@usewhisper/mcp-server 0.2.3 → 0.4.0

package/dist/chunk-TWEIYHI6.js

@@ -0,0 +1,399 @@
+import {
+  calculateTemporalRelevance,
+  parseTemporalQuery
+} from "./chunk-5KBZQHDL.js";
+import {
+  db,
+  embedSingle
+} from "./chunk-MEFLJ4PV.js";
+
+// ../src/engine/cache.ts
+import crypto from "crypto";
+var semanticCache = /* @__PURE__ */ new Map();
+var MAX_SEMANTIC_CACHE_SIZE = 500;
+var SEMANTIC_THRESHOLD = 0.92;
+var DEFAULT_CONFIG = {
+  ttl: 3600,
+  // 1 hour
+  enabled: true,
+  keyPrefix: "whisper:context:"
+};
+var cacheHits = 0;
+var cacheMisses = 0;
+function recordCacheHit() {
+  cacheHits++;
+}
+function recordCacheMiss() {
+  cacheMisses++;
+}
+function cosineSimilarity(a, b) {
+  if (a.length !== b.length) return 0;
+  let dotProduct = 0;
+  let normA = 0;
+  let normB = 0;
+  for (let i = 0; i < a.length; i++) {
+    dotProduct += a[i] * b[i];
+    normA += a[i] * a[i];
+    normB += b[i] * b[i];
+  }
+  return dotProduct / (Math.sqrt(normA) * Math.sqrt(normB));
+}
+async function getFromSemanticCache(queryEmbedding) {
+  if (!DEFAULT_CONFIG.enabled) return null;
+  const now = Date.now();
+  let bestMatch = null;
+  let bestSimilarity = 0;
+  for (const [key, entry] of semanticCache.entries()) {
+    if (entry.expiry < now) {
+      semanticCache.delete(key);
+      continue;
+    }
+    const similarity = cosineSimilarity(queryEmbedding, entry.embedding);
+    if (similarity > bestSimilarity) {
+      bestSimilarity = similarity;
+      bestMatch = { key, ...entry, similarity };
+    }
+  }
+  if (bestMatch && bestMatch.similarity >= SEMANTIC_THRESHOLD) {
+    recordCacheHit();
+    return { results: bestMatch.results, similarity: bestMatch.similarity };
+  }
+  recordCacheMiss();
+  return null;
+}
+async function setInSemanticCache(queryEmbedding, results) {
+  if (!DEFAULT_CONFIG.enabled) return;
+  const now = Date.now();
+  const key = `sem:${crypto.createHash("md5").update(JSON.stringify(queryEmbedding.slice(0, 10))).digest("hex").substring(0, 8)}`;
+  semanticCache.set(key, {
+    embedding: queryEmbedding,
+    results,
+    expiry: now + DEFAULT_CONFIG.ttl * 1e3
+  });
+  if (semanticCache.size > MAX_SEMANTIC_CACHE_SIZE) {
+    let oldestKey = null;
+    let oldestExpiry = Infinity;
+    for (const [k, v] of semanticCache.entries()) {
+      if (v.expiry < oldestExpiry) {
+        oldestExpiry = v.expiry;
+        oldestKey = k;
+      }
+    }
+    if (oldestKey) {
+      semanticCache.delete(oldestKey);
+    }
+  }
+}
+
+// ../src/engine/memory/search.ts
+var EARLY_EXIT_SIMILARITY = 0.92;
+async function searchMemories(params) {
+  const {
+    query,
+    questionDate,
+    userId,
+    projectId,
+    orgId,
+    sessionId,
+    topK = 10,
+    includeInactive = false,
+    memoryTypes
+  } = params;
+  const queryEmbedding = await embedSingle(query);
+  const cached = await getFromSemanticCache(queryEmbedding);
+  if (cached && cached.similarity >= EARLY_EXIT_SIMILARITY) {
+    console.log(`\u26A1 Semantic cache hit (similarity: ${cached.similarity.toFixed(3)})`);
+    return cached.results.slice(0, topK);
+  }
+  const temporal = params.temporalFilter || await parseTemporalQuery(query, questionDate);
+  const semanticResults = await vectorSearchMemories({
+    embedding: queryEmbedding,
+    userId,
+    projectId,
+    orgId,
+    sessionId,
+    temporal,
+    includeInactive,
+    memoryTypes,
+    limit: topK * 3
+    // Get more for reranking
+  });
+  if (semanticResults.length === 0) {
+    return [];
+  }
