@snoai/memory 1.5.1 → 1.7.0

package/dist/{rem-trigger-BKr4Qmhn.js → rem-trigger-pg_cCp5U.js}

@@ -1,4 +1,4 @@
-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-BhlMin82.js";
+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 ensureModelDownloaded, Ct as FIXED_MEMORY_CHUNKS_MEM_FACET_IDX, D as settingsToPluginConfig, Dn as DEFAULT_IMPORTANCE, Dt as FIXED_MEMORY_MEMORIES_FACT_ID_INSERT_GUARD, En as DEFAULT_BM25_WEIGHT, 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, It as FIXED_MEMORY_SNO_EXTRACT_CHAT, J as FIXED_EXTERNAL_VALUE_58, 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, 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 LocalReranker, 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, Xn as MAX_CHUNK_FETCH_LIMIT, Y as FIXED_EXTERNAL_VALUE_62, Z as FIXED_EXTERNAL_VALUE_9, Zn as MAX_LIST_LIMIT, _r as CachedEmbeddingProvider, _t as FIXED_MEMORY_CHUNKS, a as REM_MODEL_OUTPUT_TOKEN_CAP, an as PERSISTED_PROVIDER_SYSTEM, ar as RECENCY_WEIGHT, at as FIXED_IDX_MEMORY_MEMORIES_LANE_PROJECT, bn as CHUNKING_VERSION$1, bt as FIXED_MEMORY_CHUNKS_AU, cn as SUPPORTED_LOCALES, cr as RETRIEVAL_STORAGE_CHUNK_PROFILE, 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, et as FIXED_IDX_MEMORY_ACTIVE_TASK_EVIDENCE_INSTANCE, ft as FIXED_MEMORY_ACTIVE_TASK_INSTANCES, g as readRemOperationalConfig, gt as FIXED_MEMORY_ACTIVE_TASK_TRANSITIONS, hr as experimentMmrDisabled, ht as FIXED_MEMORY_ACTIVE_TASK_REVISIONS, in as PERSISTED_CONTENT_HASH_V3, it as FIXED_IDX_MEMORY_MEMORIES_FACT_ID, jr as REM_SIDECAR_TOKEN_HEADER, 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, mr as experimentCandidatePoolOverride, mt as FIXED_MEMORY_ACTIVE_TASK_MIGRATION_MANIFESTS, nn as PERSISTED_CANDIDATE_HASH_V1, nr as MULTI_HIT_BONUS_CAP, nt as FIXED_IDX_MEMORY_MEMORIES_CATEGORY, o as REM_RUN_PATH, ot as FIXED_IDX_MEMORY_MEMORIES_PROJECT, pt as FIXED_MEMORY_ACTIVE_TASK_MIGRATION_EVIDENCE, q as FIXED_EXTERNAL_VALUE_57, qt as FIXED_PROTOCOL_VALUE_73, r as REM_CORRELATION_ID_HEADER, rn as PERSISTED_CONTENT_HASH_V1, rr as MULTI_HIT_BONUS_PER_HIT, 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, tr as MMR_LAMBDA, tt as FIXED_IDX_MEMORY_ACTIVE_TASK_INSTANCES_PROJECTION, un as ADJACENCY_BONUS_CAP, ur as TIME_DECAY_FLOOR, ut as FIXED_MEMORY, vn as BONUS_TOTAL_INVARIANT, vr as LOCAL_EMBEDDING_CACHE_DIR_DEFAULT, vt as FIXED_MEMORY_CHUNKS_AD, wn as CHUNK_TARGET_TOKENS, wt as FIXED_MEMORY_CHUNKS_MEM_IDX, xn as CHUNK_MAX_TOKENS, xr as LocalEmbedProvider, xt as FIXED_MEMORY_CHUNKS_FTS, yn as CACHE_TTL_MS, yr as LOCAL_EMBEDDING_MODEL, yt as FIXED_MEMORY_CHUNKS_AI, zt as FIXED_MEMORY_UNPLACED_CANDIDATES } from "./config-BwUcq2zS.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";
@@ -6662,11 +6662,12 @@ function withProviderResponses(responses, run) {
 }
 /** Sno GPU request attribution (GPU_REQUEST_METADATA_STANDARD): self-reported labels, never identity. */
 const SNO_GPU_SOURCE = "sno-station";
-function snoGpuAttributionHeaders(clientRequestId, callId) {
+function snoGpuAttributionHeaders(clientRequestId, callId, usageClass) {
 	return {
 		"X-Sno-Source": SNO_GPU_SOURCE,
 		"X-Sno-Workload": callId ? MODEL_CALL_WORKLOAD[callId] : "station-aux",
