU, cn as SUPPORTED_LOCALES, ct as FIXED_IDX_MEMORY_MEMORIES_PROJECT_FACT_KEY_ACTIVE, d as getRemDiscoveryPath, dn as ADJACENCY_BONUS_PER_PAIR, dt as FIXED_MEMORY_ACTIVE_TASK_EVIDENCE, en as FIXED_VEC_MEMORY, er as MAX_CHUNK_FETCH_LIMIT, et as FIXED_IDX_MEMORY_ACTIVE_TASK_EVIDENCE_INSTANCE, fr as RERANK_BLEND_CROSS, ft as FIXED_MEMORY_ACTIVE_TASK_INSTANCES, g as readRemOperationalConfig, gr as TIME_DECAY_FLOOR, gt as FIXED_MEMORY_ACTIVE_TASK_TRANSITIONS, ht as FIXED_MEMORY_ACTIVE_TASK_REVISIONS, in as PERSISTED_CONTENT_HASH_V3, it as FIXED_IDX_MEMORY_MEMORIES_FACT_ID, jt as FIXED_MEMORY_PROVIDER_AGENT_MAPPINGS, k as MODEL_CALL_WORKLOAD, kt as FIXED_MEMORY_MIGRATION_MARKERS, lr as RECENCY_WEIGHT, lt as FIXED_IDX_MEMORY_MEMORIES_PROJECT_TIMESTAMP, mr as RETRIEVAL_STORAGE_CHUNK_PROFILE, mt as FIXED_MEMORY_ACTIVE_TASK_MIGRATION_MANIFESTS, nn as PERSISTED_CANDIDATE_HASH_V1, nt as FIXED_IDX_MEMORY_MEMORIES_CATEGORY, o as REM_RUN_PATH, or as MULTI_HIT_BONUS_CAP, ot as FIXED_IDX_MEMORY_MEMORIES_PROJECT, pr as RERANK_BLEND_VECTOR, pt as FIXED_MEMORY_ACTIVE_TASK_MIGRATION_EVIDENCE, q as FIXED_EXTERNAL_VALUE_57, qn as INTRA_MEMORY_COSINE_THRESHOLD, qt as FIXED_PROTOCOL_VALUE_73, r as REM_CORRELATION_ID_HEADER, rn as PERSISTED_CONTENT_HASH_V1, rt as FIXED_IDX_MEMORY_MEMORIES_CONTENT_HASH, s as REM_SIDECAR_HOST, sn as llmRoutingConfigSchema, sr as MULTI_HIT_BONUS_PER_HIT, st as FIXED_IDX_MEMORY_MEMORIES_PROJECT_CONTENT_HASH, tn as FIXED_VEC_MEMORY_CHUNKS, tr as MAX_LIST_LIMIT, tt as FIXED_IDX_MEMORY_ACTIVE_TASK_INSTANCES_PROJECTION, un as ADJACENCY_BONUS_CAP, ut as FIXED_MEMORY, vn as BONUS_TOTAL_INVARIANT, vt as FIXED_MEMORY_CHUNKS_AD, wn as CHUNK_TARGET_TOKENS, wr as LOCAL_EMBEDDING_CACHE_DIR_DEFAULT, wt as FIXED_MEMORY_CHUNKS_MEM_IDX, xn as CHUNK_MAX_TOKENS, xr as experimentMmrDisabled, xt as FIXED_MEMORY_CHUNKS_FTS, yn as CACHE_TTL_MS, yt as FIXED_MEMORY_CHUNKS_AI, zn as EVAL_QA_TRACES_FILENAME, zt as FIXED_MEMORY_UNPLACED_CANDIDATES } from "./config-D_KOwEyg.js";
+import { $ as FIXED_IDX_MEMORY_ACTIVE_TASK_CURRENT_REVISION, $n as MAX_CHUNK_FETCH_LIMIT, A as modelCallDestination, At as FIXED_MEMORY_PROFILE_RECOVERY_ENTRIES, B as FIXED_EXTERNAL_VALUE_49, C as getSnoStationMemStateDir, Cn as CHUNK_OVERLAP_TOKENS, Cr as LOCAL_EMBEDDING_CACHE_DIR_DEFAULT, Ct as FIXED_MEMORY_CHUNKS_MEM_FACET_IDX, D as settingsToPluginConfig, Dn as DEFAULT_IMPORTANCE, Dr as LocalReranker, Dt as FIXED_MEMORY_MEMORIES_FACT_ID_INSERT_GUARD, En as DEFAULT_BM25_WEIGHT, Er as LocalEmbedProvider, Et as FIXED_MEMORY_MEMORIES, Fn as EMBEDDER_DIM, G as FIXED_EXTERNAL_VALUE_55, H as FIXED_EXTERNAL_VALUE_52, Hn as HEAD_EXTRACT_OVERLAP_DROP_RATIO, In as EMBEDDER_MODEL_DEFAULT, Ir as REM_SIDECAR_TOKEN_HEADER, It as FIXED_MEMORY_SNO_EXTRACT_CHAT, J as FIXED_EXTERNAL_VALUE_58, Jn as LIGHTWEIGHT_FUSION_WEIGHT, K as FIXED_EXTERNAL_VALUE_56, Kn as INTRA_MEMORY_COSINE_THRESHOLD, Kt as FIXED_PROTOCOL_VALUE_69, Ln as EMBEDDER_PROVIDER_DEFAULT, Mn as DEFAULT_RERANK_TIMEOUT_MS, Mt as FIXED_MEMORY_PROVIDER_PROJECT_AGENTS, Nt as FIXED_MEMORY_PROVIDER_PROJECT_MAPPINGS, O as MODEL_CALLS, On as DEFAULT_RECALL_LIFECYCLE, Or as ensureModelDownloaded, Ot as FIXED_MEMORY_MEMORIES_FACT_ID_UPDATE_GUARD, P as observeAgentId, Pn