@falai/agent 0.4.0 → 0.5.0

package/examples/business-onboarding.ts

@@ -1,13 +1,13 @@
 /**
  * Business Onboarding Example
+ * Updated for v2 architecture with session state management and schema-first data extraction
  *
  * Real-world example showing:
- * - Complex multi-step onboarding flow
- * - Tools with contextUpdate for automatic state management
- * - Lifecycle hooks for persistence
- * - Branching logic (physical vs online business)
- * - Both step-by-step and chained transition approaches
- * - Agent caching with lifecycle hooks
+ * - Complex multi-step onboarding flow with schema-based data extraction
+ * - Tools with enhanced context access for automatic state management
+ * - Lifecycle hooks for data validation and persistence
+ * - Branching logic (physical vs online business) with skipIf conditions
+ * - Code-based state progression instead of fuzzy conditions
  */
 
 import {
@@ -17,6 +17,7 @@ import {
   EventSource,
   createMessageEvent,
   OpenAIProvider,
+  createSession,
   type ToolContext,
 } from "../src/index";
 
@@ -690,10 +691,21 @@ async function main() {
 
   // Generate response (requires valid API key)
   try {
-    const response = await agent.respond({ history });
+    // Initialize session state for multi-turn conversation
+    let session = createSession<OnboardingData>();
+
+    const response = await agent.respond({ history, session });
     console.log("Agent:", response.message);
-    console.log("\nRoute:", response.route?.title);
-    console.log("State:", response.state?.description);
+    console.log("\nRoute:", response.session?.currentRoute?.title);
+    console.log("Extracted:", response.session?.extracted);
+
+    // Update session with progress
+    session = response.session!;
+
+    console.log("\n✅ Session state benefits:");
+    console.log("   - Data extraction tracked across turns");
+    console.log("   - State progression managed automatically");
+    console.log("   - Always-on routing respects intent changes");
   } catch (error: any) {
     console.log("\n(Skipping AI response - requires valid API key)");
     console.log("Error:", error.message);

