@betterdb/semantic-cache 0.1.0 → 0.2.0

package/dist/SemanticCache.js

@@ -2,10 +2,14 @@
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.SemanticCache = void 0;
 const node_crypto_1 = require("node:crypto");
+const node_crypto_2 = require("node:crypto");
 const api_1 = require("@opentelemetry/api");
 const errors_1 = require("./errors");
 const telemetry_1 = require("./telemetry");
 const utils_1 = require("./utils");
+const defaultCostTable_1 = require("./defaultCostTable");
+const cluster_1 = require("./cluster");
+const analytics_1 = require("./analytics");
 const INVALIDATE_BATCH_SIZE = 1000;
 function errMsg(err) {
     return err instanceof Error ? err.message : String(err);
@@ -17,15 +21,27 @@ class SemanticCache {
     indexName;
     entryPrefix;
     statsKey;
+    similarityWindowKey;
     defaultThreshold;
     defaultTtl;
     categoryThresholds;
     uncertaintyBand;
     telemetry;
+    costTable;
+    embeddingCacheEnabled;
+    embeddingCacheTtl;
+    embedKeyPrefix;
     _initialized = false;
     _dimension = 0;
+    _hasBinaryRefs = false;
     _initPromise = null;
     _initGeneration = 0;
+    analyticsOpts;
+    usesDefaultCostTable;
+    analytics = analytics_1.NOOP_ANALYTICS;
+    statsTimer;
+    shutdownCalled = false;
+    analyticsInitiated = false;
     /**
      * Creates a new SemanticCache instance.
      *
@@ -42,17 +58,35 @@ class SemanticCache {
         this.indexName = `${this.name}:idx`;
         this.entryPrefix = `${this.name}:entry:`;
         this.statsKey = `${this.name}:__stats`;
+        this.similarityWindowKey = `${this.name}:__similarity_window`;
+        this.embedKeyPrefix = `${this.name}:embed:`;
         this.defaultThreshold = options.defaultThreshold ?? 0.1;
         this.defaultTtl = options.defaultTtl;
         this.categoryThresholds = options.categoryThresholds ?? {};
         this.uncertaintyBand = options.uncertaintyBand ?? 0.05;
+        // Build effective cost table
+        const useDefault = options.useDefaultCostTable ?? true;
+        if (!useDefault && !options.costTable) {
+            this.costTable = undefined;
+        }
+        else if (!useDefault) {
+            this.costTable = options.costTable;
+        }
+        else {
+            this.costTable = { ...defaultCostTable_1.DEFAULT_COST_TABLE, ...(options.costTable ?? {}) };
+        }
+        // Embedding cache config
+        this.embeddingCacheEnabled = options.embeddingCache?.enabled ?? true;
+        this.embeddingCacheTtl = options.embeddingCache?.ttl ?? 86400;
         this.telemetry = (0, telemetry_1.createTelemetry)({
             prefix: options.telemetry?.metricsPrefix ?? 'semantic_cache',
             tracerName: options.telemetry?.tracerName ?? '@betterdb/semantic-cache',
             registry: options.telemetry?.registry,
         });
+        this.analyticsOpts = options.analytics;
+        this.usesDefaultCostTable = useDefault;
     }
-    // ── Lifecycle ──────────────────────────────────────────────
+    // -- Lifecycle --
     async initialize() {
         if (!this._initPromise) {
             this._initPromise = this._doInitialize().catch((err) => {
@@ -65,7 +99,6 @@ class SemanticCache {
     async flush() {
         // Mark uninitialized immediately so concurrent check()/store() calls get
         // a clear SemanticCacheUsageError instead of cryptic Valkey errors.
-        // Bump generation so any in-flight _doInitialize() won't overwrite this state.
         this._initialized = false;
         this._initPromise = null;
         this._initGeneration++;
@@ -79,33 +112,59 @@ class SemanticCache {
                 throw new errors_1.ValkeyCommandError('FT.DROPINDEX', err);
             }
         }
-        const entryPattern = `${this.name}:entry:*`;
-        let cursor = '0';
-        do {
-            const [nextCursor, keys] = await this.client.scan(cursor, 'MATCH', entryPattern, 'COUNT', '100');
-            cursor = nextCursor;
-            if (keys.length > 0)
-                await this.client.del(keys);
-        } while (cursor !== '0');
+        // Cluster-aware SCAN for entry keys and embed cache keys
+        const patterns = [
+            `${this.name}:entry:*`,
+            `${this.name}:embed:*`,
+        ];
+        for (const pattern of patterns) {
+            await (0, cluster_1.clusterScan)(this.client, pattern, async (keys, nodeClient) => {
+                await nodeClient.del(keys);
+            });
+        }
         await this.client.del(this.statsKey);
+        await this.client.del(this.similarityWindowKey);
+        this.analytics.capture('cache_flush');
+    }
+    /** Shut down the analytics client and cancel the stats timer. */
+    async shutdown() {
+        this.shutdownCalled = true;
+        if (this.statsTimer) {
+            clearInterval(this.statsTimer);
+            this.statsTimer = undefined;
+        }
+        await this.analytics.shutdown();
     }
-    // ── Public operations ──────────────────────────────────────
+    // -- Public operations --
     async check(prompt, options) {
         this.assertInitialized('check');
         return this.traced('check', async (span) => {
             const category = options?.category ?? '';
-            const k = options?.k ?? 1;
             const threshold = options?.threshold ??