as DEFAULT_VECTOR_WEIGHT, Q as FIXED_IDX_MEMORY, Rn as EVAL_QA_TRACES_FILENAME, Rt as FIXED_MEMORY_TASK_LIFECYCLE_COMMANDS, Sn as CHUNK_MIN_TOKENS, Sr as CachedEmbeddingProvider, St as FIXED_MEMORY_CHUNKS_MEMORY_ID_IDX, T as readSettings, Tn as CHUNK_TOKENIZER_MODE, Tt as FIXED_MEMORY_EXTRACTION_TIMESTAMPS, U as FIXED_EXTERNAL_VALUE_53, Un as HEAD_EXTRACT_TOKEN_BUDGET, V as FIXED_EXTERNAL_VALUE_51, Vn as HEAD_EXTRACT_MIN_CONTENT_TOKENS, W as FIXED_EXTERNAL_VALUE_54, Wn as IMPORTANCE_WEIGHT_BASE, X as FIXED_EXTERNAL_VALUE_8, Y as FIXED_EXTERNAL_VALUE_62, Yn as LIGHTWEIGHT_RERANK_PENALTY, Z as FIXED_EXTERNAL_VALUE_9, _t as FIXED_MEMORY_CHUNKS, a as REM_MODEL_OUTPUT_TOKEN_CAP, an as PERSISTED_PROVIDER_SYSTEM, ar as MULTI_HIT_BONUS_CAP, at as FIXED_IDX_MEMORY_MEMORIES_LANE_PROJECT, bn as CHUNKING_VERSION$1, br as experimentMmrDisabled, bt as FIXED_MEMORY_CHUNKS_AU, cn as SUPPORTED_LOCALES, cr as RECENCY_WEIGHT, ct as FIXED_IDX_MEMORY_MEMORIES_PROJECT_FACT_KEY_ACTIVE, d as getRemDiscoveryPath, dn as ADJACENCY_BONUS_PER_PAIR, dr as RERANK_BLEND_CROSS, dt as FIXED_MEMORY_ACTIVE_TASK_EVIDENCE, en as FIXED_VEC_MEMORY, er as MAX_LIST_LIMIT, et as FIXED_IDX_MEMORY_ACTIVE_TASK_EVIDENCE_INSTANCE, fr as RERANK_BLEND_VECTOR, ft as FIXED_MEMORY_ACTIVE_TASK_INSTANCES, g as readRemOperationalConfig, gt as FIXED_MEMORY_ACTIVE_TASK_TRANSITIONS, hr as TIME_DECAY_FLOOR, ht as FIXED_MEMORY_ACTIVE_TASK_REVISIONS, in as PERSISTED_CONTENT_HASH_V3, ir as MMR_LAMBDA, it as FIXED_IDX_MEMORY_MEMORIES_FACT_ID, jt as FIXED_MEMORY_PROVIDER_AGENT_MAPPINGS, k as MODEL_CALL_WORKLOAD, kt as FIXED_MEMORY_MIGRATION_MARKERS, lt as FIXED_IDX_MEMORY_MEMORIES_PROJECT_TIMESTAMP, mt as FIXED_MEMORY_ACTIVE_TASK_MIGRATION_MANIFESTS, nn as PERSISTED_CANDIDATE_HASH_V1, nt as FIXED_IDX_MEMORY_MEMORIES_CATEGORY, o as REM_RUN_PATH, or as MULTI_HIT_BONUS_PER_HIT, ot as FIXED_IDX_MEMORY_MEMORIES_PROJECT, pr as RETRIEVAL_STORAGE_CHUNK_PROFILE, pt as FIXED_MEMORY_ACTIVE_TASK_MIGRATION_EVIDENCE, q as FIXED_EXTERNAL_VALUE_57, qn as LIGHTWEIGHT_COSINE_WEIGHT, qt as FIXED_PROTOCOL_VALUE_73, r as REM_CORRELATION_ID_HEADER, rn as PERSISTED_CONTENT_HASH_V1, rt as FIXED_IDX_MEMORY_MEMORIES_CONTENT_HASH, s as REM_SIDECAR_HOST, sn as llmRoutingConfigSchema, st as FIXED_IDX_MEMORY_MEMORIES_PROJECT_CONTENT_HASH, tn as FIXED_VEC_MEMORY_CHUNKS, tt as FIXED_IDX_MEMORY_ACTIVE_TASK_INSTANCES_PROJECTION, un as ADJACENCY_BONUS_CAP, ut as FIXED_MEMORY, vn as BONUS_TOTAL_INVARIANT, vt as FIXED_MEMORY_CHUNKS_AD, wn as CHUNK_TARGET_TOKENS, wr as LOCAL_EMBEDDING_MODEL, wt as FIXED_MEMORY_CHUNKS_MEM_IDX, xn as CHUNK_MAX_TOKENS, xt as FIXED_MEMORY_CHUNKS_FTS, yn as CACHE_TTL_MS, yr as experimentCandidatePoolOverride, yt as FIXED_MEMORY_CHUNKS_AI, zt as FIXED_MEMORY_UNPLACED_CANDIDATES } from "./config-DWeuetYo.js";
 import { createLogger, createLogger as createLogger$1, currentLogContext, privateLogReference, withLogContext } from "@snoai/utils/logger";
 import path, { dirname, join, resolve } from "node:path";
 import { appendFileSync, closeSync, createReadStream, existsSync, fstatSync, fsyncSync, ftruncateSync, mkdirSync, openSync, readFileSync, readSync, realpathSync, renameSync, rmSync, statSync, writeFileSync, writeSync } from "node:fs";