+  if (semanticResults.length > 0 && semanticResults[0].similarity >= EARLY_EXIT_SIMILARITY) {
+    console.log(`\u26A1 Early exit at ${semanticResults[0].similarity.toFixed(3)}`);
+    const topMemories2 = semanticResults.slice(0, topK);
+    await setInSemanticCache(queryEmbedding, topMemories2);
+    return topMemories2;
+  }
+  const enriched = await enrichWithRelations(semanticResults, topK * 2);
+  const scored = enriched.map((memory) => ({
+    ...memory,
+    temporalScore: memory.documentDate ? calculateTemporalRelevance(memory.documentDate, questionDate) : 0.5
+  }));
+  const combined = scored.map((m) => ({
+    ...m,
+    finalScore: m.similarity * 0.7 + m.temporalScore * 0.3
+  }));
+  combined.sort((a, b) => b.finalScore - a.finalScore);
+  const topMemories = combined.slice(0, topK);
+  const results = await injectSourceChunks(topMemories);
+  await setInSemanticCache(queryEmbedding, results);
+  return results;
+}
+async function vectorSearchMemories(params) {
+  const {
+    embedding,
+    userId,
+    projectId,
+    orgId,
+    sessionId,
+    temporal,
+    includeInactive,
+    memoryTypes,
+    limit
+  } = params;
+  const whereConditions = [
+    { projectId }
+  ];
+  if (orgId) {
+    whereConditions.push({ orgId });
+  }
+  if (userId) {
+    whereConditions.push({ userId });
+  }
+  if (sessionId) {
+    whereConditions.push({ sessionId });
+  }
+  if (!includeInactive) {
+    whereConditions.push({ isActive: true });
+  }
+  whereConditions.push({
+    OR: [
+      { validUntil: null },
+      { validUntil: { gt: /* @__PURE__ */ new Date() } }
+    ]
+  });
+  if (memoryTypes && memoryTypes.length > 0) {
+    whereConditions.push({
+      memoryType: { in: memoryTypes }
+    });
+  }
+  if (temporal.hasTemporalConstraint && temporal.dateRange) {
+    whereConditions.push({
+      documentDate: {
+        gte: temporal.dateRange.start,
+        lte: temporal.dateRange.end
+      }
+    });
+  }
+  const embeddingStr = `[${embedding.join(",")}]`;
+  let whereClause = `"projectId" = '${projectId}'`;
+  if (orgId) whereClause += ` AND "orgId" = '${orgId}'`;
+  if (userId) whereClause += ` AND "userId" = '${userId}'`;
+  if (sessionId) whereClause += ` AND "sessionId" = '${sessionId}'`;
+  if (!includeInactive) whereClause += ` AND "isActive" = true`;
+  whereClause += ` AND ("validUntil" IS NULL OR "validUntil" > NOW())`;
+  if (memoryTypes && memoryTypes.length > 0) {
+    const typesStr = memoryTypes.map((t) => `'${t.replace(/'/g, "''")}'`).join(",");
+    whereClause += ` AND "memoryType" IN (${typesStr})`;
+  }
+  if (temporal.hasTemporalConstraint && temporal.dateRange) {
+    whereClause += ` AND "documentDate" >= '${temporal.dateRange.start.toISOString()}' AND "documentDate" <= '${temporal.dateRange.end.toISOString()}'`;
+  }
+  const results = await db.$queryRawUnsafe(`
+    SELECT
+      id,
+      content,
+      "memoryType" as "memoryType",
+      confidence,
+      version,
+      "documentDate" as "documentDate",
+      "eventDate" as "eventDate",
+      "validFrom" as "validFrom",
+      "validUntil" as "validUntil",
+      "sourceChunkId" as "sourceChunkId",
+      metadata,
+      1 - (embedding <=> '${embeddingStr}'::vector) as similarity
+    FROM memories
+    WHERE ${whereClause}
+    ORDER BY embedding <=> '${embeddingStr}'::vector
+    LIMIT ${limit}
+  `);
+  return results;
+}
+async function enrichWithRelations(memories, maxTotal) {
+  if (memories.length === 0) {
+    return [];
+  }
+  const memoryIds = memories.map((m) => m.id);
+  const memoryIdsList = memoryIds.map((id) => `'${id}'`).join(",");