                 (category && this.categoryThresholds[category] !== undefined
                     ? this.categoryThresholds[category]
                     : this.defaultThreshold);
-            const { vector: embedding, durationSec: embedSec } = await this.embed(prompt);
+            // Resolve text and binary refs from prompt
+            const { text: promptText, binaryRefs } = await this.resolvePrompt(prompt);
+            // Stale model detection
+            const checkStale = (options?.staleAfterModelChange ?? false) && !!options?.currentModel;
+            // Rerank option
+            const rerankOpts = options?.rerank;
+            const k = rerankOpts ? rerankOpts.k : (options?.k ?? 1);
+            const { vector: embedding, durationSec: embedSec } = await this.embed(promptText);
             this.assertDimension(embedding);
-            // FT.SEARCH — Valkey Search 1.2 rejects KNN aliases in RETURN/SORTBY,
-            // so we omit both. Results include all fields and are pre-sorted by distance.
+            // Build filter
+            const userFilter = options?.filter;
+            // AND semantics: each ref must be present — chain separate TAG clauses.
+            const binaryFilter = binaryRefs.length > 0 && this._hasBinaryRefs
+                ? (binaryRefs.length === 1
+                    ? `@binary_refs:{${(0, utils_1.escapeTag)(binaryRefs[0])}}`
+                    : binaryRefs.map((r) => `@binary_refs:{${(0, utils_1.escapeTag)(r)}}`).join(' '))
+                : null;
+            const combinedFilter = [userFilter, binaryFilter].filter(Boolean).join(' ');
+            const filterExpr = combinedFilter ? `(${combinedFilter})` : '*';
+            const query = `${filterExpr}=>[KNN ${k} @embedding $vec AS __score]`;
             const searchStart = performance.now();
-            const filter = options?.filter;
-            const query = `${filter ? `(${filter})` : '*'}=>[KNN ${k} @embedding $vec AS __score]`;
             let rawResult;
             try {
                 rawResult = await this.client.call('FT.SEARCH', this.indexName, query, 'PARAMS', '2', 'vec', (0, utils_1.encodeFloat32)(embedding), 'LIMIT', '0', String(k), 'DIALECT', '2');
@@ -136,6 +195,9 @@ class SemanticCache {
             }
             // Miss (no usable score, or score exceeds threshold)
             if (isNaN(score) || score > threshold) {
+                if (!isNaN(score)) {
+                    await this.recordSimilarityWindow(score, 'miss', category);
+                }
                 await this.recordStat('misses');
                 this.telemetry.metrics.requestsTotal
                     .labels({ cache_name: this.name, result: 'miss', category: categoryLabel }).inc();
@@ -151,42 +213,148 @@ class SemanticCache {
                 }
                 return result;
             }
-            // Hit
-            const confidence = score >= threshold - this.uncertaintyBand ? 'uncertain' : 'high';
+            // Rerank: apply rerankFn to all candidates above threshold
+            let winnerParsedIndex = 0;
+            if (rerankOpts && parsed.length > 0) {
+                // Preserve the original parsed[] index alongside each candidate so we
+                // can map back even when NaN-scored entries are filtered out.
+                const indexedCandidates = parsed
+                    .map((r, i) => ({ i, s: parseFloat(r.fields['__score'] ?? 'NaN') }))
+                    .filter(({ s }) => !isNaN(s))
+                    .map(({ i, s }) => ({
+                    origIdx: i,
+                    candidate: { response: parsed[i].fields['response'] ?? '', similarity: s },
+                }));
+                const picked = await rerankOpts.rerankFn(promptText, indexedCandidates.map((x) => x.candidate));
+                // Explicit bounds check: -1 means "reject all"; out-of-range is a caller bug
+                // treated as a miss rather than silently falling back to the top candidate.
