dopple-ai 0.1.0

package/dist/chunk-FA7ZWJOA.js

@@ -0,0 +1,365 @@
+#!/usr/bin/env node
+import {
+  findPatterns,
+  getNeighborhood,
+  getPersonaContext,
+  graphSummary,
+  queryGraph
+} from "./chunk-PGZVVIL6.js";
+
+// src/graph/graph.ts
+import { readFile, writeFile, mkdir } from "fs/promises";
+import { dirname } from "path";
+
+// src/graph/builder.ts
+import { randomUUID } from "crypto";
+async function buildGraph(llm, userData, contextTexts) {
+  const nodes = [];
+  const edges = [];
+  const sources = /* @__PURE__ */ new Set();
+  extractFromUserData(userData, nodes, edges, sources);
+  if (contextTexts && contextTexts.length > 0) {
+    await extractFromText(llm, contextTexts, nodes, edges, sources);
+  }
+  if (nodes.length > 2) {
+    await discoverRelationships(llm, nodes, edges);
+  }
+  const deduped = deduplicateNodes(nodes, edges);
+  return {
+    nodes: deduped.nodes,
+    edges: deduped.edges,
+    metadata: {
+      createdAt: (/* @__PURE__ */ new Date()).toISOString(),
+      sources: [...sources],
+      nodeCount: deduped.nodes.length,
+      edgeCount: deduped.edges.length
+    }
+  };
+}
+function extractFromUserData(records, nodes, edges, sources) {
+  const eventCounts = {};
+  const propertyValues = {};
+  for (const record of records) {
+    sources.add("user-data");
+    const userId = `user-${record.id}`;
+    nodes.push({
+      id: userId,
+      type: "user",
+      label: String(record.properties.name ?? record.properties.email ?? record.id),
+      properties: record.properties,
+      source: "user-data"
+    });
+    for (const event of record.events ?? []) {
+      eventCounts[event.name] = (eventCounts[event.name] ?? 0) + 1;
+      const eventNodeId = `event-${event.name}`;
+      if (!nodes.some((n) => n.id === eventNodeId)) {
+        nodes.push({
+          id: eventNodeId,
+          type: "event",
+          label: event.name,
+          properties: { totalOccurrences: 0 },
+          source: "user-data"
+        });
+      }
+      edges.push({
+        id: randomUUID(),
+        from: userId,
+        to: eventNodeId,
+        type: "triggered",
+        weight: 0.5,
+        properties: { timestamp: event.timestamp },
+        source: "user-data"
+      });
+    }
+    for (const payment of record.payments ?? []) {
+      const planId = `plan-${payment.plan ?? "unknown"}`;
+      if (!nodes.some((n) => n.id === planId)) {
+        nodes.push({
+          id: planId,
+          type: "product",
+          label: payment.plan ?? "unknown plan",
+          properties: { amount: payment.amount, currency: payment.currency },
+          source: "payment-data"
+        });
+      }
+      edges.push({
+        id: randomUUID(),
+        from: userId,
+        to: planId,
+        type: payment.status === "cancelled" ? "cancelled" : "subscribes_to",
+        weight: payment.status === "active" ? 0.8 : 0.3,
+        properties: {
+          status: payment.status,
+          cancelReason: payment.cancelReason
+        },
+        source: "payment-data"
+      });
+      if (payment.cancelReason) {
+        const complaintId = `complaint-${randomUUID().slice(0, 8)}`;
+        nodes.push({
+          id: complaintId,
+          type: "complaint",
+          label: payment.cancelReason,
+          properties: {},
+          source: "payment-data"
+        });
+        edges.push({
+          id: randomUUID(),
+          from: userId,
+          to: complaintId,
+          type: "complained_about",
+          weight: 0.9,
+          properties: {},
+          source: "payment-data"
+        });
+      }
+    }
+    for (const conv of record.conversations ?? []) {
+      const convId = `conv-${randomUUID().slice(0, 8)}`;
+      nodes.push({
+        id: convId,
+        type: "complaint",
+        label: conv.slice(0, 100),
+        properties: { fullText: conv },
+        source: "support-data"
+      });
+      edges.push({
+        id: randomUUID(),
+        from: userId,
+        to: convId,
+        type: "said",
+        weight: 0.7,
+        properties: {},
+        source: "support-data"
+      });
+    }
+    for (const [key, value] of Object.entries(record.properties)) {
+      if (!propertyValues[key]) propertyValues[key] = {};
+      const v = String(value);
+      propertyValues[key][v] = (propertyValues[key][v] ?? 0) + 1;
+    }
+  }
+  for (const node of nodes) {
+    if (node.type === "event" && eventCounts[node.label]) {
+      node.properties.totalOccurrences = eventCounts[node.label];
+    }
+  }
+  const totalUsers = records.length;
+  if (totalUsers > 0) {
+    nodes.push({
+      id: "metric-total-users",
+      type: "metric",
+      label: `${totalUsers} total users`,
+      properties: { value: totalUsers },
+      source: "user-data"
+    });
+  }
+}
+async function extractFromText(llm, texts, nodes, edges, sources) {
+  sources.add("context");
+  const combined = texts.join("\n\n").slice(0, 8e3);
+  const extracted = await llm.generateJSON(
+    "You extract entities and relationships from text to build a knowledge graph.",
+    `Extract entities and relationships from this text.
+
+${combined}
+
+Return JSON:
+{
+  "entities": [{"label": "<name>", "type": "user|behavior|feature|complaint|segment|event|product|metric|entity", "properties": {}}],
+  "relationships": [{"from": "<entity label>", "to": "<entity label>", "type": "<relationship>", "weight": <0-1>}]
+}
+
+Focus on: people, products, features, complaints, behaviors, events, and how they relate.
+Use descriptive relationship types like "uses", "complained_about", "wants", "avoids", "correlates_with".`
+  );
+  const labelToId = {};
+  for (const entity of extracted.entities ?? []) {
+    const id = `ctx-${randomUUID().slice(0, 8)}`;
+    labelToId[entity.label] = id;
+    nodes.push({
+      id,
+      type: entity.type,
+      label: entity.label,
+      properties: entity.properties ?? {},
+      source: "context"
+    });
+  }
+  for (const rel of extracted.relationships ?? []) {
+    const fromId = labelToId[rel.from];
+    const toId = labelToId[rel.to];
+    if (fromId && toId) {
+      edges.push({
+        id: randomUUID(),
+        from: fromId,
+        to: toId,
+        type: rel.type,
+        weight: rel.weight ?? 0.5,
+        properties: {},
+        source: "context"
+      });
+    }
+  }
+}
+async function discoverRelationships(llm, nodes, edges) {
+  const significantNodes = nodes.filter((n) => n.type !== "user").slice(0, 30).map((n) => `${n.type}:${n.label}`);
+  if (significantNodes.length < 3) return;
+  const discovered = await llm.generateJSON(
+    "You identify non-obvious relationships between entities in a product knowledge graph.",
+    `Given these entities from a product's user data, identify relationships between them that aren't already obvious.