 import { arch, availableParallelism, platform } from "node:os";
 import { fileURLToPath } from "node:url";
+import { createHash, createHmac, createPublicKey, randomUUID, randomUUID as randomUUID$1, timingSafeEqual, verify } from "node:crypto";
 import { LRUCache } from "lru-cache";
 import { z } from "zod";
 import { CHUNKING_VERSION, aggregateChunksToMemories, buildChunkId, buildDensePayload, chunk, chunk as chunk$1, countTokens, expandSnippetWindow, extractMetadataHeader, headExtract, shouldDropSummary } from "@snoai/chunking";
 import { AsyncLocalStorage } from "node:async_hooks";
-import { createHash, createHmac, createPublicKey, randomUUID, randomUUID as randomUUID$1, timingSafeEqual, verify } from "node:crypto";
 import { getDek, openEncryptedDb, openEncryptedDbReadonly } from "@snoai/sqlite-crypto";
 import { appendFile, mkdir, open, readFile, rename, rm } from "node:fs/promises";
 import { Temporal } from "@js-temporal/polyfill";
@@ -7413,6 +7413,7 @@ function createLlmClient(config) {
 			const hostInFlight = ++hostRequestsInFlight;
 			const result = await withLogContext({ attempt_id: randomUUID() }, async () => {
 				let outcome = "failed";
+				let hostReport = {};
 				try {
 					const completion = await agentPort.complete({
 						system: systemContent,
@@ -7423,6 +7424,7 @@ function createLlmClient(config) {
 						...request.signal ? { signal: request.signal } : {}
 					});
 					outcome = completion.kind;
+					if (completion.kind === "ok") hostReport = completion;
 					return completion;
 				} finally {
 					hostRequestsInFlight -= 1;
@@ -7434,8 +7436,13 @@ function createLlmClient(config) {
 						duration_ms: performance.now() - hostStarted,
 						in_flight_at_dispatch: hostInFlight,
 						requested_model: "unavailable",
-						returned_model: "unavailable",
-						usage: { source: "unavailable" },
+						returned_model: hostReport.model ?? "unavailable",
+						usage: hostReport.usage ? {
+							source: "provider_reported",
+							inputTokens: hostReport.usage.inputTokens,
+							outputTokens: hostReport.usage.outputTokens,
+							totalTokens: hostReport.usage.inputTokens + hostReport.usage.outputTokens
+						} : { source: "unavailable" },
 						prompt_template_hash: request.promptTemplateHash ?? "unavailable",
 						extraction_skill_hash: request.extractionSkillHash ?? "unavailable",
 						prompt_template_unavailable_reason: request.promptTemplateHash ? void 0 : "caller_does_not_provide_template_identity",
@@ -7454,7 +7461,7 @@ function createLlmClient(config) {
 						dispatch_to_headers_ms: "unavailable",
 						headers_to_body_ms: "unavailable",
 						first_token_ms: "unavailable",
-						provider_unavailable_reason: "host_port_does_not_report_provider_fields"
+						...hostReport.usage || hostReport.model ? {} : { provider_unavailable_reason: "host_port_does_not_report_provider_fields" }
 					}, {
 						event_name: "llm.request.completed",
 						file: "packages/memory/src/model/llm-client.ts",
@@ -26393,6 +26400,7 @@ var Embedder = class {
 	model;
 	warmupPromise = null;
 	modelDownload;
+	reranker;
 	/**
 	* Initializes embedding generation collaborators while keeping runtime work in explicit
 	* methods.