package/examples/company-qna-agent.ts

@@ -0,0 +1,508 @@
+/**
+ * Example: Company Q&A Agent (Stateless, Knowledge-Based)
+ *
+ * This demonstrates:
+ * 1. Schema-first architecture for stateless Q&A routes (no gatherSchema)
+ * 2. Tools for context enrichment (not data extraction)
+ * 3. Session state management even for stateless conversations
+ * 4. Always-on routing with context awareness
+ * 5. Three-phase pipeline: PREPARATION → ROUTING → RESPONSE
+ */
+
+import {
+  Agent,
+  createSession,
+  EventSource,
+  createMessageEvent,
+  EventKind,
+} from "../src";
+import type { Event } from "../src/types";
+import type { ToolRef } from "../src/types/tool";
+
+// ==============================================================================
+// CONTEXT: Company Knowledge Base
+// ==============================================================================
+
+interface CompanyContext {
+  companyInfo: {
+    name: string;
+    founded: number;
+    employees: number;
+    headquarters: string;
+  };
+  products: Array<{
+    id: string;
+    name: string;
+    description: string;
+    price: string;
+    category: string;
+  }>;
+  policies: {
+    returnPolicy: string;
+    shippingPolicy: string;
+    warrantyPolicy: string;
+  };
+  faqs: Array<{
+    question: string;
+    answer: string;
+    category: string;
+  }>;
+  recentNews?: Array<{
+    title: string;
+    date: string;
+    summary: string;
+  }>;
+}
+
+// ==============================================================================
+// TOOLS: Context Enrichment (PREPARATION Phase)
+// ==============================================================================
+
+// Tool: Fetch latest company news (context enrichment)
+const fetchNewsTool: ToolRef<CompanyContext, [], void> = {
+  id: "fetch_news",
+  name: "Fetch Company News",
+  description: "Retrieve latest company news and updates",
+  handler: async (context) => {
+    // Simulate API call to news service
+    const news = [
+      {
+        title: "New Product Launch: Acme Widget Pro",
+        date: "2025-10-10",
+        summary: "We're excited to announce the Acme Widget Pro...",
+      },
+      {
+        title: "Company Expands to European Market",
+        date: "2025-10-01",
+        summary: "Acme Corp opens new offices in London and Berlin...",
+      },
+    ];
+
+    console.log(`[Tool] Fetched ${news.length} news articles`);
+
+    return {
+      data: undefined,
+      contextUpdate: {
+        recentNews: news,
+      },
+    };
+  },
+};
+
+// Tool: Search knowledge base (context enrichment)
+const searchKnowledgeTool: ToolRef<CompanyContext, [], string> = {
+  id: "search_knowledge",
+  name: "Search Knowledge Base",
+  description: "Search FAQs and documentation",
+  handler: async (context) => {
+    const { history } = context;
+
+    // Get last user message
+    const lastMessage = history
+      .filter(
+        (e) => e.kind === EventKind.MESSAGE && e.source === EventSource.CUSTOMER
+      )
+      .pop();
+
+    if (!lastMessage) {
+      return { data: "No query found" };
+    }
+
+    const query = (lastMessage.data as any).message.toLowerCase();
+
+    // Simple keyword matching (in real app, use vector search)
+    const relevantFaqs = context.context.faqs.filter(
+      (faq) =>
+        faq.question.toLowerCase().includes(query) ||
+        faq.answer.toLowerCase().includes(query)
+    );
+
+    console.log(`[Tool] Found ${relevantFaqs.length} relevant FAQs`);
+
+    return {
+      data: relevantFaqs
+        .map((faq) => `Q: ${faq.question}\nA: ${faq.answer}`)
+        .join("\n\n"),
+    };
+  },
+};
+
+// ==============================================================================
+// AGENT SETUP
+// ==============================================================================
+
+const agent = new Agent<CompanyContext>({
+  name: "Acme Support Agent",
+  goal: "Answer questions about Acme Corp and our products",
+  description:
+    "I'm here to help you learn about Acme Corp, our products, and policies",
+  personality:
+    "Friendly, helpful, and knowledgeable. Always professional but approachable.",
+  ai: null as any, // Replace with actual AI provider
+
+  // Initialize with company knowledge
+  context: {
+    companyInfo: {
+      name: "Acme Corporation",
+      founded: 2010,
+      employees: 500,
+      headquarters: "San Francisco, CA",
+    },
+    products: [
+      {
+        id: "widget-1",
+        name: "Acme Widget",
+        description: "Our flagship product for all your widget needs",
+        price: "$99.99",
+        category: "widgets",
+      },
+      {
+        id: "gadget-1",
+        name: "Acme Gadget",
+        description: "The revolutionary gadget that changed everything",
+        price: "$149.99",
+        category: "gadgets",
+      },
+    ],
+    policies: {
+      returnPolicy: "30-day money-back guarantee, no questions asked",
+      shippingPolicy:
+        "Free shipping on orders over $50. 2-5 business days delivery",
+      warrantyPolicy: "1-year warranty on all products",
+    },
+    faqs: [
+      {
+        question: "What is your return policy?",
+        answer: "We offer a 30-day money-back guarantee on all products.",
+        category: "returns",
+      },
+      {
+        question: "Do you ship internationally?",
+        answer: "Yes, we ship to over 50 countries worldwide.",
+        category: "shipping",
+      },
+      {
+        question: "How long is the warranty?",
+        answer: "All products come with a 1-year manufacturer warranty.",
+        category: "warranty",
+      },
+    ],
+  },
+
+  // Add domain terms for better understanding
+  terms: [
+    {
+      name: "Widget",
+      description: "Our core product line for task automation",
+    },
+    {
+      name: "Gadget",
+      description: "Advanced tools for power users",
+    },
+  ],
+