+
+Entities:
+${significantNodes.join("\n")}
+
+Return a JSON array of relationships:
+[{"from": "<entity label>", "to": "<entity label>", "type": "<relationship>", "weight": <0-1>}]
+
+Focus on: correlations ("users who do X also do Y"), causal patterns ("feature X leads to churn"), and segments ("these behaviors cluster together").
+Return at most 10 relationships. Only include ones with real analytical value.`
+  );
+  for (const rel of discovered ?? []) {
+    const fromNode = nodes.find((n) => n.label === rel.from.replace(/^\w+:/, ""));
+    const toNode = nodes.find((n) => n.label === rel.to.replace(/^\w+:/, ""));
+    if (fromNode && toNode) {
+      edges.push({
+        id: randomUUID(),
+        from: fromNode.id,
+        to: toNode.id,
+        type: rel.type,
+        weight: rel.weight ?? 0.5,
+        properties: { discovered: true },
+        source: "llm-analysis"
+      });
+    }
+  }
+}
+function deduplicateNodes(nodes, edges) {
+  const seen = /* @__PURE__ */ new Map();
+  const idMapping = /* @__PURE__ */ new Map();
+  for (const node of nodes) {
+    const key = `${node.type}:${node.label.toLowerCase()}`;
+    if (seen.has(key)) {
+      idMapping.set(node.id, seen.get(key).id);
+    } else {
+      seen.set(key, node);
+    }
+  }
+  const remappedEdges = edges.map((e) => ({
+    ...e,
+    from: idMapping.get(e.from) ?? e.from,
+    to: idMapping.get(e.to) ?? e.to
+  }));
+  const cleanEdges = remappedEdges.filter((e) => e.from !== e.to);
+  return { nodes: [...seen.values()], edges: cleanEdges };
+}
+
+// src/graph/graph.ts
+var DoppleGraph = class {
+  graph = null;
+  /**
+   * Build the graph from user data and optional context texts.
+   */
+  async build(llm, userData, contextTexts) {
+    this.graph = await buildGraph(llm, userData, contextTexts);
+  }
+  /**
+   * Get nodes within N hops of a starting node.
+   */
+  query(nodeId, depth = 2) {
+    this.ensureBuilt();
+    return getNeighborhood(this.graph, nodeId, depth);
+  }
+  /**
+   * Get context strings relevant to a persona's traits.
+   */
+  getPersonaContext(traits) {
+    this.ensureBuilt();
+    return getPersonaContext(this.graph, traits);
+  }
+  /**
+   * Find patterns in the graph (hubs, churn, complaints, usage).
+   */
+  patterns() {
+    this.ensureBuilt();
+    return findPatterns(this.graph);
+  }
+  /**
+   * Human-readable summary.
+   */
+  get summary() {
+    this.ensureBuilt();
+    return graphSummary(this.graph);
+  }
+  /**
+   * Get the raw graph data.
+   */
+  get raw() {
+    this.ensureBuilt();
+    return this.graph;
+  }
+  /**
+   * Ask the graph a natural language question.
+   * LLM interprets the question, finds relevant data, returns grounded answer.
+   */
+  async ask(llm, question) {
+    this.ensureBuilt();
+    return queryGraph(llm, this.graph, question);
+  }
+  get isBuilt() {
+    return this.graph !== null;
+  }
+  /**
+   * Save graph to JSONL file.
+   */
+  async save(path) {
+    this.ensureBuilt();
+    await mkdir(dirname(path), { recursive: true });
+    const lines = [];
+    lines.push(JSON.stringify({ _type: "metadata", ...this.graph.metadata }));
+    for (const node of this.graph.nodes) {
+      lines.push(JSON.stringify({ _type: "node", ...node }));
+    }
+    for (const edge of this.graph.edges) {
+      lines.push(JSON.stringify({ _type: "edge", ...edge }));
+    }
+    await writeFile(path, lines.join("\n") + "\n");
+  }
+  /**
+   * Load graph from JSONL file.
+   */
+  async load(path) {
+    const content = await readFile(path, "utf-8");
+    const lines = content.trim().split("\n").filter(Boolean);
+    const nodes = [];
+    const edges = [];
+    let metadata = {
+      createdAt: "",
+      sources: [],
+      nodeCount: 0,
+      edgeCount: 0
+    };
+    for (const line of lines) {
+      const obj = JSON.parse(line);
+      const type = obj._type;
+      delete obj._type;
+      if (type === "metadata") {
+        metadata = obj;
+      } else if (type === "node") {
+        nodes.push(obj);
+      } else if (type === "edge") {
+        edges.push(obj);
+      }
+    }
+    this.graph = { nodes, edges, metadata };
+  }
+  ensureBuilt() {
+    if (!this.graph) {
+      throw new Error("Graph not built yet. Call build() or load() first.");
+    }
+  }
+};
+
+export {
+  DoppleGraph
+};
+//# sourceMappingURL=chunk-FA7ZWJOA.js.map

package/dist/chunk-FA7ZWJOA.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/graph/graph.ts","../src/graph/builder.ts"],"sourcesContent":["/**\n * DoppleGraph — in-memory knowledge graph with persistence.\n */\n\nimport { readFile, writeFile, mkdir } from \"node:fs/promises\";\nimport { join, dirname } from \"node:path\";\nimport type { LLMProvider, UserRecord, OceanVector } from \"../types.js\";\nimport type { KnowledgeGraph, GraphNode, GraphPattern } from \"./types.js\";\nimport { buildGraph } from \"./builder.js\";\nimport {\n  getNeighborhood,\n  findPatterns,\n  getPersonaContext,\n  graphSummary,\n  queryGraph,\n  type GraphQueryResult,\n} from \"./query.js\";\n\nexport class DoppleGraph {\n  private graph: KnowledgeGraph | null = null;\n\n  /**\n   * Build the graph from user data and optional context texts.