@@ -26410,6 +26418,12 @@ var Embedder = class {
 		this.dimensions = inner.dimension;
 		this.providerKind = providerKind;
 		this.model = config.model && config.model.trim().length > 0 ? config.model : LOCAL_EMBEDDING_MODEL;
+		this.reranker = new LocalReranker({
+			cacheDir: resolveEmbeddingCacheDir(config),
+			offline: config.offline === true,
+			...config.mirror ? { mirror: config.mirror } : {},
+			intraOpNumThreads: config.sessionOptions?.intraOpNumThreads ?? Math.max(1, Math.min(4, Math.floor(availableParallelism() / 2)))
+		});
 		this.provider = new CachedEmbeddingProvider(inner, {
 			maxSize: 256,
 			ttlMs: CACHE_TTL_MS
@@ -26579,11 +26593,19 @@ var Embedder = class {
 		});
 		return this.warmupPromise;
 	}
+	/**
+	* Local relevance scores (raw logits, higher is more relevant) for each text against the query, or undefined while
+	* the local reranker is not loaded; the first call starts loading it in the background.
+	*/
+	scoreRelevance(query, texts, signal) {
+		return this.reranker.score(query, texts, signal);
+	}
 	/** Implements dispose as the local embedding generation operation. */
 	async dispose() {
 		const warmup = this.warmupPromise;
 		if (warmup) await warmup.catch(() => void 0);
 		await this.provider.dispose();
+		await this.reranker.dispose();
 	}
 };
 /** Creates the embedder facade around the local ONNX provider. */
@@ -27194,9 +27216,7 @@ const DEFAULT_RETRIEVAL_CONFIG = {
 	recencyWeight: RECENCY_WEIGHT,
 	temporalExpiry: false,
 	temporalDecay: true,
-	rerankModel: DEFAULT_RERANK_MODEL,
 	rerankTimeoutMs: DEFAULT_RERANK_TIMEOUT_MS,
-	rerankProvider: "voyage",
 	lengthNormAnchor: 0,
 	hardMinScore: 0,
 	timeDecayHalfLifeDays: 60,
@@ -27501,11 +27521,10 @@ Object.assign(MemoryRetriever.prototype, {
 		this.throwIfAborted(context);
 		const rerankStageMetadata = rerankOutcome.fallback ? {
 			...rerankOutcome.stats,
-			rerankFallbackReason: rerankOutcome.fallback.reason,
-			rerankFallbackProvider: rerankOutcome.fallback.provider
+			rerankFallbackReason: rerankOutcome.fallback.reason
 		} : rerankOutcome.stats;
 		trace?.endStage(fused.map((r) => r.entry.id), fused.map((r) => r.score), rerankStageMetadata);
-		fused = this.applyScoringPipeline(fused, trace, context.limit);
+		fused = this.applyScoringPipeline(fused, trace, context.limit, rerankOutcome.modelOrder);
 		trace?.startStage("min_score_filter", fused.map((result) => result.entry.id));
 		fused = fused.filter((result) => result.score >= (context.minScore ?? this.config.minScore));
 		trace?.endStage(fused.map((result) => result.entry.id), fused.map((result) => result.score));
@@ -27574,100 +27593,31 @@ Object.assign(MemoryRetriever.prototype, {
 });
 //#endregion
 //#region src/engine/retrieval/retrieval-rerank-provider.ts
-/** @file retrieval-rerank-provider.ts
-* @purpose Adapts external rerank provider request and response formats.
-* @boundary HTTP payload shaping and response parsing only.