+  const relationsQuery = await db.$queryRawUnsafe(`
+    SELECT 
+      r.id as relation_id,
+      r."fromMemoryId",
+      r."toMemoryId",
+      r."relationType",
+      m.id,
+      m.content,
+      m."memoryType",
+      m.confidence,
+      m.version,
+      m."documentDate",
+      m."eventDate",
+      m."validFrom",
+      m."validUntil",
+      m."sourceChunkId",
+      m.metadata
+    FROM "memory_relations" r
+    LEFT JOIN memories m ON m.id = r."toMemoryId"
+    WHERE r."fromMemoryId" IN (${memoryIdsList})
+      AND m."isActive" = true
+    LIMIT 100
+  `);
+  if (!relationsQuery || relationsQuery.length === 0) {
+    return memories;
+  }
+  const relatedIds = /* @__PURE__ */ new Set();
+  const relationMap = /* @__PURE__ */ new Map();
+  for (const row of relationsQuery) {
+    if (row.toMemoryId && !memoryIds.includes(row.toMemoryId)) {
+      relatedIds.add(row.toMemoryId);
+      if (!relationMap.has(row.fromMemoryId)) {
+        relationMap.set(row.fromMemoryId, []);
+      }
+      relationMap.get(row.fromMemoryId)?.push({
+        memoryId: row.toMemoryId,
+        relationType: row.relationType,
+        content: row.content
+      });
+    }
+  }
+  if (relatedIds.size === 0) {
+    return memories;
+  }
+  const relatedIdsList = Array.from(relatedIds).map((id) => `'${id}'`).join(",");
+  const relatedMemories = await db.$queryRawUnsafe(`
+    SELECT 
+      id,
+      content,
+      "memoryType" as "memoryType",
+      confidence,
+      version,
+      "documentDate" as "documentDate",
+      "eventDate" as "eventDate",
+      "validFrom" as "validFrom",
+      "validUntil" as "validUntil",
+      "sourceChunkId" as "sourceChunkId",
+      metadata
+    FROM memories
+    WHERE id IN (${relatedIdsList})
+      AND "isActive" = true
+    LIMIT ${maxTotal - memories.length}
+  `);
+  const relatedWithScores = relatedMemories.map((m) => ({
+    ...m,
+    similarity: 0.6,
+    isRelated: true,
+    relations: relationMap.get(m.id) || []
+  }));
+  return [...memories, ...relatedWithScores];
+}
+async function injectSourceChunks(memories) {
+  const chunkIds = memories.map((m) => m.sourceChunkId).filter((id) => id !== null);
+  if (chunkIds.length === 0) {
+    return memories.map((m) => ({
+      memory: {
+        id: m.id,
+        content: m.content,
+        memoryType: m.memoryType,
+        entityMentions: m.entityMentions || [],
+        confidence: m.confidence,
+        version: m.version,
+        temporal: {
+          documentDate: m.documentDate,
+          eventDate: m.eventDate,
+          validFrom: m.validFrom,
+          validUntil: m.validUntil
+        }
+      },
+      similarity: m.similarity
+    }));
+  }
+  const chunks = await db.chunk.findMany({
+    where: {
+      id: { in: chunkIds }
+    },
+    select: {
+      id: true,
+      content: true,
+      metadata: true
+    }
+  });
+  const chunkMap = new Map(chunks.map((c) => [c.id, c]));
+  return memories.map((m) => ({
+    memory: {
+      id: m.id,
+      content: m.content,
+      memoryType: m.memoryType,
+      entityMentions: m.entityMentions || [],
+      confidence: m.confidence,
+      version: m.version,
+      temporal: {
+        documentDate: m.documentDate,
+        eventDate: m.eventDate,
+        validFrom: m.validFrom,
+        validUntil: m.validUntil
+      }
+    },
+    chunk: m.sourceChunkId && chunkMap.has(m.sourceChunkId) ? {
+      id: chunkMap.get(m.sourceChunkId).id,
+      content: chunkMap.get(m.sourceChunkId).content,