\n   */\n  async build(\n    llm: LLMProvider,\n    userData: UserRecord[],\n    contextTexts?: string[]\n  ): Promise<void> {\n    this.graph = await buildGraph(llm, userData, contextTexts);\n  }\n\n  /**\n   * Get nodes within N hops of a starting node.\n   */\n  query(nodeId: string, depth: number = 2): GraphNode[] {\n    this.ensureBuilt();\n    return getNeighborhood(this.graph!, nodeId, depth);\n  }\n\n  /**\n   * Get context strings relevant to a persona's traits.\n   */\n  getPersonaContext(traits: OceanVector): string[] {\n    this.ensureBuilt();\n    return getPersonaContext(this.graph!, traits);\n  }\n\n  /**\n   * Find patterns in the graph (hubs, churn, complaints, usage).\n   */\n  patterns(): GraphPattern[] {\n    this.ensureBuilt();\n    return findPatterns(this.graph!);\n  }\n\n  /**\n   * Human-readable summary.\n   */\n  get summary(): string {\n    this.ensureBuilt();\n    return graphSummary(this.graph!);\n  }\n\n  /**\n   * Get the raw graph data.\n   */\n  get raw(): KnowledgeGraph {\n    this.ensureBuilt();\n    return this.graph!;\n  }\n\n  /**\n   * Ask the graph a natural language question.\n   * LLM interprets the question, finds relevant data, returns grounded answer.\n   */\n  async ask(llm: LLMProvider, question: string): Promise<GraphQueryResult> {\n    this.ensureBuilt();\n    return queryGraph(llm, this.graph!, question);\n  }\n\n  get isBuilt(): boolean {\n    return this.graph !== null;\n  }\n\n  /**\n   * Save graph to JSONL file.\n   */\n  async save(path: string): Promise<void> {\n    this.ensureBuilt();\n    await mkdir(dirname(path), { recursive: true });\n\n    const lines: string[] = [];\n    lines.push(JSON.stringify({ _type: \"metadata\", ...this.graph!.metadata }));\n    for (const node of this.graph!.nodes) {\n      lines.push(JSON.stringify({ _type: \"node\", ...node }));\n    }\n    for (const edge of this.graph!.edges) {\n      lines.push(JSON.stringify({ _type: \"edge\", ...edge }));\n    }\n\n    await writeFile(path, lines.join(\"\\n\") + \"\\n\");\n  }\n\n  /**\n   * Load graph from JSONL file.\n   */\n  async load(path: string): Promise<void> {\n    const content = await readFile(path, \"utf-8\");\n    const lines = content.trim().split(\"\\n\").filter(Boolean);\n\n    const nodes: GraphNode[] = [];\n    const edges: import(\"./types.js\").GraphEdge[] = [];\n    let metadata: KnowledgeGraph[\"metadata\"] = {\n      createdAt: \"\",\n      sources: [],\n      nodeCount: 0,\n      edgeCount: 0,\n    };\n\n    for (const line of lines) {\n      const obj = JSON.parse(line) as Record<string, unknown>;\n      const type = obj._type;\n      delete obj._type;\n\n      if (type === \"metadata\") {\n        metadata = obj as unknown as KnowledgeGraph[\"metadata\"];\n      } else if (type === \"node\") {\n        nodes.push(obj as unknown as GraphNode);\n      } else if (type === \"edge\") {\n        edges.push(obj as unknown as import(\"./types.js\").GraphEdge);\n      }\n    }\n\n    this.graph = { nodes, edges, metadata };\n  }\n\n  private ensureBuilt(): void {\n    if (!this.graph) {\n      throw new Error(\"Graph not built yet. Call build() or load() first.\");\n    }\n  }\n}\n","/**\n * Knowledge graph builder.\n *\n * Two modes:\n * 1. Structured: UserRecord[] from adapters → deterministic entity extraction → graph\n * 2. Unstructured: freeform text → LLM entity extraction → graph\n *\n * No external graph database dependency — in-memory with JSONL persistence.\n */\n\nimport { randomUUID } from \"node:crypto\";\nimport type { LLMProvider, UserRecord } from \"../types.js\";\nimport type { GraphNode, GraphEdge, KnowledgeGraph, NodeType } from \"./types.js\";\n\n/**\n * Build a knowledge graph from structured user data and optional unstructured context.\n */\nexport async function buildGraph(\n  llm: LLMProvider,\n  userData: UserRecord[],\n  contextTexts?: string[]\n): Promise<KnowledgeGraph> {\n  const nodes: GraphNode[] = [];\n  const edges: GraphEdge[] = [];\n  const sources = new Set<string>();\n\n  // Phase 1: Extract entities from structured data (deterministic)\n  extractFromUserData(userData, nodes, edges, sources);\n\n  // Phase 2: Extract entities from unstructured text (LLM-assisted)\n  if (contextTexts && contextTexts.length > 0) {\n    await extractFromText(llm, contextTexts, nodes, edges, sources);\n  }\n\n  // Phase 3: Ask LLM to discover relationships between existing nodes\n  if (nodes.length > 2) {\n    await discoverRelationships(llm, nodes, edges);\n  }\n\n  // Deduplicate nodes by label (case-insensitive)\n  const deduped = deduplicateNodes(nodes, edges);\n\n  return {\n    nodes: deduped.nodes,\n    edges: deduped.edges,\n    metadata: {\n      createdAt: new Date().toISOString(),\n      sources: [...sources],\n      nodeCount: deduped.nodes.length,\n      edgeCount: deduped.edges.length,\n    },\n  };\n}\n\n// ---------------------------------------------------------------------------\n// Phase 1: Structured data extraction\n// ---------------------------------------------------------------------------\n\nfunction extractFromUserData(\n  records: UserRecord[],\n  nodes: GraphNode[],\n  edges: GraphEdge[],\n  sources: Set<string>\n): void {\n  // Track feature/event usage across users\n  const eventCounts: Record<string, number> = {};\n  const propertyValues: Record<string, Record<string, number>> = {};\n\n  for (const record of records) {\n    sources.add(\"user-data\");\n\n    // Create user node\n    const userId = `user-${record.id}`;\n    nodes.push({\n      id: userId,\n      type: \"user\",\n      label: String(record.properties.name ?? record.properties.email ?? record.id),\n      properties: record.properties,\n      source: \"user-data\",\n    });\n\n    // Extract events as behavior/feature nodes\n    for (const event of record.events ?? []) {\n      eventCounts[event.name] = (eventCounts[event.name] ?? 