-		"X-Request-ID": clientRequestId
+		"X-Request-ID": clientRequestId,
+		...usageClass ? { "X-Sno-Usage-Class": usageClass } : {}
 	};
 }
 async function observeRequest(input, run) {
@@ -6882,7 +6883,7 @@ async function callSnoProfileCompletion(request) {
 				headers: {
 					"Content-Type": "application/json",
 					"X-Internal-Token": request.apiKey,
-					...snoGpuAttributionHeaders(diagnostic.clientRequestId, request.callId),
+					...snoGpuAttributionHeaders(diagnostic.clientRequestId, request.callId, request.usageClass),
 					...request.operationId ? { "X-Sno-Operation": request.operationId } : {}
 				},
 				body: JSON.stringify({
@@ -6943,6 +6944,32 @@ function isCcproxyOpenAiBaseUrl(value) {
 		return false;
 	}
 }
+/** Joins a chat-completion SSE stream into the body a non-streamed reply carries; `: ping` comments are skipped. */
+function assembleChatStream(text) {
+	const body = {};
+	let message;
+	for (const line of text.split("\n")) {
+		if (!line.startsWith("data:")) continue;
+		const data = line.slice(5).trim();
+		if (data === "[DONE]") break;
+		const chunk = JSON.parse(data);
+		body.id ??= chunk.id;
+		body.model ??= chunk.model;
+		if (chunk.usage) body.usage = chunk.usage;
+		const choice = chunk.choices?.[0];
+		if (!choice) continue;
+		message ??= { content: "" };
+		if (typeof choice.delta?.content === "string") message.content += choice.delta.content;
+		if (choice.finish_reason) message.finishReason = choice.finish_reason;
+	}
+	return {
+		...body,
+		choices: message ? [{
+			message: { content: message.content },
+			finish_reason: message.finishReason
+		}] : []
+	};
+}
 async function snoStationMemDispatch(ctx) {
 	const cfgExt = ctx.config;
 	const provider = ctx.provider;
@@ -6982,7 +7009,7 @@ async function snoStationMemDispatch(ctx) {
 		if (typeof cfgExt.requestId === "string") headers["X-Request-ID"] = cfgExt.requestId;
 		if (provider === "sno-gpu") {
 			headers["X-Internal-Token"] = ctx.apiKey;
-			Object.assign(headers, snoGpuAttributionHeaders(diagnostic.clientRequestId, cfgExt.callId));
+			Object.assign(headers, snoGpuAttributionHeaders(diagnostic.clientRequestId, cfgExt.callId, cfgExt.usageClass));
 			if (typeof cfgExt.operationId === "string") headers["X-Sno-Operation"] = cfgExt.operationId;
 		} else headers.Authorization = `Bearer ${ctx.apiKey}`;
 		let response;
@@ -6994,7 +7021,11 @@ async function snoStationMemDispatch(ctx) {
 				body: JSON.stringify({
 					model: ctx.model,
 					messages: ctx.messages,
-					...callBody
+					...callBody,
+					...provider === "sno-gpu" ? {
+						stream: true,
+						stream_options: { include_usage: true }
+					} : {}
 				})
 			});
 			diagnostic.headersAt = performance.now();
@@ -7012,7 +7043,7 @@ async function snoStationMemDispatch(ctx) {
 		}
 		let body;
 		try {
-			body = await response.json();
+			body = response.headers.get("content-type")?.includes("text/event-stream") ? assembleChatStream(await response.text()) : await response.json();
 			diagnostic.bodyAt = performance.now();
 		} catch (parseErr) {
 			if (abortedProviderCategory(requestAbort)) throwTerminalAbort(parseErr, requestAbort);
@@ -7058,6 +7089,7 @@ async function callProvider(ctx) {