+      metadata: chunkMap.get(m.sourceChunkId).metadata
+    } : void 0,
+    similarity: m.similarity,
+    relations: m.isRelated ? [] : void 0
+  }));
+}
+async function getSessionMemories(params) {
+  const { sessionId, projectId, limit = 50, sinceDate } = params;
+  const where = {
+    sessionId,
+    projectId,
+    isActive: true
+  };
+  if (sinceDate) {
+    where.createdAt = { gte: sinceDate };
+  }
+  return db.memory.findMany({
+    where,
+    orderBy: {
+      createdAt: "desc"
+    },
+    take: limit
+  });
+}
+async function getUserProfile(params) {
+  const { userId, projectId, memoryTypes, limit = 50 } = params;
+  const where = {
+    userId,
+    projectId,
+    isActive: true,
+    scope: "USER"
+  };
+  if (memoryTypes) {
+    where.memoryType = { in: memoryTypes };
+  }
+  return db.memory.findMany({
+    where,
+    orderBy: { importance: "desc" },
+    take: Math.min(limit, 100)
+  });
+}
+
+export {
+  searchMemories,
+  getSessionMemories,
+  getUserProfile
+};

package/dist/consolidation-2GCKI4RE.js

@@ -0,0 +1,220 @@
+import {
+  db,
+  embedSingle
+} from "./chunk-MEFLJ4PV.js";
+import "./chunk-QGM4M3NI.js";
+
+// ../src/engine/memory/consolidation.ts
+import OpenAI from "openai";
+var openai = new OpenAI({
+  apiKey: process.env.OPENAI_API_KEY || ""
+});
+async function findDuplicateMemories(params) {
+  const {
+    projectId,
+    userId,
+    similarityThreshold = 0.95,
+    limit = 50
+  } = params;
+  const maxMemories = Math.min(Math.max(limit, 10), 100);
+  const memories = await db.memory.findMany({
+    where: {
+      projectId,
+      userId,
+      isActive: true,
+      validUntil: null
+    },
+    orderBy: { importance: "desc" },
+    take: maxMemories
+  });
+  const clusters = [];
+  const processed = /* @__PURE__ */ new Set();
+  for (let i = 0; i < memories.length; i++) {
+    const memory = memories[i];
+    if (processed.has(memory.id)) continue;
+    const similar = [];
+    const candidates = memories.slice(i + 1);
+    const batchSimilarities = await calculateBatchSimilarity(memory.id, candidates.map((c) => c.id));
+    for (let j = 0; j < candidates.length; j++) {
+      const other = candidates[j];
+      if (processed.has(other.id)) continue;
+      const similarity = batchSimilarities[j];
+      if (similarity >= similarityThreshold) {
+        similar.push({ ...other, similarity });
+        processed.add(other.id);
+      }
+    }
+    if (similar.length > 0) {
+      clusters.push({
+        representative: memory,
+        duplicates: similar,
+        similarity: similar.reduce((sum, m) => sum + m.similarity, 0) / similar.length
+      });
+      processed.add(memory.id);
+    }
+  }
+  return clusters;
+}
+async function calculateBatchSimilarity(memoryId, otherIds) {
+  if (otherIds.length === 0) return [];
+  const placeholders = otherIds.map((_, i) => `(m1.embedding <=> $${i + 2}::vector)`).join(" + ");
+  const conditions = otherIds.map((id, i) => `m2.id = $${i + 2}`).join(" OR ");
+  const result = await db.$queryRaw`
+    SELECT
+      1 - (m1.embedding <=> m2.embedding) as similarity,
+      m2.id as id
+    FROM memories m1, memories m2
+    WHERE m1.id = ${memoryId} AND (${conditions})
+  `;
+  const similarityMap = new Map(result.map((r) => [r.id, r.similarity]));
+  return otherIds.map((id) => similarityMap.get(id) || 0);
+}
+async function mergeDuplicateMemories(cluster) {
+  const memories = [cluster.representative, ...cluster.duplicates];
+  const prompt = `You are merging duplicate memories into a single, comprehensive memory.