0) + 1;\n\n      const eventNodeId = `event-${event.name}`;\n      if (!nodes.some((n) => n.id === eventNodeId)) {\n        nodes.push({\n          id: eventNodeId,\n          type: \"event\",\n          label: event.name,\n          properties: { totalOccurrences: 0 },\n          source: \"user-data\",\n        });\n      }\n\n      // Edge: user → uses → event\n      edges.push({\n        id: randomUUID(),\n        from: userId,\n        to: eventNodeId,\n        type: \"triggered\",\n        weight: 0.5,\n        properties: { timestamp: event.timestamp },\n        source: \"user-data\",\n      });\n    }\n\n    // Extract payment data\n    for (const payment of record.payments ?? []) {\n      const planId = `plan-${payment.plan ?? \"unknown\"}`;\n      if (!nodes.some((n) => n.id === planId)) {\n        nodes.push({\n          id: planId,\n          type: \"product\",\n          label: payment.plan ?? \"unknown plan\",\n          properties: { amount: payment.amount, currency: payment.currency },\n          source: \"payment-data\",\n        });\n      }\n\n      edges.push({\n        id: randomUUID(),\n        from: userId,\n        to: planId,\n        type: payment.status === \"cancelled\" ? \"cancelled\" : \"subscribes_to\",\n        weight: payment.status === \"active\" ? 0.8 : 0.3,\n        properties: {\n          status: payment.status,\n          cancelReason: payment.cancelReason,\n        },\n        source: \"payment-data\",\n      });\n\n      // Create complaint node from cancel reason\n      if (payment.cancelReason) {\n        const complaintId = `complaint-${randomUUID().slice(0, 8)}`;\n        nodes.push({\n          id: complaintId,\n          type: \"complaint\",\n          label: payment.cancelReason,\n          properties: {},\n          source: \"payment-data\",\n        });\n        edges.push({\n          id: randomUUID(),\n          from: userId,\n          to: complaintId,\n          type: \"complained_about\",\n          weight: 0.9,\n          properties: {},\n          source: \"payment-data\",\n        });\n      }\n    }\n\n    // Extract conversations as complaint/feature nodes\n    for (const conv of record.conversations ?? []) {\n      const convId = `conv-${randomUUID().slice(0, 8)}`;\n      nodes.push({\n        id: convId,\n        type: \"complaint\",\n        label: conv.slice(0, 100),\n        properties: { fullText: conv },\n        source: \"support-data\",\n      });\n      edges.push({\n        id: randomUUID(),\n        from: userId,\n        to: convId,\n        type: \"said\",\n        weight: 0.7,\n        properties: {},\n        source: \"support-data\",\n      });\n    }\n\n    // Track property distributions\n    for (const [key, value] of Object.entries(record.properties)) {\n      if (!propertyValues[key]) propertyValues[key] = {};\n      const v = String(value);\n      propertyValues[key][v] = (propertyValues[key][v] ?? 0) + 1;\n    }\n  }\n\n  // Update event occurrence counts\n  for (const node of nodes) {\n    if (node.type === \"event\" && eventCounts[node.label]) {\n      node.properties.totalOccurrences = eventCounts[node.label];\n    }\n  }\n\n  // Create metric nodes for key aggregates\n  const totalUsers = records.length;\n  if (totalUsers > 0) {\n    nodes.push({\n      id: \"metric-total-users\",\n      type: \"metric\",\n      label: `${totalUsers} total users`,\n      properties: { value: totalUsers },\n      source: \"user-data\",\n    });\n  }\n}\n\n// ---------------------------------------------------------------------------\n// Phase 2: Unstructured text extraction\n// ---------------------------------------------------------------------------\n\nasync function extractFromText(\n  llm: LLMProvider,\n  texts: string[],\n  nodes: GraphNode[],\n  edges: GraphEdge[],\n  sources: Set<string>\n): Promise<void> {\n  sources.add(\"context\");\n\n  const combined = texts.join(\"\\n\\n\").slice(0, 8000); // Cap context length\n\n  const extracted = await llm.generateJSON<{\n    entities: Array<{\n      label: string;\n      type: NodeType;\n      properties: Record<string, unknown>;\n    }>;\n    relationships: Array<{\n      from: string;\n      to: string;\n      type: string;\n      weight: number;\n    }>;\n  }>(\n    \"You extract entities and relationships from text to build a knowledge graph.\",\n    `Extract entities and relationships from this text.