 		if (typeof raw.endpointUrl !== "string" || typeof raw.prompt !== "string" || typeof raw.timeoutMs !== "number") throw new Error("sno-station-mem llm-client: invalid raw profile completion config");
 		return {
 			...await callSnoProfileCompletion({
+				...cfgExt.usageClass ? { usageClass: cfgExt.usageClass } : {},
 				...cfgExt.promptTemplateHash ? { promptTemplateHash: cfgExt.promptTemplateHash } : {},
 				...cfgExt.extractionSkillHash ? { extractionSkillHash: cfgExt.extractionSkillHash } : {},
 				endpointUrl: raw.endpointUrl,
@@ -7220,6 +7252,7 @@ function buildPipelineConfig(preset, config, endpoint, request, destination, ope
 		destination,
 		...request.requestId ? { requestId: request.requestId } : {},
 		...operationId ? { operationId } : {},
+		...config.usageClass ? { usageClass: config.usageClass } : {},
 		...request.promptTemplateHash ? { promptTemplateHash: request.promptTemplateHash } : {},
 		...request.extractionSkillHash ? { extractionSkillHash: request.extractionSkillHash } : {},
 		...request.signal ? { signal: request.signal } : {},
@@ -7332,12 +7365,12 @@ function createLlmClient(config) {
 	let lastError = null;
 	let lastUsage = null;
 	const resolveRequestTimeoutMs = (request) => {
-		if (destinationFor(request) !== "sno-gpu" || !request.callId.startsWith("E") || MODEL_CALLS[request.callId].occasion !== "memoryExtract") return request.timeoutMs ?? config.timeoutMs ?? 3e4;
+		if (!request.callId.startsWith("E") || MODEL_CALLS[request.callId].occasion !== "memoryExtract") return request.timeoutMs ?? config.timeoutMs ?? 3e4;
 		const operationDeadline = memoryOperationDeadline();
 		if (operationDeadline === void 0) return request.timeoutMs ?? 3e5;
 		const remainingMs = operationDeadline - performance.now();
 		if (remainingMs <= 0) throw new LlmClientTerminalError("timeout", "Memory operation deadline exhausted before extraction");
-		return Math.min(request.timeoutMs ?? remainingMs, remainingMs);
+		return Math.floor(Math.min(request.timeoutMs ?? remainingMs, remainingMs));
 	};
 	const destinationFor = (request) => {
 		if (!config.routing) return "host";
@@ -7665,7 +7698,7 @@ function createLlmClient(config) {
 			const attemptLimit = request.emptyReplyAttempts ?? EMPTY_REPLY_ATTEMPTS;
 			for (let attempt = 2; attempt <= attemptLimit; attempt += 1) {
 				if (content === null || content.raw.trim().length > 0 || budget.remaining <= 0) break;
-				const remainingMs = deadlineMs - performance.now();
+				const remainingMs = Math.floor(deadlineMs - performance.now());
 				if (remainingMs <= 0) break;
 				log$20.debug("Retrying empty model reply", {
 					call_id: request.callId,
@@ -26704,9 +26737,8 @@ function buildConfigSnapshot() {
 		HEAD_EXTRACT_OVERLAP_DROP_RATIO,
 		HEAD_EXTRACT_TOKEN_BUDGET,
 		INTRA_MEMORY_COSINE_THRESHOLD,
-		LIGHTWEIGHT_COSINE_WEIGHT,
-		LIGHTWEIGHT_FUSION_WEIGHT,
-		LIGHTWEIGHT_RERANK_PENALTY,
+		LIGHTWEIGHT_COSINE_WEIGHT: 1,
+		LIGHTWEIGHT_FUSION_WEIGHT: 0,
 		MAX_CHUNKS_PER_PARENT: 16,
 		MMR_LAMBDA,
 		MULTI_HIT_BONUS_CAP,
@@ -27219,7 +27251,7 @@ const DEFAULT_RETRIEVAL_CONFIG = {
 	vectorWeight: DEFAULT_VECTOR_WEIGHT,
 	bm25Weight: DEFAULT_BM25_WEIGHT,
 	minScore: 0,
-	rerank: "cross-encoder",