-*/
-const RERANK_DEFAULT_ENDPOINTS = {
-	voyage: "https://api.voyageai.com/v1/rerank",
-	jina: "https://api.jina.ai/v1/rerank",
-	pinecone: "https://api.pinecone.io/rerank",
-	dashscope: "https://dashscope.aliyuncs.com/api/v1/services/rerank"
-};
-function buildRerankHttpError(status) {
-	if (classifyLlmFailure({ status }).category === "auth") return new RetrievalError(`Rerank API failed with status ${status}`);
-}
-function buildRerankRequest(provider, apiKey, model, query, candidates, topN) {
-	switch (provider) {
-		case "tei":
-		case "sno-gpu": return {
-			headers: {
-				"Content-Type": "application/json",
-				Authorization: `Bearer ${apiKey}`,
-				...provider === "sno-gpu" ? {
-					"X-Sno-Source": "sno-station",
-					"X-Sno-Workload": "station-rerank"
-				} : {}
-			},
-			body: {
-				query,
-				texts: candidates,
-				raw_scores: true
-			}
-		};
-		case "dashscope": return {
-			headers: {
-				"Content-Type": "application/json",
-				Authorization: `Bearer ${apiKey}`
-			},
-			body: {
-				model,
-				input: {
-					query,
-					documents: candidates
-				}
-			}
-		};
-		case "pinecone": return {
-			headers: {
-				"Content-Type": "application/json",
-				"Api-Key": apiKey,
-				"X-Pinecone-API-Version": "2024-10"
-			},
-			body: {
-				model,
-				query,
-				documents: candidates.map((text) => ({ text })),
-				top_n: topN,
-				rank_fields: ["text"]
-			}
-		};
-		case "voyage": return {
-			headers: {
-				"Content-Type": "application/json",
-				Authorization: `Bearer ${apiKey}`
-			},
-			body: {
-				model,
-				query,
-				documents: candidates,
-				top_k: topN
-			}
-		};
-		default: return {
-			headers: {
-				"Content-Type": "application/json",
-				Authorization: `Bearer ${apiKey}`
-			},
-			body: {
-				model,
-				query,
-				documents: candidates,
-				top_n: topN
-			}
-		};
-	}
+function buildRerankRequest(apiKey, query, candidates) {
+	return {
+		headers: {
+			"Content-Type": "application/json",
+			Authorization: `Bearer ${apiKey}`,
+			"X-Sno-Source": "sno-station",
+			"X-Sno-Workload": "station-rerank"
+		},
+		body: {
+			query,
+			texts: candidates,
+			raw_scores: true
+		}
+	};
 }
-function parseRerankResponse(provider, data) {
-	/** Parses items into the normalized shape used by precision recall retrieval ranking. */
-	const parseItems = (items, scoreKeys) => {
+/** Sno's reranker answers a list of `{index, score}`, bare or under `results` / `data`. */
+function parseRerankResponse(data) {
+	const parseItems = (items) => {
 		if (!Array.isArray(items)) return void 0;
 		const parsed = [];
 		for (const raw of items) {
 			const index = typeof raw?.index === "number" ? raw.index : Number(raw?.index);
 			if (!Number.isFinite(index)) continue;
 			let score;
-			for (const key of scoreKeys) {
+			for (const key of ["score", "relevance_score"]) {
 				const value = raw?.[key];
 				const n = typeof value === "number" ? value : Number(value);
 				if (Number.isFinite(n)) {
@@ -27684,18 +27634,7 @@ function parseRerankResponse(provider, data) {
 		return parsed.length > 0 ? parsed : void 0;
 	};
 	const objectData = data && typeof data === "object" && !Array.isArray(data) ? data : void 0;
-	switch (provider) {
-		case "tei":
-		case "sno-gpu": return parseItems(data, ["score", "relevance_score"]) ?? parseItems(objectData?.results, ["score", "relevance_score"]) ?? parseItems(objectData?.data, ["score", "relevance_score"]);
-		case "dashscope": {
-			const output = objectData?.output;
-			if (output) return parseItems(output.results, ["relevance_score", "score"]);
-			return parseItems(objectData?.results, ["relevance_score", "score"]);
-		}
-		case "pinecone": return parseItems(objectData?.data, ["score", "relevance_score"]) ?? parseItems(objectData?.results, ["score", "relevance_score"]);
-		case "voyage": return parseItems(objectData?.data, ["relevance_score", "score"]) ?? parseItems(objectData?.results, ["relevance_score", "score"]);
-		default: return parseItems(objectData?.results, ["relevance_score", "score"]) ?? parseItems(objectData?.data, ["relevance_score", "score"]);
-	}
+	return parseItems(data) ?? parseItems(objectData?.results) ?? parseItems(objectData?.data);
 }
 //#endregion
 //#region src/engine/retrieval/retriever-rerank.ts
@@ -27763,11 +27702,9 @@ async function fetchRerankWithRetry(send, stop) {
 	for (let attempt = 1; attempt <= RERANK_TRANSIENT_ATTEMPTS; attempt += 1) {
 		let retryDelayMs = RERANK_TRANSIENT_BACKOFF_MS * attempt;
 		try {
-			const response = await send();
-			if (response.ok) return {
-				response,
-				data: await response.json()
-			};
+			const reply = await send();
+			const { response } = reply;
+			if (response.ok) return reply;