+                if (picked === -1 || picked < 0 || picked >= indexedCandidates.length) {
+                    await this.recordSimilarityWindow(score, 'miss', category);
+                    await this.recordStat('misses');
+                    this.telemetry.metrics.requestsTotal
+                        .labels({ cache_name: this.name, result: 'miss', category: categoryLabel }).inc();
+                    span.setAttributes({ 'cache.hit': false, 'cache.name': this.name, 'cache.reranked': true });
+                    return { hit: false, confidence: 'miss' };
+                }
+                // Map back to the original parsed[] index (not the candidates[] index)
+                winnerParsedIndex = indexedCandidates[picked].origIdx;
+            }
+            const winner = parsed[winnerParsedIndex] ?? parsed[0];
+            const winnerScore = parseFloat(winner.fields['__score'] ?? String(score));
+            // Stale model check: if winner's model differs from currentModel, evict and treat as miss
+            if (checkStale) {
+                const storedModel = winner.fields['model'] ?? '';
+                if (storedModel && storedModel !== options.currentModel) {
+                    // Evict stale entry
+                    try {
+                        await this.client.del(winner.key);
+                    }
+                    catch { /* best effort */ }
+                    await this.recordSimilarityWindow(winnerScore, 'miss', category);
+                    this.telemetry.metrics.staleModelEvictions.labels({ cache_name: this.name }).inc();
+                    await this.recordStat('misses');
+                    this.telemetry.metrics.requestsTotal
+                        .labels({ cache_name: this.name, result: 'miss', category: categoryLabel }).inc();
+                    span.setAttributes({ 'cache.hit': false, 'cache.stale_evicted': true });
+                    return { hit: false, confidence: 'miss' };
+                }
+            }
+            // All checks passed — record as a genuine hit
+            await this.recordSimilarityWindow(winnerScore, 'hit', category);
+            const confidence = winnerScore >= threshold - this.uncertaintyBand ? 'uncertain' : 'high';
             await this.recordStat('hits');
             const metricResult = confidence === 'uncertain' ? 'uncertain_hit' : 'hit';
             this.telemetry.metrics.requestsTotal
                 .labels({ cache_name: this.name, result: metricResult, category: categoryLabel }).inc();
-            const matchedKey = parsed[0].key;
+            const matchedKey = winner.key;
             if (this.defaultTtl !== undefined && matchedKey) {
                 await this.client.expire(matchedKey, this.defaultTtl);
             }
+            // Cost saved
+            let costSaved;
+            const costMicrosStr = winner.fields['cost_micros'];
+            if (costMicrosStr) {
+                const costMicros = parseInt(costMicrosStr, 10);
+                if (!isNaN(costMicros) && costMicros > 0) {
+                    costSaved = costMicros / 1_000_000;
+                    // Atomically increment cost_saved_micros in stats
+                    await this.client.hincrby(this.statsKey, 'cost_saved_micros', costMicros);
+                    this.telemetry.metrics.costSavedTotal
+                        .labels({ cache_name: this.name, category: categoryLabel }).inc(costSaved);
+                }
+            }
+            // Content blocks
+            let contentBlocks;
+            const contentBlocksStr = winner.fields['content_blocks'];
+            if (contentBlocksStr) {
+                try {
+                    contentBlocks = JSON.parse(contentBlocksStr);
+                }
+                catch { /* ignore parse errors */ }
+            }
             span.setAttributes({
-                'cache.hit': true, 'cache.similarity': score, 'cache.threshold': threshold,
+                'cache.hit': true, 'cache.similarity': winnerScore, 'cache.threshold': threshold,
                 'cache.confidence': confidence, 'cache.matched_key': matchedKey,
                 'cache.category': categoryLabel, ...timingAttrs,
             });
-            return {
-                hit: true, response: parsed[0].fields['response'],
-                similarity: score, confidence, matchedKey,
+            const result = {
+                hit: true, response: winner.fields['response'],
+                similarity: winnerScore, confidence, matchedKey,
             };
+            if (costSaved !== undefined)
+                result.costSaved = costSaved;
+            if (contentBlocks)
+                result.contentBlocks = contentBlocks;
+            return result;
         });
     }
     async store(prompt, response, options) {
         this.assertInitialized('store');
         return this.traced('store', async (span) => {
-            const { vector: embedding, durationSec: embedSec } = await this.embed(prompt);
+            const { text: promptText, binaryRefs } = await this.resolvePrompt(prompt);
+            const { vector: embedding, durationSec: embedSec } = await this.embed(promptText);
             this.assertDimension(embedding);
             const entryKey = `${this.entryPrefix}${(0, node_crypto_1.randomUUID)()}`;
             const category = options?.category ?? '';
             const model = options?.model ?? '';
+            // Compute cost if tokens and model provided
+            let costMicros;
+            if (options?.model &&
+                options?.inputTokens !== undefined &&
+                options?.outputTokens !== undefined &&
+                this.costTable) {
+                const pricing = this.costTable[options.model];
+                if (pricing) {
+                    costMicros = Math.round((options.inputTokens * pricing.inputPer1k / 1000 +
+                        options.outputTokens * pricing.outputPer1k / 1000) * 1_000_000);
+                }
+            }
+            const hashFields = {
+                prompt: promptText,
+                response,
+                model,
+                category,
+                inserted_at: Date.now().toString(),
+                metadata: JSON.stringify(options?.metadata ?? {}),
+                embedding: (0, utils_1.encodeFloat32)(embedding),
+            };
+            if (binaryRefs.length > 0) {
+                hashFields['binary_refs'] = binaryRefs.join(',');
+            }
+            if (costMicros !== undefined && costMicros > 0) {
+                hashFields['cost_micros'] = String(costMicros);
+            }
+            if (options?.temperature !== undefined) {
+                hashFields['temperature'] = String(options.temperature);
+            }
+            if (options?.topP !== undefined) {
+                hashFields['top_p'] = String(options.topP);
+            }
+            if (options?.seed !== undefined) {
+                hashFields['seed'] = String(options.seed);
+            }
             try {
-                await this.client.hset(entryKey, {
-                    prompt, response, model, category,
-                    inserted_at: Date.now().toString(),
-                    metadata: JSON.stringify(options?.metadata ?? {}),
-                    embedding: (0, utils_1.encodeFloat32)(embedding),
-                });
+                await this.client.hset(entryKey, hashFields);
             }
             catch (err) {
                 throw new errors_1.ValkeyCommandError('HSET', err);
@@ -202,11 +370,195 @@ class SemanticCache {
             return entryKey;
         });
     }
+    /**
+     * Store structured content blocks as the cached response.
+     * Populates both the response field (from TextBlock text) and content_blocks (full JSON).