+	rerank: "lightweight",
 	candidatePoolSize: 64,
 	recencyHalfLifeDays: 14,
 	recencyWeight: RECENCY_WEIGHT,
@@ -27229,10 +27261,8 @@ const DEFAULT_RETRIEVAL_CONFIG = {
 	lengthNormAnchor: 0,
 	hardMinScore: 0,
 	timeDecayHalfLifeDays: 60,
-	rerankBlendVector: RERANK_BLEND_VECTOR,
-	rerankBlendCross: RERANK_BLEND_CROSS,
-	lightweightFusionWeight: LIGHTWEIGHT_FUSION_WEIGHT,
-	lightweightCosineWeight: LIGHTWEIGHT_COSINE_WEIGHT,
+	lightweightFusionWeight: 0,
+	lightweightCosineWeight: 1,
 	importanceWeightBase: IMPORTANCE_WEIGHT_BASE,
 	timeDecayFloor: TIME_DECAY_FLOOR,
 	mmrLambda: MMR_LAMBDA,
@@ -27528,11 +27558,7 @@ Object.assign(MemoryRetriever.prototype, {
 		});
 		fused = rerankOutcome.candidates;
 		this.throwIfAborted(context);
-		const rerankStageMetadata = rerankOutcome.fallback ? {
-			...rerankOutcome.stats,
-			rerankFallbackReason: rerankOutcome.fallback.reason
-		} : rerankOutcome.stats;
-		trace?.endStage(fused.map((r) => r.entry.id), fused.map((r) => r.score), rerankStageMetadata);
+		trace?.endStage(fused.map((r) => r.entry.id), fused.map((r) => r.score));
 		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));
@@ -27601,423 +27627,23 @@ Object.assign(MemoryRetriever.prototype, {
 	}
 });
 //#endregion
-//#region src/engine/retrieval/retrieval-rerank-provider.ts
-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
-		}
-	};
-}
-/** 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 ["score", "relevance_score"]) {
-				const value = raw?.[key];
-				const n = typeof value === "number" ? value : Number(value);
-				if (Number.isFinite(n)) {
-					score = n;
-					break;
-				}
-			}
-			if (score === void 0) continue;
-			parsed.push({
-				index,
-				score
-			});
-		}
-		return parsed.length > 0 ? parsed : void 0;
-	};
-	const objectData = data && typeof data === "object" && !Array.isArray(data) ? data : void 0;
-	return parseItems(data) ?? parseItems(objectData?.results) ?? parseItems(objectData?.data);
-}
-//#endregion
 //#region src/engine/retrieval/retriever-rerank.ts
 /** @file retriever-rerank.ts
-* @purpose Runs remote and local reranking.
+* @purpose Runs local cosine ordering and local model reranking.
 * @boundary Prototype-mounted MemoryRetriever methods; no constructor state ownership.
 */
-const RERANK_TRANSIENT_ATTEMPTS = 3;
-const RERANK_TRANSIENT_BACKOFF_MS = 200;
-const RERANK_TRANSIENT_STATUSES = /* @__PURE__ */ new Set([
-	429,
-	502,
-	503,
-	504
-]);
-const RERANK_RETRY_AFTER_MAX_MS = 5e3;
-function isRerankTimeout(error) {
-	return error instanceof Error && (error.name === "TimeoutError" || error.name === "AbortError");
-}
-/**
-* The reranker's refusal when a pair is over its window, as the endpoint states it:
-* `{"detail":{"code":"input_token_limit_exceeded","max_input_tokens":512,
-* "candidates":[{"index":0,"input_tokens":728,"excess_tokens":216}]}}` (measured 2026-09-14).
-*
-* It names every offending text, so the batch can be repaired and sent again instead of
-* losing the scores of the 49 texts that were fine. The counts are the SERVER's, taken with
-* the ranking model's tokenizer — which is not the embedder's — so they are the only exact
-* measure available here of how far over the window a text is.