\n\n${combined}\n\nReturn JSON:\n{\n  \"entities\": [{\"label\": \"<name>\", \"type\": \"user|behavior|feature|complaint|segment|event|product|metric|entity\", \"properties\": {}}],\n  \"relationships\": [{\"from\": \"<entity label>\", \"to\": \"<entity label>\", \"type\": \"<relationship>\", \"weight\": <0-1>}]\n}\n\nFocus on: people, products, features, complaints, behaviors, events, and how they relate.\nUse descriptive relationship types like \"uses\", \"complained_about\", \"wants\", \"avoids\", \"correlates_with\".`\n  );\n\n  const labelToId: Record<string, string> = {};\n\n  for (const entity of extracted.entities ?? []) {\n    const id = `ctx-${randomUUID().slice(0, 8)}`;\n    labelToId[entity.label] = id;\n    nodes.push({\n      id,\n      type: entity.type,\n      label: entity.label,\n      properties: entity.properties ?? {},\n      source: \"context\",\n    });\n  }\n\n  for (const rel of extracted.relationships ?? []) {\n    const fromId = labelToId[rel.from];\n    const toId = labelToId[rel.to];\n    if (fromId && toId) {\n      edges.push({\n        id: randomUUID(),\n        from: fromId,\n        to: toId,\n        type: rel.type,\n        weight: rel.weight ?? 0.5,\n        properties: {},\n        source: \"context\",\n      });\n    }\n  }\n}\n\n// ---------------------------------------------------------------------------\n// Phase 3: Relationship discovery\n// ---------------------------------------------------------------------------\n\nasync function discoverRelationships(\n  llm: LLMProvider,\n  nodes: GraphNode[],\n  edges: GraphEdge[]\n): Promise<void> {\n  // Only use non-user nodes for relationship discovery (to keep prompt small)\n  const significantNodes = nodes\n    .filter((n) => n.type !== \"user\")\n    .slice(0, 30)\n    .map((n) => `${n.type}:${n.label}`);\n\n  if (significantNodes.length < 3) return;\n\n  const discovered = await llm.generateJSON<\n    Array<{ from: string; to: string; type: string; weight: number }>\n  >(\n    \"You identify non-obvious relationships between entities in a product knowledge graph.\",\n    `Given these entities from a product's user data, identify relationships between them that aren't already obvious.\n\nEntities:\n${significantNodes.join(\"\\n\")}\n\nReturn a JSON array of relationships:\n[{\"from\": \"<entity label>\", \"to\": \"<entity label>\", \"type\": \"<relationship>\", \"weight\": <0-1>}]\n\nFocus on: correlations (\"users who do X also do Y\"), causal patterns (\"feature X leads to churn\"), and segments (\"these behaviors cluster together\").\nReturn at most 10 relationships. Only include ones with real analytical value.`\n  );\n\n  for (const rel of discovered ?? []) {\n    const fromNode = nodes.find((n) => n.label === rel.from.replace(/^\\w+:/, \"\"));\n    const toNode = nodes.find((n) => n.label === rel.to.replace(/^\\w+:/, \"\"));\n    if (fromNode && toNode) {\n      edges.push({\n        id: randomUUID(),\n        from: fromNode.id,\n        to: toNode.id,\n        type: rel.type,\n        weight: rel.weight ?? 0.5,\n        properties: { discovered: true },\n        source: \"llm-analysis\",\n      });\n    }\n  }\n}\n\n// ---------------------------------------------------------------------------\n// Deduplication\n// ---------------------------------------------------------------------------\n\nfunction deduplicateNodes(\n  nodes: GraphNode[],\n  edges: GraphEdge[]\n): { nodes: GraphNode[]; edges: GraphEdge[] } {\n  const seen = new Map<string, GraphNode>();\n  const idMapping = new Map<string, string>();\n\n  for (const node of nodes) {\n    const key = `${node.type}:${node.label.toLowerCase()}`;\n    if (seen.has(key)) {\n      idMapping.set(node.id, seen.get(key)!.id);\n    } else {\n      seen.set(key, node);\n    }\n  }\n\n  // Remap edge references\n  const remappedEdges = edges.map((e) => ({\n    ...e,\n    from: idMapping.get(e.from) ?? e.from,\n    to: idMapping.get(e.to) ?? e.to,\n  }));\n\n  // Remove self-referencing edges\n  const cleanEdges = remappedEdges.filter((e) => e.from !== e.to);\n\n  return { nodes: [...seen.values()], edges: cleanEdges };\n}\n"],"mappings":";;;;;;;;;;AAIA,SAAS,UAAU,WAAW,aAAa;AAC3C,SAAe,eAAe;;;ACK9B,SAAS,kBAAkB;AAO3B,eAAsB,WACpB,KACA,UACA,cACyB;AACzB,QAAM,QAAqB,CAAC;AAC5B,QAAM,QAAqB,CAAC;AAC5B,QAAM,UAAU,oBAAI,IAAY;AAGhC,sBAAoB,UAAU,OAAO,OAAO,OAAO;AAGnD,MAAI,gBAAgB,aAAa,SAAS,GAAG;AAC3C,UAAM,gBAAgB,KAAK,cAAc,OAAO,OAAO,OAAO;AAAA,EAChE;AAGA,MAAI,MAAM,SAAS,GAAG;AACpB,UAAM,sBAAsB,KAAK,OAAO,KAAK;AAAA,EAC/C;AAGA,QAAM,UAAU,iBAAiB,OAAO,KAAK;AAE7C,SAAO;AAAA,IACL,OAAO,QAAQ;AAAA,IACf,OAAO,QAAQ;AAAA,IACf,UAAU;AAAA,MACR,YAAW,oBAAI,KAAK,GAAE,YAAY;AAAA,MAClC,SAAS,CAAC,GAAG,OAAO;AAAA,MACpB,WAAW,QAAQ,MAAM;AAAA,MACzB,WAAW,QAAQ,MAAM;AAAA,IAC3B;AAAA,EACF;AACF;AAMA,SAAS,oBACP,SACA,OACA,OACA,SACM;AAEN,QAAM,cAAsC,CAAC;AAC7C,QAAM,iBAAyD,CAAC;AAEhE,aAAW,UAAU,SAAS;AAC5B,YAAQ,IAAI,WAAW;AAGvB,UAAM,SAAS,QAAQ,OAAO,EAAE;AAChC,UAAM,KAAK;AAAA,MACT,IAAI;AAAA,MACJ,MAAM;AAAA,MACN,OAAO,OAAO,OAAO,WAAW,QAAQ,OAAO,WAAW,SAAS,OAAO,EAAE;AAAA,MAC5E,YAAY,OAAO;AAAA,MACnB,QAAQ;AAAA,IACV,CAAC;AAGD,eAAW,SAAS,OAAO,UAAU,CAAC,GAAG;AACvC,kBAAY,MAAM,IAAI,KAAK,YAAY,MAAM,IAAI,KAAK,KAAK;AAE3D,YAAM,cAAc,SAAS,MAAM,IAAI;AACvC,UAAI,CAAC,MAAM,KAAK,CAAC,MAAM,EAAE,OAAO,WAAW,GAAG;AAC5C,cAAM,KAAK;AAAA,UACT,IAAI;