+     */
+    async storeMultipart(prompt, blocks, options) {
+        this.assertInitialized('storeMultipart');
+        return this.traced('storeMultipart', async (span) => {
+            const { text: promptText, binaryRefs } = await this.resolvePrompt(prompt);
+            const { vector: embedding, durationSec: embedSec } = await this.embed(promptText);
+            this.assertDimension(embedding);
+            // Derive text response from blocks for backward compat
+            const textResponse = (0, utils_1.extractText)(blocks);
+            const entryKey = `${this.entryPrefix}${(0, node_crypto_1.randomUUID)()}`;
+            const category = options?.category ?? '';
+            const model = options?.model ?? '';
+            let costMicros;
+            if (options?.model && options?.inputTokens !== undefined && options?.outputTokens !== undefined && this.costTable) {
+                const pricing = this.costTable[options.model];
+                if (pricing) {
+                    costMicros = Math.round((options.inputTokens * pricing.inputPer1k / 1000 +
+                        options.outputTokens * pricing.outputPer1k / 1000) * 1_000_000);
+                }
+            }
+            const hashFields = {
+                prompt: promptText,
+                response: textResponse,
+                model,
+                category,
+                inserted_at: Date.now().toString(),
+                metadata: JSON.stringify(options?.metadata ?? {}),
+                embedding: (0, utils_1.encodeFloat32)(embedding),
+                content_blocks: JSON.stringify(blocks),
+            };
+            if (binaryRefs.length > 0) {
+                hashFields['binary_refs'] = binaryRefs.join(',');
+            }
+            if (costMicros !== undefined && costMicros > 0) {
+                hashFields['cost_micros'] = String(costMicros);
+            }
+            if (options?.temperature !== undefined)
+                hashFields['temperature'] = String(options.temperature);
+            if (options?.topP !== undefined)
+                hashFields['top_p'] = String(options.topP);
+            if (options?.seed !== undefined)
+                hashFields['seed'] = String(options.seed);
+            try {
+                await this.client.hset(entryKey, hashFields);
+            }
+            catch (err) {
+                throw new errors_1.ValkeyCommandError('HSET', err);
+            }
+            const ttl = options?.ttl ?? this.defaultTtl;
+            if (ttl !== undefined)
+                await this.client.expire(entryKey, ttl);
+            span.setAttributes({
+                'cache.name': this.name, 'cache.key': entryKey, 'cache.ttl': ttl ?? -1,
+                'cache.category': category || 'none', 'cache.model': model || 'none',
+                'embedding_latency_ms': embedSec * 1000,
+            });
+            return entryKey;
+        });
+    }
+    /**
+     * Check multiple prompts in parallel, using pipelined FT.SEARCH calls.
+     * Returns results in input order.
+     */
+    async checkBatch(prompts, options) {
+        this.assertInitialized('checkBatch');
+        if (prompts.length === 0)
+            return [];
+        if (options?.rerank) {
+            throw new errors_1.SemanticCacheUsageError("checkBatch() does not support the 'rerank' option. Use check() for reranking individual prompts.");
+        }
+        if (options?.staleAfterModelChange) {
+            throw new errors_1.SemanticCacheUsageError("checkBatch() does not support 'staleAfterModelChange'. Use check() for stale-model eviction.");
+        }
+        return this.traced('checkBatch', async (span) => {
+            // Resolve all prompts and embed in parallel
+            const resolved = await Promise.all(prompts.map((p) => this.resolvePrompt(p)));
+            const embeddings = await Promise.all(resolved.map(({ text }) => this.embed(text)));
+            const category = options?.category ?? '';
+            const threshold = options?.threshold ??
+                (category && this.categoryThresholds[category] !== undefined
+                    ? this.categoryThresholds[category]
+                    : this.defaultThreshold);
+            const k = options?.k ?? 1;
+            const userFilter = options?.filter;
+            // Pipeline all FT.SEARCH calls
+            const pipeline = this.client.pipeline();
+            for (let i = 0; i < prompts.length; i++) {
+                const { binaryRefs } = resolved[i];
+                const { vector: embedding } = embeddings[i];
+                const binaryFilter = binaryRefs.length > 0 && this._hasBinaryRefs
+                    ? (binaryRefs.length === 1
+                        ? `@binary_refs:{${(0, utils_1.escapeTag)(binaryRefs[0])}}`
+                        : binaryRefs.map((r) => `@binary_refs:{${(0, utils_1.escapeTag)(r)}}`).join(' '))
+                    : null;
+                const combinedFilter = [userFilter, binaryFilter].filter(Boolean).join(' ');
+                const filterExpr = combinedFilter ? `(${combinedFilter})` : '*';
+                const query = `${filterExpr}=>[KNN ${k} @embedding $vec AS __score]`;
+                pipeline.call('FT.SEARCH', this.indexName, query, 'PARAMS', '2', 'vec', (0, utils_1.encodeFloat32)(embedding), 'LIMIT', '0', String(k), 'DIALECT', '2');
+            }
+            const pipelineResults = await pipeline.exec();
+            span.setAttributes({ 'cache.batch_size': prompts.length, 'cache.name': this.name });
+            const results = [];
+            const categoryLabel = category || 'none';
+            for (let i = 0; i < prompts.length; i++) {
+                const pipelineEntry = pipelineResults?.[i];
+                const err = pipelineEntry?.[0];
+                const rawResult = pipelineEntry?.[1];
+                if (err) {
+                    await this.recordStat('misses');
+                    this.telemetry.metrics.requestsTotal
+                        .labels({ cache_name: this.name, result: 'miss', category: categoryLabel }).inc();