-*/
-const RERANK_OVER_WINDOW_CODE = "input_token_limit_exceeded";
-function parseOverWindowRefusal(body) {
-	let parsed;
-	try {
-		parsed = JSON.parse(body);
-	} catch {
-		return;
-	}
-	if (typeof parsed !== "object" || parsed === null) return void 0;
-	const detail = parsed.detail;
-	if (typeof detail !== "object" || detail === null) return void 0;
-	const shape = detail;
-	if (shape.code !== RERANK_OVER_WINDOW_CODE) return void 0;
-	if (typeof shape.max_input_tokens !== "number" || !Array.isArray(shape.candidates)) return;
-	const overLong = [];
-	for (const candidate of shape.candidates) {
-		if (typeof candidate !== "object" || candidate === null) continue;
-		const named = candidate;
-		if (typeof named.index !== "number" || typeof named.input_tokens !== "number") continue;
-		overLong.push({
-			index: named.index,
-			inputTokens: named.input_tokens
-		});
-	}
-	return overLong.length === 0 ? void 0 : {
-		maxInputTokens: shape.max_input_tokens,
-		overLong
-	};
-}
-function isTransientRerankError(error) {
-	if (isRerankTimeout(error)) return true;
-	return error instanceof TypeError && error.cause !== void 0;
-}
-async function fetchRerankWithRetry(send, stop) {
-	let lastError;
-	for (let attempt = 1; attempt <= RERANK_TRANSIENT_ATTEMPTS; attempt += 1) {
-		let retryDelayMs = RERANK_TRANSIENT_BACKOFF_MS * attempt;
-		try {
-			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");
-				if (retryAfter !== null && /^\d+$/.test(retryAfter)) retryDelayMs = Math.min(Number.parseInt(retryAfter, 10) * 1e3, RERANK_RETRY_AFTER_MAX_MS);
-			}
-			await response.body?.cancel();
-			lastError = /* @__PURE__ */ new Error(`rerank API responded ${response.status}`);
-		} catch (error) {
-			if (!isTransientRerankError(error) || attempt === RERANK_TRANSIENT_ATTEMPTS || stop?.aborted) throw error;
-			lastError = error;
-		}
-		if (stop?.aborted) throw lastError;
-		await new Promise((resolve) => {
-			const onAbort = () => {
-				clearTimeout(timer);
-				resolve();
-			};
-			const timer = setTimeout(() => {
-				stop?.removeEventListener("abort", onAbort);
-				resolve();
-			}, retryDelayMs);
-			stop?.addEventListener("abort", onAbort, { once: true });
-		});
-		if (stop?.aborted) throw lastError;
-	}
-	throw lastError;
-}
 Object.assign(MemoryRetriever.prototype, {
 	async rerank(query, candidates, queryVector, options) {
-		if (candidates.length === 0) return { candidates };
-		if (this.config.rerank === "none") return { candidates };
-		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 reranker address", {}, {
-				event_name: "memory.retriever_rerank.diagnostic",
-				file: "packages/memory/src/engine/retrieval/retriever-rerank.ts",
-				function: "rerank",
-				site_id: "retrieval.retriever-rerank.rerank.6fbc6b3a22"
-			});
-			return {
-				candidates: unranked(),
-				fallback: { reason: "no_endpoint" }
-			};
-		}
-		const rerankTimeoutMs = this.config.rerankTimeoutMs ?? 15e3;
-		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;
-		const toRerank = overCap ? candidates.slice(0, maxCandidates) : candidates;
-		const beyondCap = overCap ? candidates.slice(maxCandidates) : [];
-		const batchSize = 50;
-		const queryTokens = this.embedder.countTokens(query);
-		const documentTokenBudget = 509 - queryTokens;
-		if (documentTokenBudget <= 0) {
-			log$4.warn("rerank skipped: the query alone fills the reranker token window", {
-				queryTokens,
-				window: 512,
-				templateTokens: 3
-			}, {
-				event_name: "memory.retriever_rerank.diagnostic",
-				file: "packages/memory/src/engine/retrieval/retriever-rerank.ts",
-				function: "rerank",
-				site_id: "retrieval.retriever-rerank.rerank.query_over_window"