AAAA,UACJ,MAAM;AAAA,UACN,OAAO,MAAM;AAAA,UACb,YAAY,EAAE,kBAAkB,EAAE;AAAA,UAClC,QAAQ;AAAA,QACV,CAAC;AAAA,MACH;AAGA,YAAM,KAAK;AAAA,QACT,IAAI,WAAW;AAAA,QACf,MAAM;AAAA,QACN,IAAI;AAAA,QACJ,MAAM;AAAA,QACN,QAAQ;AAAA,QACR,YAAY,EAAE,WAAW,MAAM,UAAU;AAAA,QACzC,QAAQ;AAAA,MACV,CAAC;AAAA,IACH;AAGA,eAAW,WAAW,OAAO,YAAY,CAAC,GAAG;AAC3C,YAAM,SAAS,QAAQ,QAAQ,QAAQ,SAAS;AAChD,UAAI,CAAC,MAAM,KAAK,CAAC,MAAM,EAAE,OAAO,MAAM,GAAG;AACvC,cAAM,KAAK;AAAA,UACT,IAAI;AAAA,UACJ,MAAM;AAAA,UACN,OAAO,QAAQ,QAAQ;AAAA,UACvB,YAAY,EAAE,QAAQ,QAAQ,QAAQ,UAAU,QAAQ,SAAS;AAAA,UACjE,QAAQ;AAAA,QACV,CAAC;AAAA,MACH;AAEA,YAAM,KAAK;AAAA,QACT,IAAI,WAAW;AAAA,QACf,MAAM;AAAA,QACN,IAAI;AAAA,QACJ,MAAM,QAAQ,WAAW,cAAc,cAAc;AAAA,QACrD,QAAQ,QAAQ,WAAW,WAAW,MAAM;AAAA,QAC5C,YAAY;AAAA,UACV,QAAQ,QAAQ;AAAA,UAChB,cAAc,QAAQ;AAAA,QACxB;AAAA,QACA,QAAQ;AAAA,MACV,CAAC;AAGD,UAAI,QAAQ,cAAc;AACxB,cAAM,cAAc,aAAa,WAAW,EAAE,MAAM,GAAG,CAAC,CAAC;AACzD,cAAM,KAAK;AAAA,UACT,IAAI;AAAA,UACJ,MAAM;AAAA,UACN,OAAO,QAAQ;AAAA,UACf,YAAY,CAAC;AAAA,UACb,QAAQ;AAAA,QACV,CAAC;AACD,cAAM,KAAK;AAAA,UACT,IAAI,WAAW;AAAA,UACf,MAAM;AAAA,UACN,IAAI;AAAA,UACJ,MAAM;AAAA,UACN,QAAQ;AAAA,UACR,YAAY,CAAC;AAAA,UACb,QAAQ;AAAA,QACV,CAAC;AAAA,MACH;AAAA,IACF;AAGA,eAAW,QAAQ,OAAO,iBAAiB,CAAC,GAAG;AAC7C,YAAM,SAAS,QAAQ,WAAW,EAAE,MAAM,GAAG,CAAC,CAAC;AAC/C,YAAM,KAAK;AAAA,QACT,IAAI;AAAA,QACJ,MAAM;AAAA,QACN,OAAO,KAAK,MAAM,GAAG,GAAG;AAAA,QACxB,YAAY,EAAE,UAAU,KAAK;AAAA,QAC7B,QAAQ;AAAA,MACV,CAAC;AACD,YAAM,KAAK;AAAA,QACT,IAAI,WAAW;AAAA,QACf,MAAM;AAAA,QACN,IAAI;AAAA,QACJ,MAAM;AAAA,QACN,QAAQ;AAAA,QACR,YAAY,CAAC;AAAA,QACb,QAAQ;AAAA,MACV,CAAC;AAAA,IACH;AAGA,eAAW,CAAC,KAAK,KAAK,KAAK,OAAO,QAAQ,OAAO,UAAU,GAAG;AAC5D,UAAI,CAAC,eAAe,GAAG,EAAG,gBAAe,GAAG,IAAI,CAAC;AACjD,YAAM,IAAI,OAAO,KAAK;AACtB,qBAAe,GAAG,EAAE,CAAC,KAAK,eAAe,GAAG,EAAE,CAAC,KAAK,KAAK;AAAA,IAC3D;AAAA,EACF;AAGA,aAAW,QAAQ,OAAO;AACxB,QAAI,KAAK,SAAS,WAAW,YAAY,KAAK,KAAK,GAAG;AACpD,WAAK,WAAW,mBAAmB,YAAY,KAAK,KAAK;AAAA,IAC3D;AAAA,EACF;AAGA,QAAM,aAAa,QAAQ;AAC3B,MAAI,aAAa,GAAG;AAClB,UAAM,KAAK;AAAA,MACT,IAAI;AAAA,MACJ,MAAM;AAAA,MACN,OAAO,GAAG,UAAU;AAAA,MACpB,YAAY,EAAE,OAAO,WAAW;AAAA,MAChC,QAAQ;AAAA,IACV,CAAC;AAAA,EACH;AACF;AAMA,eAAe,gBACb,KACA,OACA,OACA,OACA,SACe;AACf,UAAQ,IAAI,SAAS;AAErB,QAAM,WAAW,MAAM,KAAK,MAAM,EAAE,MAAM,GAAG,GAAI;AAEjD,QAAM,YAAY,MAAM,IAAI;AAAA,IAa1B;AAAA,IACA;AAAA;AAAA,EAEF,QAAQ;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAUR;AAEA,QAAM,YAAoC,CAAC;AAE3C,aAAW,UAAU,UAAU,YAAY,CAAC,GAAG;AAC7C,UAAM,KAAK,OAAO,WAAW,EAAE,MAAM,GAAG,CAAC,CAAC;AAC1C,cAAU,OAAO,KAAK,IAAI;AAC1B,UAAM,KAAK;AAAA,MACT;AAAA,MACA,MAAM,OAAO;AAAA,MACb,OAAO,OAAO;AAAA,MACd,YAAY,OAAO,cAAc,CAAC;AAAA,MAClC,QAAQ;AAAA,IACV,CAAC;AAAA,EACH;AAEA,aAAW,OAAO,UAAU,iBAAiB,CAAC,GAAG;AAC/C,UAAM,SAAS,UAAU,IAAI,IAAI;AACjC,UAAM,OAAO,UAAU,IAAI,EAAE;AAC7B,QAAI,UAAU,MAAM;AAClB,YAAM,KAAK;AAAA,QACT,IAAI,WAAW;AAAA,QACf,MAAM;AAAA,QACN,IAAI;AAAA,QACJ,MAAM,IAAI;AAAA,QACV,QAAQ,IAAI,UAAU;AAAA,QACtB,YAAY,CAAC;AAAA,QACb,QAAQ;AAAA,MACV,CAAC;AAAA,IACH;AAAA,EACF;AACF;AAMA,eAAe,sBACb,KACA,OACA,OACe;AAEf,QAAM,mBAAmB,MACtB,OAAO,CAAC,MAAM,EAAE,SAAS,MAAM,EAC/B,MAAM,GAAG,EAAE,EACX,IAAI,CAAC,MAAM,GAAG,EAAE,IAAI,IAAI,EAAE,KAAK,EAAE;AAEpC,MAAI,iBAAiB,SAAS,EAAG;AAEjC,QAAM,aAAa,MAAM,IAAI;AAAA,IAG3B;AAAA,IACA;AAAA;AAAA;AAAA,EAGF,iBAAiB,KAAK,IAAI,CAAC;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAO3B;AAEA,aAAW,OAAO,cAAc,CAAC,GAAG;AAClC,UAAM,WAAW,MAAM,KAAK,CAAC,MAAM,EAAE,UAAU,IAAI,KAAK,QAAQ,SAAS,EAAE,CAAC;AAC5E,UAAM,SAAS,MAAM,KAAK,CAAC,MAAM,EAAE,UAAU,IAAI,GAAG,QAAQ,SAAS,EAAE,CAAC;AACxE,QAAI,YAAY,QAAQ;AACtB,YAAM,KAAK;AAAA,QACT,IAAI,WAAW;AAAA,QACf,MAAM,SAAS;AAAA,QACf,IAAI,OAAO;AAAA,QACX,MAAM,IAAI;AAAA,QACV,QAAQ,IAAI,UAAU;AAAA,QACtB,YAAY,EAAE,YAAY,KAAK;AAAA,QAC/B,QAAQ;AAAA,MACV,CAAC;AAAA,IACH;AAAA,EACF;AACF;AAMA,SAAS,iBACP,OACA,OAC4C;AAC5C,QAAM,OAAO,oBAAI,IAAuB;AACxC,QAAM,YAAY,oBAAI,IAAoB;AAE1C,aAAW,QAAQ,OAAO;AACxB,UAAM,MAAM,GAAG,KAAK,IAAI,IAAI,KAAK,MAAM,YAAY,CAAC;AACpD,QAAI,KAAK,IAAI,GAAG,GAAG;AACjB,gBAAU,IAAI,KAAK,IAAI,KAAK,IAAI,GAAG,EAAG,EAAE;AAAA,IAC1C,OAAO;AACL,WAAK,IAAI,KAAK,IAAI;AAAA,IACpB;AAAA,EACF;AAGA,QAAM,gBAAgB,MAAM,IAAI,CAAC,OAAO;AAAA,IACtC,GAAG;AAAA,IACH,MAAM,UAAU,IAAI,EAAE,IAAI,KAAK,EAAE;AAAA,IACjC,IAAI,UAAU,IAAI,EAAE,EAAE,KAAK,EAAE;AAAA,EAC/B,EAAE;AAGF,QAAM,aAAa,cAAc,OAAO,CAAC,MAAM,EAAE,SAAS,EAAE,EAAE;AAE9D,SAAO,EAAE,OAAO,CAAC,GAAG,KAAK,OAAO,CAAC,GAAG,OAAO,WAAW;AACxD;;;ADtVO,IAAM,cAAN,MAAkB;AAAA,EACf,QAA+B;AAAA;AAAA;AAAA;AAAA,EAKvC,MAAM,MACJ,KACA,UACA,cACe;AACf,SAAK,QAAQ,MAAM,WAAW,KAAK,UAAU,YAAY;AAAA,EAC3D;AAAA;AAAA;AAAA;AAAA,EAKA,MAAM,QAAgB,QAAgB,GAAgB;AACpD,SAAK,YAAY;AACjB,WAAO,gBAAgB,KAAK,OAAQ,QAAQ,KAAK;AAAA,EACnD;AAAA;AAAA;AAAA;AAAA,EAKA,kBAAkB,QAA+B;AAC/C,SAAK,YAAY;AACjB,WAAO,kBAAkB,KAAK,OAAQ,MAAM;AAAA,EAC9C;AAAA;AAAA;AAAA;AAAA,EAKA,WAA2B;AACzB,SAAK,YAAY;AACjB,WAAO,aAAa,KAAK,KAAM;AAAA,EACjC;AAAA;AAAA;AAAA;AAAA,EAKA,IAAI,UAAkB;AACpB,SAAK,YAAY;AACjB,WAAO,aAAa,KAAK,KAAM;AAAA,EACjC;AAAA;AAAA;AAAA;AAAA,EAKA,IAAI,MAAsB;AACxB,SAAK,YAAY;AACjB,WAAO,KAAK;AAAA,EACd;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,MAAM,IAAI,KAAkB,UAA6C;AACvE,SAAK,YAAY;AACjB,WAAO,WAAW,KAAK,KAAK,OAAQ,QAAQ;AAAA,EAC9C;AAAA,EAEA,IAAI,UAAmB;AACrB,WAAO,KAAK,UAAU;AAAA,EACxB;AAAA;AAAA;AAAA;AAAA,EAKA,MAAM,KAAK,MAA6B;AACtC,SAAK,YAAY;AACjB,UAAM,MAAM,QAAQ,IAAI,GAAG,EAAE,WAAW,KAAK,CAAC;AAE9C,UAAM,QAAkB,CAAC;AACzB,UAAM,KAAK,KAAK,UAAU,EAAE,OAAO,YAAY,GAAG,KAAK,MAAO,SAAS,CAAC,CAAC;AACzE,eAAW,QAAQ,KAAK,MAAO,OAAO;AACpC,YAAM,KAAK,KAAK,UAAU,EAAE,OAAO,QAAQ,GAAG,KAAK,CAAC,CAAC;AAAA,IACvD;AACA,eAAW,QAAQ,KAAK,MAAO,OAAO;AACpC,YAAM,KAAK,KAAK,UAAU,EAAE,OAAO,QAAQ,GAAG,KAAK,CAAC,CAAC;AAAA,IACvD;AAEA,UAAM,UAAU,MAAM,MAAM,KAAK,IAAI,IAAI,IAAI;AAAA,EAC/C;AAAA;AAAA;AAAA;AAAA,EAKA,MAAM,KAAK,MAA6B;AACtC,UAAM,UAAU,MAAM,SAAS,MAAM,OAAO;AAC5C,UAAM,QAAQ,QAAQ,KAAK,EAAE,MAAM,IAAI,EAAE,OAAO,OAAO;AAEvD,UAAM,QAAqB,CAAC;AAC5B,UAAM,QAA0C,CAAC;AACjD,QAAI,WAAuC;AAAA,MACzC,WAAW;AAAA,MACX,SAAS,CAAC;AAAA,MACV,WAAW;AAAA,MACX,WAAW;AAAA,IACb;AAEA,eAAW,QAAQ,OAAO;AACxB,YAAM,MAAM,KAAK,MAAM,IAAI;AAC3B,YAAM,OAAO,IAAI;AACjB,aAAO,IAAI;AAEX,UAAI,SAAS,YAAY;AACvB,mBAAW;AAAA,MACb,WAAW,SAAS,QAAQ;AAC1B,cAAM,KAAK,GAA2B;AAAA,MACxC,WAAW,SAAS,QAAQ;AAC1B,cAAM,KAAK,GAAgD;AAAA,MAC7D;AAAA,IACF;AAEA,SAAK,QAAQ,EAAE,OAAO,OAAO,SAAS;AAAA,EACxC;AAAA,EAEQ,cAAoB;AAC1B,QAAI,CAAC,KAAK,OAAO;AACf,YAAM,IAAI,MAAM,oDAAoD;AAAA,IACtE;AAAA,EACF;AACF;","names":[]}

package/dist/chunk-KEWXLWAO.js

@@ -0,0 +1,247 @@
+// src/graph/query.ts
+function getNeighborhood(graph, nodeId, depth = 2) {
+  const visited = /* @__PURE__ */ new Set();
+  const queue = [{ id: nodeId, d: 0 }];
+  while (queue.length > 0) {
+    const { id, d } = queue.shift();
+    if (visited.has(id) || d > depth) continue;
+    visited.add(id);
+    for (const edge of graph.edges) {
+      if (edge.from === id && !visited.has(edge.to)) {
+        queue.push({ id: edge.to, d: d + 1 });
+      }
+      if (edge.to === id && !visited.has(edge.from)) {
+        queue.push({ id: edge.from, d: d + 1 });
+      }
+    }
+  }
+  return graph.nodes.filter((n) => visited.has(n.id));
+}
+function findPatterns(graph) {
+  const patterns = [];
+  const degree = {};
+  for (const edge of graph.edges) {
+    degree[edge.from] = (degree[edge.from] ?? 0) + 1;
+    degree[edge.to] = (degree[edge.to] ?? 0) + 1;
+  }
+  const hubs = Object.entries(degree).filter(([, d]) => d >= 3).sort((a, b) => b[1] - a[1]).slice(0, 5);
+  for (const [nodeId, d] of hubs) {
+    const node = graph.nodes.find((n) => n.id === nodeId);
+    if (node) {
+      patterns.push({
+        description: `"${node.label}" is a hub node (${d} connections) \u2014 many users/entities relate to it`,
+        nodes: [nodeId],
+        edges: graph.edges.filter((e) => e.from === nodeId || e.to === nodeId).map((e) => e.id),
+        frequency: d,
+        significance: Math.min(1, d / 10)
+      });
+    }
+  }
+  const churnEdges = graph.edges.filter((e) => e.type === "cancelled" || e.type === "churned_because");
+  if (churnEdges.length > 0) {
+    const churnReasons = churnEdges.map((e) => {
+      const target = graph.nodes.find((n) => n.id === e.to);
+      return target?.label;
+    }).filter(Boolean);
+    if (churnReasons.length > 0) {
+      patterns.push({
+        description: `Churn pattern: ${churnReasons.slice(0, 3).join(", ")}`,
+        nodes: churnEdges.map((e) => e.to),
+        edges: churnEdges.map((e) => e.id),
+        frequency: churnEdges.length,
+        significance: Math.min(1, churnEdges.length / 5)
+      });
+    }
+  }
+  const complaintNodes = graph.nodes.filter((n) => n.type === "complaint");
+  if (complaintNodes.length >= 2) {
+    patterns.push({
+      description: `${complaintNodes.length} complaints identified: ${complaintNodes.slice(0, 3).map((n) => `"${n.label}"`).join(", ")}`,
+      nodes: complaintNodes.map((n) => n.id),
+      edges: [],
+      frequency: complaintNodes.length,
+      significance: Math.min(1, complaintNodes.length / 5)
+    });
+  }
+  const eventNodes = graph.nodes.filter((n) => n.type === "event");
+  const topEvents = eventNodes.sort((a, b) => (Number(b.properties.totalOccurrences) ?? 0) - (Number(a.properties.totalOccurrences) ?? 0)).slice(0, 5);
+  if (topEvents.length >= 2) {
+    patterns.push({
+      description: `Top features: ${topEvents.map((n) => `${n.label} (${n.properties.totalOccurrences}x)`).join(", ")}`,
+      nodes: topEvents.map((n) => n.id),
+      edges: [],
+      frequency: topEvents.length,
+      significance: 0.7
+    });
+  }
+  return patterns.sort((a, b) => b.significance - a.significance);
+}
+function getPersonaContext(graph, traits) {
+  const context = [];
+  const metricNodes = graph.nodes.filter((n) => n.type === "metric");
+  for (const m of metricNodes) {
+    context.push(m.label);
+  }
+  if (traits.openness > 0.6) {
+    const featureNodes = graph.nodes.filter(
+      (n) => n.type === "feature" || n.type === "event"
+    );
+    for (const f of featureNodes.slice(0, 5)) {
+      const count = f.properties.totalOccurrences;
+      context.push(
+        count ? `Feature "${f.label}" is used ${count} times` : `Feature: ${f.label}`
+      );
+    }
+  }
+  if (traits.neuroticism > 0.6) {
+    const complaints = graph.nodes.filter((n) => n.type === "complaint");
+    for (const c of complaints.slice(0, 5)) {
+      context.push(`User complaint: "${c.label}"`);
+    }
+    const churnEdges = graph.edges.filter(
+      (e) => e.type === "cancelled" || e.type === "churned_because"
+    );
+    if (churnEdges.length > 0) {
+      context.push(`${churnEdges.length} users have cancelled`);
+    }
+  }
+  if (traits.conscientiousness > 0.6) {