 			if (!RERANK_TRANSIENT_STATUSES.has(response.status) || attempt === RERANK_TRANSIENT_ATTEMPTS) return { response };
 			if (response.status === 429) {
 				const retryAfter = response.headers.get("retry-after");
@@ -27799,13 +27736,12 @@ Object.assign(MemoryRetriever.prototype, {
 	async rerank(query, candidates, queryVector, options) {
 		if (candidates.length === 0) return { candidates };
 		if (this.config.rerank === "none") return { candidates };
-		const provider = this.config.rerankProvider ?? "none";
-		const unranked = () => provider === "sno-gpu" ? this.rerankLightweight(candidates, queryVector) : candidates;
-		if (this.config.rerank === "cross-encoder" && !this.config.rerankApiKey) throw new RetrievalError(`rerank "cross-encoder" requires retrieval.rerankApiKey (provider: ${provider}); set the key, or set retrieval.rerank to "lightweight" to choose the local ranker`);
-		if (this.config.rerank === "lightweight" || !this.config.rerankApiKey) return { candidates: this.rerankLightweight(candidates, queryVector) };
-		const rerankEndpoint = this.config.rerankEndpoint ?? RERANK_DEFAULT_ENDPOINTS[provider];
+		const unranked = () => this.rerankLightweight(candidates, queryVector);
+		if (this.config.rerank === "cross-encoder" && !this.config.rerankApiKey) throw new RetrievalError("rerank \"cross-encoder\" requires retrieval.rerankApiKey; set the key, or set retrieval.rerank to \"lightweight\" to choose the local ranker");
+		if (this.config.rerank === "lightweight" || !this.config.rerankApiKey) return await this.rerankLocalModel(query, this.rerankLightweight(candidates, queryVector), options?.deadlineAt, options?.signal);
+		const rerankEndpoint = this.config.rerankEndpoint;
 		if (!rerankEndpoint) {
-			log$4.warn("rerank skipped: no endpoint for provider", { provider }, {
+			log$4.warn("rerank skipped: no reranker address", {}, {
 				event_name: "memory.retriever_rerank.diagnostic",
 				file: "packages/memory/src/engine/retrieval/retriever-rerank.ts",
 				function: "rerank",
@@ -27813,15 +27749,12 @@ Object.assign(MemoryRetriever.prototype, {
 			});
 			return {
 				candidates: unranked(),
-				fallback: {
-					reason: "no_endpoint",
-					provider
-				}
+				fallback: { reason: "no_endpoint" }
 			};
 		}
 		const rerankTimeoutMs = this.config.rerankTimeoutMs ?? 15e3;
-		const stepEnds = [options?.signal, provider === "sno-gpu" ? AbortSignal.timeout(Math.max(0, Math.floor(options?.deadlineAt === void 0 ? rerankTimeoutMs : options.deadlineAt - Date.now()))) : void 0].filter((each) => each !== void 0);
-		const stepStop = stepEnds.length > 0 ? AbortSignal.any(stepEnds) : void 0;
+		const stepBudget = AbortSignal.timeout(Math.max(0, Math.floor(options?.deadlineAt === void 0 ? rerankTimeoutMs : options.deadlineAt - Date.now())));
+		const stepStop = options?.signal ? AbortSignal.any([options.signal, stepBudget]) : stepBudget;
 		const rerankApiKey = this.config.rerankApiKey;
 		const maxCandidates = this.config.rerankMaxCandidates;
 		const overCap = maxCandidates !== void 0 && candidates.length > maxCandidates;
@@ -27843,10 +27776,7 @@ Object.assign(MemoryRetriever.prototype, {
 			});
 			return {
 				candidates: unranked(),
-				fallback: {
-					reason: "query_over_window",
-					provider
-				}
+				fallback: { reason: "query_over_window" }
 			};
 		}
 		let truncatedCandidates = 0;
@@ -27869,20 +27799,54 @@ Object.assign(MemoryRetriever.prototype, {
 			const runBatch = async (start) => {
 				let batchTexts = rerankTexts.slice(start, start + batchSize);
 				for (let pass = 0;; pass += 1) {
-					const { headers, body } = buildRerankRequest(provider, rerankApiKey, this.config.rerankModel ?? "rerank-2", query, batchTexts, batchTexts.length);
-					const { response, data } = await fetchRerankWithRetry(() => fetch(rerankEndpoint, {
-						method: "POST",
-						headers: provider === "sno-gpu" ? {
-							...headers,
-							"X-Request-ID": randomUUID()
-						} : headers,
-						body: JSON.stringify(body),
-						signal: stepStop ? AbortSignal.any([AbortSignal.timeout(rerankTimeoutMs), stepStop]) : AbortSignal.timeout(rerankTimeoutMs)
-					}), stepStop);
+					const { headers, body } = buildRerankRequest(rerankApiKey, query, batchTexts);
+					const sendAttempt = async () => {
+						const clientRequestId = randomUUID();
+						const attemptStarted = performance.now();
+						let outcome = "failed";
+						let statusCode;
+						try {
+							const attempt = await fetch(rerankEndpoint, {
+								method: "POST",
+								headers: {
+									...headers,
+									"X-Request-ID": clientRequestId
+								},
+								body: JSON.stringify(body),
+								signal: AbortSignal.any([AbortSignal.timeout(rerankTimeoutMs), stepStop])