-			});
-			return {
-				candidates: unranked(),
-				fallback: { reason: "query_over_window" }
-			};
-		}
-		let truncatedCandidates = 0;
-		const rerankTexts = toRerank.map((c) => {
-			const bounded = this.embedder.truncateToTokens(c.entry.text, documentTokenBudget);
-			if (bounded !== c.entry.text) truncatedCandidates += 1;
-			return bounded;
-		});
-		if (truncatedCandidates > 0) log$4.warn("rerank candidates were cut to the document budget the query left", {
-			truncatedCandidates,
-			documentTokenBudget,
-			queryTokens
-		}, {
-			event_name: "memory.retriever_rerank.diagnostic",
-			file: "packages/memory/src/engine/retrieval/retriever-rerank.ts",
-			function: "rerank",
-			site_id: "retrieval.retriever-rerank.rerank.candidate_over_ceiling"
-		});
-		try {
-			const runBatch = async (start) => {
-				let batchTexts = rerankTexts.slice(start, start + batchSize);
-				for (let pass = 0;; pass += 1) {
-					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 errorBody = await response.text().catch(() => "<unreadable body>");
-						const refusal = pass === 0 ? parseOverWindowRefusal(errorBody) : void 0;
-						if (refusal) {
-							const allowed = refusal.maxInputTokens - 3;
-							batchTexts = batchTexts.map((text, index) => {
-								const named = refusal.overLong.find((c) => c.index === index);
-								if (!named) return text;
-								const localTokens = this.embedder.countTokens(text);
-								const localPairTokens = queryTokens + localTokens;
-								const target = Math.max(1, Math.floor(allowed * localPairTokens / named.inputTokens) - queryTokens);
-								return this.embedder.truncateToTokens(text, target);
-							});
-							log$4.warn("rerank batch refused over its token window; resending the named texts cut to fit", {
-								status: response.status,
-								overLongCount: refusal.overLong.length,
-								maxInputTokens: refusal.maxInputTokens,
-								sentToRerankCount: batchTexts.length,
-								batchStart: start
-							}, {
-								event_name: "memory.retriever_rerank.diagnostic",
-								file: "packages/memory/src/engine/retrieval/retriever-rerank.ts",
-								function: "runBatch",
-								site_id: "retrieval.retriever-rerank.runBatch.over_window_repair"
-							});
-							continue;
-						}
-						log$4.warn("rerank API failed, using pre-rerank results", {
-							status: response.status,
-							endpoint_hash: createHash("sha256").update(rerankEndpoint).digest("hex"),
-							retryAfter,
-							sentToRerankCount: batchTexts.length,
-							batchStart: start,
-							body_length: errorBody.length
-						}, {
-							event_name: "memory.retriever_rerank.diagnostic",
-							file: "packages/memory/src/engine/retrieval/retriever-rerank.ts",
-							function: "runBatch",
-							site_id: "retrieval.retriever-rerank.runBatch.76d7d1fac6"
-						});
-						return "http_error";
-					}
-					const parsed = parseRerankResponse(data);
-					if (!parsed) {
-						log$4.warn("rerank API returned unparseable response", {
-							endpoint_hash: createHash("sha256").update(rerankEndpoint).digest("hex"),
-							batchStart: start
-						}, {
-							event_name: "memory.retriever_rerank.diagnostic",
-							file: "packages/memory/src/engine/retrieval/retriever-rerank.ts",
-							function: "runBatch",
-							site_id: "retrieval.retriever-rerank.runBatch.4429b8bd8a"
-						});
-						return "invalid_response";
-					}
-					const batchItems = [];
-					for (const item of parsed) {
-						if (item.index < 0 || item.index >= batchTexts.length) continue;
-						batchItems.push({
-							index: start + item.index,
-							score: item.score
-						});