+                    results.push({ hit: false, confidence: 'miss' });
+                    continue;
+                }
+                const parsed = (0, utils_1.parseFtSearchResponse)(rawResult);
+                if (parsed.length === 0) {
+                    await this.recordStat('misses');
+                    this.telemetry.metrics.requestsTotal
+                        .labels({ cache_name: this.name, result: 'miss', category: categoryLabel }).inc();
+                    results.push({ hit: false, confidence: 'miss' });
+                    continue;
+                }
+                const scoreStr = parsed[0].fields['__score'];
+                const score = scoreStr !== undefined ? parseFloat(scoreStr) : NaN;
+                if (isNaN(score) || score > threshold) {
+                    if (!isNaN(score)) {
+                        await this.recordSimilarityWindow(score, 'miss', category);
+                    }
+                    await this.recordStat('misses');
+                    this.telemetry.metrics.requestsTotal
+                        .labels({ cache_name: this.name, result: 'miss', category: categoryLabel }).inc();
+                    const result = { hit: false, confidence: 'miss' };
+                    if (!isNaN(score)) {
+                        result.similarity = score;
+                        result.nearestMiss = { similarity: score, deltaToThreshold: score - threshold };
+                    }
+                    results.push(result);
+                    continue;
+                }
+                await this.recordSimilarityWindow(score, 'hit', category);
+                const confidence = score >= threshold - this.uncertaintyBand ? 'uncertain' : 'high';
+                await this.recordStat('hits');
+                const metricResult = confidence === 'uncertain' ? 'uncertain_hit' : 'hit';
+                this.telemetry.metrics.requestsTotal
+                    .labels({ cache_name: this.name, result: metricResult, category: categoryLabel }).inc();
+                const matchedKey = parsed[0].key;
+                if (this.defaultTtl !== undefined && matchedKey) {
+                    await this.client.expire(matchedKey, this.defaultTtl);
+                }
+                let costSaved;
+                const costMicrosStr = parsed[0].fields['cost_micros'];
+                if (costMicrosStr) {
+                    const costMicros = parseInt(costMicrosStr, 10);
+                    if (!isNaN(costMicros) && costMicros > 0) {
+                        costSaved = costMicros / 1_000_000;
+                        await this.client.hincrby(this.statsKey, 'cost_saved_micros', costMicros);
+                        this.telemetry.metrics.costSavedTotal
+                            .labels({ cache_name: this.name, category: categoryLabel }).inc(costSaved);
+                    }
+                }
+                let contentBlocks;
+                const contentBlocksStr = parsed[0].fields['content_blocks'];
+                if (contentBlocksStr) {
+                    try {
+                        contentBlocks = JSON.parse(contentBlocksStr);
+                    }
+                    catch { /* ignore */ }
+                }
+                const result = {
+                    hit: true, response: parsed[0].fields['response'],
+                    similarity: score, confidence, matchedKey,
+                };
+                if (costSaved !== undefined)
+                    result.costSaved = costSaved;
+                if (contentBlocks)
+                    result.contentBlocks = contentBlocks;
+                results.push(result);
+            }
+            return results;
+        });
+    }
     /**
      * Deletes all entries matching a valkey-search filter expression.
      *
      * **Security note:** `filter` is passed directly to FT.SEARCH. Only pass
-     * trusted, programmatically-constructed expressions — never unsanitised
+     * trusted, programmatically-constructed expressions - never unsanitised
      * user input.
      */
     async invalidate(filter) {
@@ -242,13 +594,34 @@ class SemanticCache {
             return { deleted: keys.length, truncated };
         });
     }
+    /** Delete all entries tagged with the given model name. */
+    async invalidateByModel(model) {
+        let total = 0;
+        let result;
+        do {
+            result = await this.invalidate(`@model:{${(0, utils_1.escapeTag)(model)}}`);
+            total += result.deleted;
+        } while (result.truncated);
+        return total;
+    }
+    /** Delete all entries tagged with the given category. */
+    async invalidateByCategory(category) {
+        let total = 0;
+        let result;
+        do {
+            result = await this.invalidate(`@category:{${(0, utils_1.escapeTag)(category)}}`);
+            total += result.deleted;
+        } while (result.truncated);
+        return total;
+    }
     async stats() {
         this.assertInitialized('stats');
         const raw = await this.client.hgetall(this.statsKey);
-        const hits = parseInt(raw.hits ?? '0', 10);
-        const misses = parseInt(raw.misses ?? '0', 10);
-        const total = parseInt(raw.total ?? '0', 10);
-        return { hits, misses, total, hitRate: total === 0 ? 0 : hits / total };
+        const hits = parseInt(raw?.hits ?? '0', 10);
+        const misses = parseInt(raw?.misses ?? '0', 10);
+        const total = parseInt(raw?.total ?? '0', 10);
+        const costSavedMicros = parseInt(raw?.cost_saved_micros ?? '0', 10);
+        return { hits, misses, total, hitRate: total === 0 ? 0 : hits / total, costSavedMicros };
     }
     async indexInfo() {
         this.assertInitialized('indexInfo');
@@ -271,27 +644,217 @@ class SemanticCache {
         }
         return { name: this.indexName, numDocs, dimension: this._dimension, indexingState };
     }
-    // ── Private helpers ────────────────────────────────────────
+    /**
+     * Analyze the rolling similarity score window and recommend threshold adjustments.