+    const discovered = graph.edges.filter(
+      (e) => e.properties.discovered === true
+    );
+    for (const d of discovered.slice(0, 3)) {
+      const from = graph.nodes.find((n) => n.id === d.from);
+      const to = graph.nodes.find((n) => n.id === d.to);
+      if (from && to) {
+        context.push(`Pattern: "${from.label}" ${d.type} "${to.label}"`);
+      }
+    }
+  }
+  if (traits.agreeableness > 0.6) {
+    const hubEdges = Object.entries(
+      graph.edges.reduce(
+        (acc, e) => {
+          acc[e.to] = (acc[e.to] ?? 0) + 1;
+          return acc;
+        },
+        {}
+      )
+    ).sort((a, b) => b[1] - a[1]).slice(0, 3);
+    for (const [nodeId, count] of hubEdges) {
+      const node = graph.nodes.find((n) => n.id === nodeId);
+      if (node && count >= 2) {
+        context.push(`${count} users interact with "${node.label}"`);
+      }
+    }
+  }
+  if (traits.extraversion > 0.6) {
+    const socialEvents = graph.nodes.filter(
+      (n) => n.type === "event" && /share|invite|collab|team|social/i.test(n.label)
+    );
+    for (const s of socialEvents.slice(0, 3)) {
+      context.push(`Social feature: "${s.label}" (${s.properties.totalOccurrences ?? 0} uses)`);
+    }
+  }
+  return context;
+}
+function graphSummary(graph) {
+  const typeCounts = {};
+  for (const node of graph.nodes) {
+    typeCounts[node.type] = (typeCounts[node.type] ?? 0) + 1;
+  }
+  const edgeTypes = {};
+  for (const edge of graph.edges) {
+    edgeTypes[edge.type] = (edgeTypes[edge.type] ?? 0) + 1;
+  }
+  const lines = [
+    `Knowledge graph: ${graph.metadata.nodeCount} nodes, ${graph.metadata.edgeCount} edges`,
+    `Sources: ${graph.metadata.sources.join(", ")}`,
+    `Node types: ${Object.entries(typeCounts).map(([t, c]) => `${t}(${c})`).join(", ")}`,
+    `Edge types: ${Object.entries(edgeTypes).map(([t, c]) => `${t}(${c})`).join(", ")}`
+  ];
+  return lines.join("\n");
+}
+async function queryGraph(llm, graph, question) {
+  const graphContext = serializeGraphForLLM(graph);
+  const patterns = findPatterns(graph);
+  const patternText = patterns.length > 0 ? `
+Discovered patterns:
+${patterns.map((p) => `- ${p.description} (significance: ${p.significance.toFixed(2)})`).join("\n")}` : "";
+  const result = await llm.generateJSON(
+    `You are a data analyst answering questions about a product's user base using a knowledge graph. The graph contains real user data \u2014 users, events, features, complaints, payments, and their relationships.
+
+Ground every claim in specific graph data. If the graph doesn't contain enough information to answer confidently, say so and explain what data would help.`,
+    `Knowledge Graph:
+${graphContext}
+${patternText}
+
+Question: "${question}"
+
+Analyze the graph to answer this question. Return JSON:
+{
+  "answer": "<2-4 sentence answer grounded in the graph data>",
+  "evidence": [
+    {"type": "node|edge|pattern", "label": "<entity or relationship>", "detail": "<how this supports the answer>", "relevance": <0-1>}
+  ],
+  "confidence": <0.0-1.0 based on how much graph data supports the answer>,
+  "relatedNodes": ["<labels of the most relevant nodes for follow-up questions>"]
+}`
+  );
+  return {
+    question,
+    ...result
+  };
+}
+function serializeGraphForLLM(graph) {
+  const lines = [];
+  lines.push(`${graph.metadata.nodeCount} nodes, ${graph.metadata.edgeCount} edges`);
+  const byType = {};
+  for (const node of graph.nodes) {
+    if (!byType[node.type]) byType[node.type] = [];
+    byType[node.type].push(node);
+  }
+  for (const [type, nodes] of Object.entries(byType)) {
+    const limit = type === "user" ? 5 : 15;
+    const shown = nodes.slice(0, limit);
+    const remaining = nodes.length - shown.length;
+    lines.push(`
+[${type.toUpperCase()}] (${nodes.length} total)`);
+    for (const n of shown) {
+      const props = Object.entries(n.properties).filter(([k]) => !k.startsWith("_") && k !== "fullText").slice(0, 3).map(([k, v]) => `${k}=${JSON.stringify(v)}`).join(", ");
+      lines.push(`  - ${n.label}${props ? ` (${props})` : ""}`);
+    }
+    if (remaining > 0) {
+      lines.push(`  ... and ${remaining} more`);
+    }
+  }
+  const edgesByType = {};
+  for (const edge of graph.edges) {
+    if (!edgesByType[edge.type]) edgesByType[edge.type] = [];
+    const from = graph.nodes.find((n) => n.id === edge.from);
+    const to = graph.nodes.find((n) => n.id === edge.to);
+    if (from && to) {
+      edgesByType[edge.type].push({ from: from.label, to: to.label, weight: edge.weight });
+    }
+  }
+  lines.push("\n[RELATIONSHIPS]");
+  for (const [type, edges] of Object.entries(edgesByType)) {
+    const shown = edges.slice(0, 8);
+    lines.push(`  ${type} (${edges.length} total):`);
+    for (const e of shown) {
+      lines.push(`    "${e.from}" \u2192 "${e.to}" (weight: ${e.weight.toFixed(1)})`);
+    }
+    if (edges.length > shown.length) {
+      lines.push(`    ... and ${edges.length - shown.length} more`);
+    }
+  }
+  return lines.join("\n");
+}
+
+export {
+  getNeighborhood,
+  findPatterns,
+  getPersonaContext,
+  graphSummary,
+  queryGraph
+};
+//# sourceMappingURL=chunk-KEWXLWAO.js.map