-					}
-					return batchItems;
-				}
-			};
-			const batchStarts = [];
-			for (let start = 0; start < rerankTexts.length; start += batchSize) batchStarts.push(start);
-			const items = [];
-			for (let wave = 0; wave < batchStarts.length; wave += 4) {
-				const waveStarts = batchStarts.slice(wave, wave + 4);
-				const settled = await Promise.allSettled(waveStarts.map(runBatch));
-				for (const outcome of settled) if (outcome.status === "rejected" && outcome.reason instanceof RetrievalError) throw outcome.reason;
-				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 }
-				};
-				for (const outcome of settled) if (outcome.status === "fulfilled" && typeof outcome.value !== "string") items.push(...outcome.value);
-			}
-			const blendCross = this.config.rerankBlendCross ?? .7;
-			const blendVector = this.config.rerankBlendVector ?? .3;
-			const returnedIndices = new Set(items.map((r) => r.index));
-			const reranked = [];
-			const seenIndices = /* @__PURE__ */ new Set();
-			for (const item of items) {
-				if (seenIndices.has(item.index)) continue;
-				seenIndices.add(item.index);
-				const candidate = toRerank[item.index];
-				if (!candidate) continue;
-				const sourceScore = this.getRerankSourceScore(candidate);
-				const blendedScore = clamp01(1 / (1 + Math.exp(-item.score)) * blendCross + sourceScore * blendVector, 0);
-				reranked.push({
-					...candidate,
-					score: blendedScore,
-					sources: {
-						...candidate.sources,
-						reranked: { score: item.score }
-					}
-				});
-			}
-			const unreturned = [...toRerank.filter((_, idx) => !returnedIndices.has(idx)), ...beyondCap].map((r) => ({
-				...r,
-				score: clamp01(this.getRerankSourceScore(r) * LIGHTWEIGHT_RERANK_PENALTY, 0)
-			}));
-			const merged = [...reranked, ...unreturned].sort((a, b) => b.score - a.score);
-			log$4.debug("reranked", {
-				totalCandidateCount: candidates.length,
-				sentToRerankCount: toRerank.length,
-				returnedCount: reranked.length
-			}, {
-				event_name: "memory.retriever_rerank.diagnostic",
-				file: "packages/memory/src/engine/retrieval/retriever-rerank.ts",
-				function: "rerank",
-				site_id: "retrieval.retriever-rerank.rerank.7847c6248a"
-			});
-			return {
-				candidates: merged.length > 0 ? merged : candidates,
-				stats: {
-					rerankSentCount: toRerank.length,
-					rerankReturnedCount: reranked.length,
-					rerankBeyondCapCount: beyondCap.length
-				}
-			};
-		} catch (error) {
-			if (error instanceof RetrievalError) throw error;
-			let reason;
-			if (isRerankTimeout(error)) {
-				reason = "timeout";
-				log$4.warn("rerank API timed out, using pre-rerank results", {
-					timeoutMs: rerankTimeoutMs,
-					attempts: RERANK_TRANSIENT_ATTEMPTS
-				}, {
-					event_name: "memory.retriever_rerank.diagnostic",
-					file: "packages/memory/src/engine/retrieval/retriever-rerank.ts",
-					function: "rerank",
-					site_id: "retrieval.retriever-rerank.rerank.4af9c17c75"
-				});
-			} else {
-				reason = "request_error";
-				log$4.warn("rerank error, using pre-rerank results", {
-					error,
-					attempts: RERANK_TRANSIENT_ATTEMPTS
-				}, {
-					event_name: "memory.retriever_rerank.diagnostic",
-					file: "packages/memory/src/engine/retrieval/retriever-rerank.ts",
-					function: "rerank",
-					site_id: "retrieval.retriever-rerank.rerank.2c5fc67dbf"
-				});
-			}
-			return {
-				candidates: unranked(),
-				fallback: { reason }
-			};
-		}
+		if (candidates.length === 0 || this.config.rerank === "none") return { candidates };
+		return this.rerankLocalModel(query, this.rerankLightweight(candidates, queryVector), options?.deadlineAt, options?.signal);