+     */
+    async thresholdEffectiveness(options) {
+        this.assertInitialized('thresholdEffectiveness');
+        const minSamples = options?.minSamples ?? 100;
+        const category = options?.category;
+        const threshold = category && this.categoryThresholds[category] !== undefined
+            ? this.categoryThresholds[category]
+            : this.defaultThreshold;
+        // Read all window entries
+        let rawEntries;
+        try {
+            rawEntries = (await this.client.zrange(this.similarityWindowKey, '0', '-1'));
+        }
+        catch {
+            rawEntries = [];
+        }
+        // Parse and optionally filter by category
+        const entries = [];
+        for (const raw of rawEntries) {
+            try {
+                const entry = JSON.parse(String(raw));
+                if (typeof entry.score === 'number' &&
+                    (entry.result === 'hit' || entry.result === 'miss')) {
+                    if (!category || entry.category === category) {
+                        entries.push(entry);
+                    }
+                }
+            }
+            catch { /* skip corrupt entries */ }
+        }
+        const sampleCount = entries.length;
+        const categoryLabel = category ?? 'all';
+        if (sampleCount < minSamples) {
+            return {
+                category: categoryLabel,
+                sampleCount,
+                currentThreshold: threshold,
+                hitRate: 0,
+                uncertainHitRate: 0,
+                nearMissRate: 0,
+                avgHitSimilarity: 0,
+                avgMissSimilarity: 0,
+                recommendation: 'insufficient_data',
+                reasoning: `Only ${sampleCount} samples collected; ${minSamples} required for a reliable recommendation.`,
+            };
+        }
+        const hits = entries.filter((e) => e.result === 'hit');
+        const misses = entries.filter((e) => e.result === 'miss');
+        const hitRate = hits.length / sampleCount;
+        const uncertainHits = hits.filter((e) => e.score >= threshold - this.uncertaintyBand);
+        const uncertainHitRate = hits.length > 0 ? uncertainHits.length / hits.length : 0;
+        // Near-misses are scores just ABOVE the threshold (genuine close misses).
+        // Scores below the threshold recorded as misses (rerank rejection, stale eviction)
+        // must be excluded — they produce negative avgNearMissDelta, causing
+        // recommendedThreshold = threshold + negative < threshold, contradicting "loosen".
+        const nearMisses = misses.filter((e) => e.score > threshold && e.score <= threshold + 0.03);
+        const nearMissRate = misses.length > 0 ? nearMisses.length / misses.length : 0;
+        const avgHitSimilarity = hits.length > 0 ? hits.reduce((s, e) => s + e.score, 0) / hits.length : 0;
+        const avgMissSimilarity = misses.length > 0 ? misses.reduce((s, e) => s + e.score, 0) / misses.length : 0;
+        // avgNearMissDelta: how far above the threshold near-misses are on average
+        const avgNearMissDelta = nearMisses.length > 0
+            ? nearMisses.reduce((s, e) => s + (e.score - threshold), 0) / nearMisses.length
+            : 0;
+        let recommendation;
+        let recommendedThreshold;
+        let reasoning;
+        if (uncertainHitRate > 0.2) {
+            recommendation = 'tighten_threshold';
+            recommendedThreshold = Math.max(0, threshold - this.uncertaintyBand * 1.5);
+            reasoning = `${(uncertainHitRate * 100).toFixed(1)}% of hits are in the uncertainty band - tighten the threshold to reduce false positives.`;
+        }
+        else if (nearMissRate > 0.3 && avgNearMissDelta < 0.03) {
+            recommendation = 'loosen_threshold';
+            recommendedThreshold = threshold + avgNearMissDelta;
+            reasoning = `${(nearMissRate * 100).toFixed(1)}% of misses are very close to the threshold - consider loosening to capture more hits.`;
+        }
+        else {
+            recommendation = 'optimal';
+            reasoning = `Hit rate is ${(hitRate * 100).toFixed(1)}% with ${(uncertainHitRate * 100).toFixed(1)}% uncertain hits - threshold appears well-calibrated.`;
+        }
+        return {
+            category: categoryLabel,
+            sampleCount,
+            currentThreshold: threshold,
+            hitRate,
+            uncertainHitRate,
+            nearMissRate,
+            avgHitSimilarity,
+            avgMissSimilarity,
+            recommendation,
+            recommendedThreshold,
+            reasoning,
+        };
+    }
+    /**
+     * Returns threshold effectiveness results for every category seen in the
+     * rolling window, plus one aggregate result for all categories combined.