+
+**Memories to merge:**
+${memories.map(
+    (m, i) => `${i + 1}. "${m.content}" (confidence: ${m.confidence}, date: ${m.documentDate?.toISOString() || "unknown"})`
+  ).join("\n")}
+
+**Instructions:**
+1. Combine all unique information from these memories
+2. Resolve any contradictions by keeping the most recent or most confident information
+3. Extract all unique entity mentions
+4. Use the highest confidence score
+5. Keep the most recent document date
+
+Return JSON:
+{
+  "merged_content": "comprehensive merged memory",
+  "entity_mentions": ["list", "of", "entities"],
+  "confidence": 0.0-1.0,
+  "reasoning": "brief explanation of how you merged"
+}`;
+  const response = await openai.chat.completions.create({
+    model: "gpt-4o",
+    max_tokens: 2048,
+    temperature: 0,
+    messages: [{ role: "user", content: prompt }],
+    response_format: { type: "json_object" }
+  });
+  const text = response.choices[0]?.message?.content?.trim();
+  if (!text) {
+    throw new Error("Failed to merge memories");
+  }
+  const jsonMatch = text.match(/```json\n?([\s\S]*?)\n?```/) || text.match(/\{[\s\S]*\}/);
+  const jsonStr = jsonMatch ? jsonMatch[1] || jsonMatch[0] : text;
+  const result = JSON.parse(jsonStr);
+  const embedding = await embedSingle(result.merged_content);
+  const mergedMemory = await db.memory.create({
+    data: {
+      projectId: cluster.representative.projectId,
+      orgId: cluster.representative.orgId,
+      userId: cluster.representative.userId,
+      sessionId: cluster.representative.sessionId,
+      memoryType: cluster.representative.memoryType,
+      content: result.merged_content,
+      embedding,
+      entityMentions: result.entity_mentions || [],
+      confidence: result.confidence || cluster.representative.confidence,
+      documentDate: cluster.representative.documentDate,
+      eventDate: cluster.representative.eventDate,
+      validFrom: /* @__PURE__ */ new Date(),
+      importance: Math.max(...memories.map((m) => m.importance || 0.5)),
+      metadata: {
+        mergedFrom: memories.map((m) => m.id),
+        mergeReasoning: result.reasoning,
+        mergedAt: (/* @__PURE__ */ new Date()).toISOString()
+      }
+    }
+  });
+  for (const memory of memories) {
+    await db.memory.update({
+      where: { id: memory.id },
+      data: {
+        isActive: false,
+        validUntil: /* @__PURE__ */ new Date(),
+        supersededBy: mergedMemory.id
+      }
+    });
+  }
+  return mergedMemory.id;
+}
+async function consolidateMemories(params) {
+  const { projectId, userId, similarityThreshold = 0.95, dryRun = false } = params;
+  console.log(`\u{1F50D} Finding duplicate memories in project ${projectId}...`);
+  const clusters = await findDuplicateMemories({
+    projectId,
+    userId,
+    similarityThreshold
+  });
+  console.log(`\u{1F4CA} Found ${clusters.length} memory clusters`);
+  if (dryRun) {
+    for (const cluster of clusters) {
+      console.log(`
+Cluster (similarity: ${cluster.similarity.toFixed(2)}):`);
+      console.log(`  Representative: "${cluster.representative.content}"`);
+      console.log(`  Duplicates: ${cluster.duplicates.length}`);
+      cluster.duplicates.forEach((d) => {
+        console.log(`    - "${d.content}"`);
+      });
+    }
+    return {
+      clustersFound: clusters.length,
+      memoriesMerged: 0,
+      memoriesDeactivated: 0
+    };
+  }
+  let memoriesMerged = 0;
+  let memoriesDeactivated = 0;
+  for (const cluster of clusters) {
+    try {
+      console.log(`\u{1F517} Merging cluster with ${cluster.duplicates.length + 1} memories...`);
+      await mergeDuplicateMemories(cluster);
+      memoriesMerged++;
+      memoriesDeactivated += cluster.duplicates.length + 1;
+      console.log(`\u2705 Merged successfully`);
+    } catch (error) {
+      console.error(`\u274C Failed to merge cluster:`, error);
+    }
+  }
+  console.log(
+    `
+\u2705 Consolidation complete: ${memoriesMerged} clusters merged, ${memoriesDeactivated} memories deactivated`
+  );
+  return {
+    clustersFound: clusters.length,
+    memoriesMerged,
+    memoriesDeactivated
+  };
+}
+async function scheduledConsolidation(orgId) {
+  console.log(`\u{1F504} Running scheduled consolidation for org ${orgId}...`);
+  const projects = await db.project.findMany({
+    where: { orgId }
+  });
+  for (const project of projects) {
+    try {
+      const result = await consolidateMemories({
+        projectId: project.id,
+        similarityThreshold: 0.92
+        // Slightly lower for scheduled runs
+      });
+      if (result.memoriesMerged > 0) {
+        console.log(
+          `\u{1F4CA} Project ${project.name}: merged ${result.memoriesMerged} clusters`
+        );
+      }
+    } catch (error) {
+      console.error(`Failed to consolidate project ${project.name}:`, error);
+    }
+  }
+  console.log("\u2705 Scheduled consolidation complete");
+}
+export {
+  consolidateMemories,
+  findDuplicateMemories,
+  mergeDuplicateMemories,
+  scheduledConsolidation
+};