 	},
 	rerankLightweight(candidates, queryVector) {
 		if (queryVector.length === 0) return candidates;
 		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 ?? .1;
-		const wC = this.config.lightweightCosineWeight ?? .9;
+		const wF = this.config.lightweightFusionWeight ?? 0;
+		const wC = this.config.lightweightCosineWeight ?? 1;
 		let blendedCount = 0;
 		let dimMismatchCount = 0;
 		const reranked = candidates.map((candidate) => {
@@ -28083,9 +27709,10 @@ Object.assign(MemoryRetriever.prototype, {
 	*/
 	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 head = ordered.slice(0, 32);
 		const started = performance.now();
-		const budgetMs = deadlineAt === void 0 ? this.config.rerankTimeoutMs ?? 15e3 : Math.max(0, Math.floor(deadlineAt - Date.now()));
+		const budgetMs = Math.max(0, Math.floor(Math.min(DEFAULT_RERANK_TIMEOUT_MS, this.config.rerankTimeoutMs ?? 2e3, deadlineAt === void 0 ? DEFAULT_RERANK_TIMEOUT_MS : deadlineAt - Date.now())));
+		if (budgetMs === 0) return { candidates: ordered };
 		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 }, {
@@ -28095,11 +27722,23 @@ Object.assign(MemoryRetriever.prototype, {
 				site_id: "retrieval.retriever-rerank.rerankLocalModel.failed"
 			});
 		});
-		const logits = await Promise.race([scoring, new Promise((resolve) => stop.addEventListener("abort", () => resolve(void 0), { once: true }))]);
+		let onAbort = () => {};
+		const aborted = new Promise((resolve) => {
+			onAbort = () => resolve(void 0);
+			stop.addEventListener("abort", onAbort, { once: true });
+		});
+		let logits;
+		try {
+			logits = await Promise.race([scoring, aborted]);
+		} finally {
+			stop.removeEventListener("abort", onAbort);
+		}
 		if (!logits || signal?.aborted) return { candidates: ordered };
+		const minimum = Math.min(...logits);
+		const span = Math.max(1e-9, Math.max(...logits) - minimum);
 		const blended = head.map((candidate, index) => ({
 			candidate,
-			value: logits[index] ?? Number.NEGATIVE_INFINITY
+			value: .5 * (((logits[index] ?? minimum) - minimum) / span) + .5 * candidate.score
 		}));
 		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);
@@ -28128,9 +27767,6 @@ Object.assign(MemoryRetriever.prototype, {
 			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);
 	}
 });
 //#endregion
@@ -28424,7 +28060,7 @@ Object.assign(MemoryRetriever.prototype, {
 	},
 	applyMmrDiversity(results) {
 		if (results.length === 0) return results;
-		const lambda = this.config.mmrLambda ?? .7;
+		const lambda = this.config.mmrLambda ?? .9;
 		const chunkIds = results.map((r) => r.chunkId).filter((id) => id !== void 0);
 		const vectorMap = chunkIds.length > 0 ? this.store.getVectorsByIds(chunkIds) : /* @__PURE__ */ new Map();
 		const sorted = [...results].sort((a, b) => {
@@ -30231,7 +29867,7 @@ async function runWithCanonicalStoreWriteMutex(canonicalStorePath, action) {
 async function openBatchRuntime(input) {
 	const settings = input.settings;
 	const dbPath = settings.store.path;
-	const pluginConfig = settingsToPluginConfig(settings);
+	const pluginConfig = settingsToPluginConfig(settings, process.env.SNO_USAGE_CLASS === "research" ? "research" : "product");
 	initSqliteRuntime(settings.store.encryptionKey);
 	const embedder = createEmbedder(pluginConfig.embedding, getSnoStationMemStateDir());
 	let store;
@@ -30258,6 +29894,7 @@ async function openBatchRuntime(input) {
 		loadStorageExtensions(database);
 		const llm = createLlmClient({
 			preset: FIXED_MEMORY_SNO_EXTRACT_CHAT,
+			usageClass: pluginConfig.extraction.llm.usageClass,
 			operationId: input.jobId,
 			apiKey: settings.snoGpu.apiKey,
 			baseURL: settings.snoGpu.baseUrl,