+     */
+    async thresholdEffectivenessAll(options) {
+        this.assertInitialized('thresholdEffectivenessAll');
+        let rawEntries;
+        try {
+            rawEntries = (await this.client.zrange(this.similarityWindowKey, '0', '-1'));
+        }
+        catch {
+            rawEntries = [];
+        }
+        // Collect unique categories
+        const categories = new Set();
+        for (const raw of rawEntries) {
+            try {
+                const entry = JSON.parse(raw);
+                if (entry.category)
+                    categories.add(entry.category);
+            }
+            catch { /* skip */ }
+        }
+        const results = await Promise.all([
+            this.thresholdEffectiveness({ minSamples: options?.minSamples }),
+            ...[...categories].filter(Boolean).map((cat) => this.thresholdEffectiveness({ category: cat, minSamples: options?.minSamples })),
+        ]);
+        return results;
+    }
+    // -- Internal helpers exposed to package adapters --
+    /** @internal Default similarity threshold. */
+    get _defaultThreshold() { return this.defaultThreshold; }
+    /**
+     * Execute a stable FT.SEARCH for use by adapters (e.g. LangGraph).
+     * SORTBY inserted_at ASC gives stable ordering across paginated calls.
+     * @internal
+     */
+    async _searchEntries(filterExpr, limit, offset) {
+        return this.client.call('FT.SEARCH', this.indexName, filterExpr, 'SORTBY', 'inserted_at', 'ASC', 'LIMIT', String(offset), String(limit), 'DIALECT', '2');
+    }
+    /**
+     * Embed text for use by adapters (e.g. LangGraph semantic search).
+     * @internal
+     */
+    async _embedText(text) {
+        return this.embed(text);
+    }
+    // -- Private helpers --
     async _doInitialize() {
         const gen = this._initGeneration;
         return this.traced('initialize', async () => {
-            const dim = await this.ensureIndexAndGetDimension();
-            // If flush() ran while we were initializing, don't overwrite its state.
+            const { dim, hasBinaryRefs } = await this.ensureIndexAndGetDimension();
             if (this._initGeneration !== gen)
                 return;
             this._dimension = dim;
+            this._hasBinaryRefs = hasBinaryRefs;
             this._initialized = true;
+            // Fire analytics init once (not on every flush+initialize cycle)
+            this.initAnalyticsSafe().catch(() => { });
         });
     }
+    async initAnalyticsSafe() {
+        if (this.analyticsInitiated)
+            return;
+        this.analyticsInitiated = true;
+        try {
+            const a = await (0, analytics_1.createAnalytics)(this.analyticsOpts);
+            if (this.shutdownCalled) {
+                await a.shutdown();
+                return;
+            }
+            this.analytics = a;
+            await a.init(this.client, this.name, {
+                defaultThreshold: this.defaultThreshold,
+                uncertaintyBand: this.uncertaintyBand,
+                defaultTtl: this.defaultTtl ?? null,
+                hasCostTable: !!this.costTable,
+                usesDefaultCostTable: this.usesDefaultCostTable,
+                embeddingCacheEnabled: this.embeddingCacheEnabled,
+                categoryThresholdCount: Object.keys(this.categoryThresholds).length,
+                dimension: this._dimension,
+            });
+            const intervalMs = this.analyticsOpts?.statsIntervalMs ?? 300_000;
+            if (!this.shutdownCalled && intervalMs > 0) {
+                this.statsTimer = setInterval(() => this.captureStatsSnapshot(), intervalMs);
+                this.statsTimer.unref();
+            }
+        }
+        catch {
+            // never throw from analytics
+        }
+    }
+    captureStatsSnapshot() {
+        this.stats()
+            .then((s) => {
+            this.analytics.capture('stats_snapshot', {
+                hits: s.hits,
+                misses: s.misses,
+                hit_rate: s.hitRate,
+                cost_saved_micros: s.costSavedMicros,
+            });
+        })
+            .catch(() => { });
+    }
     async ensureIndexAndGetDimension() {
         // Try reading an existing index
         try {
             const info = (await this.client.call('FT.INFO', this.indexName));
             const dim = this.parseDimensionFromInfo(info);
+            const hasBinaryRefs = this.parseHasBinaryRefsFromInfo(info);
             if (dim > 0)
-                return dim;
-            // Couldn't parse dimension from FT.INFO — fall back to probe
-            return (await this.embed('probe')).vector.length;
+                return { dim, hasBinaryRefs };
+            // Couldn't parse dimension from FT.INFO - fall back to probe
+            const probeDim = (await this.embed('probe')).vector.length;
+            return { dim: probeDim, hasBinaryRefs };
         }
         catch (err) {
             if (err instanceof errors_1.EmbeddingError)
@@ -300,18 +863,69 @@ class SemanticCache {
                 throw new errors_1.ValkeyCommandError('FT.INFO', err);
             }
         }
-        // Index doesn't exist — probe dimension and create it
+        // Index doesn't exist - probe dimension and create it
         const dim = (await this.embed('probe')).vector.length;
         try {
-            await this.client.call('FT.CREATE', this.indexName, 'ON', 'HASH', 'PREFIX', '1', this.entryPrefix, 'SCHEMA', 'prompt', 'TEXT', 'NOSTEM', 'response', 'TEXT', 'NOSTEM', 'model', 'TAG', 'category', 'TAG', 'inserted_at', 'NUMERIC', 'SORTABLE', 'embedding', 'VECTOR', 'HNSW', '6', 'TYPE', 'FLOAT32', 'DIM', String(dim), 'DISTANCE_METRIC', 'COSINE');