+							});
+							statusCode = attempt.status;
+							if (!attempt.ok) {
+								outcome = "http_error";
+								return { response: attempt };
+							}
+							const data = await attempt.json();
+							outcome = "success";
+							return {
+								response: attempt,
+								data
+							};
+						} catch (error) {
+							outcome = isRerankTimeout(error) ? "timeout" : "failed";
+							throw error;
+						} finally {
+							log$4.info("Rerank request completed", {
+								outcome,
+								status_code: statusCode ?? "unavailable",
+								duration_ms: performance.now() - attemptStarted,
+								texts_sent: batchTexts.length,
+								client_request_id: clientRequestId
+							}, {
+								event_name: "rerank.request.completed",
+								file: "packages/memory/src/engine/retrieval/retriever-rerank.ts",
+								function: "runBatch",
+								site_id: "rerank.request.completed"
+							});
+						}
+					};
+					const { response, data } = await fetchRerankWithRetry(sendAttempt, stepStop);
 					if (!response.ok) {
 						const retryAfter = response.headers.get("retry-after");
-						const fatalError = buildRerankHttpError(response.status);
-						if (fatalError && provider !== "sno-gpu") throw fatalError;
 						const errorBody = await response.text().catch(() => "<unreadable body>");
 						const refusal = pass === 0 ? parseOverWindowRefusal(errorBody) : void 0;
 						if (refusal) {
@@ -27924,10 +27888,9 @@ Object.assign(MemoryRetriever.prototype, {
 						});
 						return "http_error";
 					}
-					const parsed = parseRerankResponse(provider, data);
+					const parsed = parseRerankResponse(data);
 					if (!parsed) {
 						log$4.warn("rerank API returned unparseable response", {
-							provider,
 							endpoint_hash: createHash("sha256").update(rerankEndpoint).digest("hex"),
 							batchStart: start
 						}, {
@@ -27959,10 +27922,7 @@ Object.assign(MemoryRetriever.prototype, {
 				for (const outcome of settled) if (outcome.status === "rejected") throw outcome.reason;
 				for (const outcome of settled) if (outcome.status === "fulfilled" && typeof outcome.value === "string") return {
 					candidates: unranked(),
-					fallback: {
-						reason: outcome.value,
-						provider
-					}
+					fallback: { reason: outcome.value }
 				};
 				for (const outcome of settled) if (outcome.status === "fulfilled" && typeof outcome.value !== "string") items.push(...outcome.value);
 			}
@@ -27977,9 +27937,7 @@ Object.assign(MemoryRetriever.prototype, {
 				const candidate = toRerank[item.index];
 				if (!candidate) continue;
 				const sourceScore = this.getRerankSourceScore(candidate);
-				let crossScore = item.score;
-				if (provider === "tei" || provider === "sno-gpu") crossScore = 1 / (1 + Math.exp(-item.score));
-				const blendedScore = clamp01(crossScore * blendCross + sourceScore * blendVector, 0);
+				const blendedScore = clamp01(1 / (1 + Math.exp(-item.score)) * blendCross + sourceScore * blendVector, 0);
 				reranked.push({
 					...candidate,
 					score: blendedScore,
@@ -27997,8 +27955,7 @@ Object.assign(MemoryRetriever.prototype, {
 			log$4.debug("reranked", {
 				totalCandidateCount: candidates.length,
 				sentToRerankCount: toRerank.length,
-				returnedCount: reranked.length,
-				provider
+				returnedCount: reranked.length
 			}, {
 				event_name: "memory.retriever_rerank.diagnostic",
 				file: "packages/memory/src/engine/retrieval/retriever-rerank.ts",
@@ -28041,10 +27998,7 @@ Object.assign(MemoryRetriever.prototype, {
 			}
 			return {
 				candidates: unranked(),
-				fallback: {
-					reason,
-					provider
-				}
+				fallback: { reason }
 			};
 		}
 	},
@@ -28053,8 +28007,8 @@ Object.assign(MemoryRetriever.prototype, {
 		if (candidates.length === 0) return candidates;
 		const chunkIds = candidates.map((c) => c.chunkId).filter((id) => id !== void 0);
 		const vectorMap = chunkIds.length > 0 ? this.store.getVectorsByIds(chunkIds) : /* @__PURE__ */ new Map();
-		const wF = this.config.lightweightFusionWeight ?? .7;
-		const wC = this.config.lightweightCosineWeight ?? .3;
+		const wF = this.config.lightweightFusionWeight ?? .1;
+		const wC = this.config.lightweightCosineWeight ?? .9;
 		let blendedCount = 0;
 		let dimMismatchCount = 0;
 		const reranked = candidates.map((candidate) => {
@@ -28114,6 +28068,58 @@ Object.assign(MemoryRetriever.prototype, {
 			return a.entry.id.localeCompare(b.entry.id);
 		});
 	},
+	/**
+	* The local reranker model reorders the first candidates of the local ranking. Their order is unchanged while the
+	* model is not loaded yet (the first call starts loading it), past the caller's deadline, and after a cancel.