+            await this.client.call('FT.CREATE', this.indexName, 'ON', 'HASH', 'PREFIX', '1', this.entryPrefix, 'SCHEMA', 'prompt', 'TEXT', 'NOSTEM', 'response', 'TEXT', 'NOSTEM', 'model', 'TAG', 'category', 'TAG', 'binary_refs', 'TAG', 'inserted_at', 'NUMERIC', 'SORTABLE', 'temperature', 'NUMERIC', 'top_p', 'NUMERIC', 'seed', 'NUMERIC', 'embedding', 'VECTOR', 'HNSW', '6', 'TYPE', 'FLOAT32', 'DIM', String(dim), 'DISTANCE_METRIC', 'COSINE');
         }
         catch (err) {
             throw new errors_1.ValkeyCommandError('FT.CREATE', err);
         }
-        return dim;
+        return { dim, hasBinaryRefs: true };
+    }
+    /** Check if the index schema has a binary_refs field. */
+    parseHasBinaryRefsFromInfo(info) {
+        for (let i = 0; i < info.length - 1; i += 2) {
+            const key = String(info[i]);
+            if (key !== 'attributes' && key !== 'fields')
+                continue;
+            const attributes = info[i + 1];
+            if (!Array.isArray(attributes))
+                continue;
+            for (const attr of attributes) {
+                if (!Array.isArray(attr))
+                    continue;
+                for (let j = 0; j < attr.length - 1; j++) {
+                    if (String(attr[j]) === 'identifier' && String(attr[j + 1]) === 'binary_refs') {
+                        return true;
+                    }
+                }
+            }
+        }
+        return false;
     }
-    /** Wraps embedFn with error handling and duration tracking. */
+    /** Resolve a prompt (string or ContentBlock[]) into text + binary refs. */
+    resolvePrompt(prompt) {
+        if (typeof prompt === 'string') {
+            return { text: prompt, binaryRefs: [] };
+        }
+        const text = (0, utils_1.extractText)(prompt);
+        const binaryRefs = (0, utils_1.extractBinaryRefs)(prompt);
+        return { text, binaryRefs };
+    }
+    /** Wraps embedFn with error handling, duration tracking, and optional embedding cache. */
     async embed(text) {
+        // Check embedding cache
+        if (this.embeddingCacheEnabled && text) {
+            const hash = (0, node_crypto_2.createHash)('sha256').update(text).digest('hex');
+            const embedKey = `${this.embedKeyPrefix}${hash}`;
+            try {
+                const cached = await this.client.getBuffer(embedKey);
+                if (cached) {
+                    this.telemetry.metrics.embeddingCacheTotal
+                        .labels({ cache_name: this.name, result: 'hit' }).inc();
+                    // Decode Float32 buffer
+                    const vector = [];
+                    for (let i = 0; i < cached.length; i += 4) {
+                        vector.push(cached.readFloatLE(i));
+                    }
+                    return { vector, durationSec: 0 };
+                }
+            }
+            catch { /* ignore cache read errors */ }
+            this.telemetry.metrics.embeddingCacheTotal
+                .labels({ cache_name: this.name, result: 'miss' }).inc();
+        }
         const start = performance.now();
         let vector;
         try {
@@ -324,12 +938,22 @@ class SemanticCache {
         this.telemetry.metrics.embeddingDuration
             .labels({ cache_name: this.name })
             .observe(durationSec);
+        // Store in embedding cache
+        if (this.embeddingCacheEnabled && text) {
+            const hash = (0, node_crypto_2.createHash)('sha256').update(text).digest('hex');
+            const embedKey = `${this.embedKeyPrefix}${hash}`;
+            try {
+                const buf = (0, utils_1.encodeFloat32)(vector);
+                await this.client.set(embedKey, buf, 'EX', this.embeddingCacheTtl);
+            }
+            catch { /* ignore cache write errors */ }
+        }
         return { vector, durationSec };
     }
     /**
      * Wraps a method body in an OTel span with automatic status, end, and
      * operation duration metric. The span is passed to fn so callers can
-     * set attributes — but callers must NOT call span.end() or span.setStatus(),
+     * set attributes - but callers must NOT call span.end() or span.setStatus(),
      * as traced() handles both.
      */
     async traced(operation, fn) {
@@ -359,6 +983,24 @@ class SemanticCache {
         pipeline.hincrby(this.statsKey, field, 1);
         await pipeline.exec();
     }
+    /** Append to the rolling similarity window sorted set and trim to 10,000 entries or 7 days. */
+    async recordSimilarityWindow(score, result, category) {
+        const now = Date.now();
+        // Include a unique nonce so identical (score, result, category) tuples are
+        // each recorded as distinct ZADD members instead of overwriting each other.
+        const member = JSON.stringify({ score, result, category, _n: Math.random() });
+        const sevenDaysAgo = now - 7 * 24 * 60 * 60 * 1000;
+        try {
+            const pipeline = this.client.pipeline();
+            pipeline.zadd(this.similarityWindowKey, now, member);
+            // Trim by time: remove entries older than 7 days
+            pipeline.zremrangebyscore(this.similarityWindowKey, '-inf', sevenDaysAgo);
+            // Trim by count: keep at most 10,000 most recent
+            pipeline.zremrangebyrank(this.similarityWindowKey, 0, -10001);
+            await pipeline.exec();
+        }
+        catch { /* best effort - never fail on window writes */ }
+    }
     assertInitialized(method) {
         if (!this._initialized) {
             throw new errors_1.SemanticCacheUsageError(`SemanticCache.initialize() must be called before ${method}().`);