+	*/
+	async rerankLocalModel(query, ordered, deadlineAt, signal) {
+		if (ordered.length < 2 || signal?.aborted || deadlineAt !== void 0 && Date.now() >= deadlineAt) return { candidates: ordered };
+		const head = ordered.slice(0, 15);
+		const started = performance.now();
+		const budgetMs = deadlineAt === void 0 ? this.config.rerankTimeoutMs ?? 15e3 : Math.max(0, Math.floor(deadlineAt - Date.now()));
+		const stop = signal ? AbortSignal.any([signal, AbortSignal.timeout(budgetMs)]) : AbortSignal.timeout(budgetMs);
+		const scoring = this.embedder.scoreRelevance(query, head.map((r) => r.entry.text), stop).catch((error) => {
+			log$4.warn("local reranker failed; recall keeps the local ranking order", { error }, {
+				event_name: "memory.retriever_rerank.diagnostic",
+				file: "packages/memory/src/engine/retrieval/retriever-rerank.ts",
+				function: "rerankLocalModel",
+				site_id: "retrieval.retriever-rerank.rerankLocalModel.failed"
+			});
+		});
+		const logits = await Promise.race([scoring, new Promise((resolve) => stop.addEventListener("abort", () => resolve(void 0), { once: true }))]);
+		if (!logits || signal?.aborted) return { candidates: ordered };
+		const blended = head.map((candidate, index) => ({
+			candidate,
+			value: logits[index] ?? Number.NEGATIVE_INFINITY
+		}));
+		blended.sort((a, b) => b.value - a.value || a.candidate.entry.id.localeCompare(b.candidate.entry.id));
+		const slots = head.map((candidate) => candidate.score).sort((a, b) => b - a);
+		const rescored = blended.map(({ candidate }, index) => {
+			const score = slots[index] ?? candidate.score;
+			return {
+				...candidate,
+				score,
+				rerankScore: score,
+				sources: {
+					...candidate.sources,
+					reranked: { score }
+				}
+			};
+		});
+		log$4.debug("local reranker applied", {
+			scored: head.length,
+			duration_ms: performance.now() - started
+		}, {
+			event_name: "memory.retriever_rerank.diagnostic",
+			file: "packages/memory/src/engine/retrieval/retriever-rerank.ts",
+			function: "rerankLocalModel",
+			site_id: "retrieval.retriever-rerank.rerankLocalModel.applied"
+		});
+		return {
+			candidates: [...rescored, ...ordered.slice(head.length)],
+			modelOrder: blended.map(({ candidate }) => candidate.entry.id)
+		};
+	},
 	getRerankSourceScore(result) {
 		return clamp01(Math.max(result.sources.bm25?.score ?? 0, result.sources.vector?.score ?? 0), 0);
 	}
@@ -28305,8 +28311,21 @@ function tracedStage(trace, name, skipReason, run, input) {
 	trace.endStage(output.map((result) => result.entry.id), output.map((result) => result.score), skipReason === void 0 ? void 0 : { skipped: skipReason });
 	return output;
 }
+/** The candidates the local reranker model scored take their own scores, highest first, in its order. */
+function orderByLocalModel(results, modelOrder) {
+	const present = new Set(results.map((result) => result.entry.id));
+	const order = modelOrder.filter((id) => present.has(id));
+	if (order.length < 2) return results;
+	const judged = new Set(order);
+	const slots = results.filter((result) => judged.has(result.entry.id)).map((result) => result.score).sort((a, b) => b - a);
+	const scores = new Map(order.map((id, index) => [id, slots[index] ?? 0]));
+	return results.map((result) => ({
+		...result,
+		score: scores.get(result.entry.id) ?? result.score
+	})).sort((a, b) => b.score - a.score);
+}
 Object.assign(MemoryRetriever.prototype, {
-	applyScoringPipeline(results, trace, limit = results.length) {
+	applyScoringPipeline(results, trace, limit = results.length, modelOrder) {
 		const stage = (name, skipReason, run, input) => tracedStage(trace, name, skipReason, run, input);
 		const temporalOff = !this.config.temporalWeighting;
 		let scored = stage("recency_boost", temporalOff ? "temporalWeighting" : this.config.recencyHalfLifeDays <= 0 ? "recencyHalfLifeDays" : this.config.recencyWeight <= 0 ? "recencyWeight" : void 0, (input) => temporalOff ? input : this.applyRecencyBoost(input), results);
@@ -28314,6 +28333,7 @@ Object.assign(MemoryRetriever.prototype, {
 		scored = stage("length_normalization", this.config.lengthNormAnchor <= 0 ? "lengthNormAnchor" : void 0, (input) => this.applyLengthNormalization(input), scored);
 		scored = stage("time_decay", temporalOff ? "temporalWeighting" : this.config.timeDecayHalfLifeDays <= 0 ? "timeDecayHalfLifeDays" : void 0, (input) => temporalOff ? input : this.applyTimeDecay(input), scored);
 		scored = stage("retention_boost", temporalOff ? "temporalWeighting" : this.config.recallLifecycle?.retentionScorer === true ? void 0 : "recallLifecycle.retentionScorer", (input) => temporalOff ? input : this.applyRetentionBoost(input), scored);
+		scored = stage("local_model_order", modelOrder ? void 0 : "modelOrder", (input) => modelOrder ? orderByLocalModel(input, modelOrder) : input, scored);
 		scored = stage("hard_min_score", void 0, (input) => input.filter((result) => result.score >= this.config.hardMinScore), scored);
 		const mmrOff = experimentMmrDisabled();
 		scored = stage("mmr_diversity", mmrOff ? "experiment.disableMmr" : void 0, (input) => {