+  // General guidelines (no tools attached - just behavioral)
+  guidelines: [
+    {
+      action: "Always be polite and professional",
+      enabled: true,
+    },
+    {
+      action:
+        "If you don't know the answer, admit it and offer to connect them with a human",
+      enabled: true,
+    },
+    {
+      action: "Provide specific, accurate information from the knowledge base",
+      enabled: true,
+    },
+  ],
+});
+
+// ==============================================================================
+// ROUTES: STATELESS Q&A ROUTES (Schema-First Architecture)
+// ==============================================================================
+
+// Route 1: Company Information (stateless - no data extraction)
+const companyInfoRoute = agent.createRoute({
+  title: "Company Information",
+  description: "Answer general questions about Acme Corp",
+  conditions: [
+    "User asks about the company",
+    "Questions about company history, size, location",
+    "When was the company founded",
+    "How many employees",
+    "Where is the headquarters",
+  ],
+  // NO gatherSchema - stateless Q&A route
+  // Just use initial state with chatState for response generation
+});
+
+// Initial state: Answer from knowledge base (no data gathering needed)
+
+// Route 2: Product Information (stateless)
+const productInfoRoute = agent.createRoute({
+  title: "Product Information",
+  description: "Answer questions about products",
+  conditions: [
+    "User asks about products",
+    "Questions about features, pricing, availability",
+    "What products do you offer",
+    "Tell me about your widgets",
+  ],
+  // NO gatherSchema - just answering questions
+});
+
+// Initial state is enough - no transitions needed for simple Q&A
+
+// Route 3: Policy Questions (stateless)
+const policyRoute = agent.createRoute({
+  title: "Policy Information",
+  description: "Answer questions about company policies",
+  conditions: [
+    "User asks about policies",
+    "Return policy",
+    "Shipping information",
+    "Warranty questions",
+  ],
+  // NO gatherSchema
+});
+
+// Initial state is enough - no extra setup needed
+
+// Route 4: News & Updates (uses tool, but still stateless)
+const newsRoute = agent.createRoute({
+  title: "Company News",
+  description: "Share latest company news and updates",
+  conditions: [
+    "User asks about news",
+    "What's new",
+    "Recent updates",
+    "Latest announcements",
+  ],
+});
+
+// Add tool to initial state to fetch news
+const fetchNews = newsRoute.initialState.transitionTo({
+  toolState: fetchNewsTool,
+});
+
+const shareNews = fetchNews.transitionTo({
+  chatState: "Share the latest company news from context",
+});
+
+// Route 5: General FAQ Search (uses tool)
+const faqRoute = agent.createRoute({
+  title: "FAQ Search",
+  description: "Search FAQs for relevant answers",
+  conditions: [
+    "User has a question that might be in FAQs",
+    "How do I...",
+    "Can I...",
+    "Is there...",
+  ],
+});
+
+const searchFaqs = faqRoute.initialState.transitionTo({
+  toolState: searchKnowledgeTool,
+});
+
+const provideFaqAnswer = searchFaqs.transitionTo({
+  chatState: "Provide answer based on FAQ search results",
+});
+
+// Route 6: Fallback (generic response)
+const fallbackRoute = agent.createRoute({
+  title: "General Conversation",
+  description: "Handle general conversation or unclear questions",
+  conditions: [
+    "User message doesn't match other routes",
+    "Greetings",
+    "Small talk",
+    "Unclear intent",
+  ],
+});
+
+// Initial state is enough for fallback conversations
+
+// ==============================================================================
+// USAGE EXAMPLES: Three-Phase Pipeline Demonstration
+// ==============================================================================
+
+async function exampleConversations() {
+  let session = createSession();
+
+  // =========================================================================
+  // Example 1: Simple company info question (stateless)
+  // =========================================================================
+  console.log("\n=== EXAMPLE 1: Company Info (Stateless Q&A) ===");
+  const history1: Event[] = [
+    createMessageEvent(
+      EventSource.CUSTOMER,
+      "User",
+      "How many employees does Acme have?"
+    ),
+  ];
+
+  const response1 = await agent.respond({ history: history1, session });
+  console.log("AI:", response1.message);
+  console.log("Route:", response1.session?.currentRoute?.title);
+
+  /*
+   * ARCHITECTURE FLOW:
+   * 1. PREPARATION: No tools needed for simple Q&A
+   * 2. ROUTING: Framework routes to "Company Information" (score: 95)
+   * 3. RESPONSE: AI answers from context knowledge
+   *    - Route: "Company Information"
+   *    - Session: Updated with route/state (even for stateless)
+   *    - No data extraction (stateless route)
+   */
+
+  // =========================================================================
+  // Example 2: Product question (stateless)
+  // =========================================================================
+  console.log("\n=== EXAMPLE 2: Product Info ===");
+  const history2: Event[] = [
+    createMessageEvent(
+      EventSource.CUSTOMER,
+      "User",
+      "What products do you offer?"
+    ),
+  ];
+
+  const response2 = await agent.respond({ history: history2, session });
+  console.log("AI:", response2.message);
+  console.log("Route:", response2.session?.currentRoute?.title);
+  // Expected: "We offer two main products: Acme Widget ($99.99)..."
+  // Route: "Product Information"
+
+  // =========================================================================
+  // Example 3: Policy question (stateless)
+  // =========================================================================
+  console.log("\n=== EXAMPLE 3: Policy Question ===");
+  const history3: Event[] = [
+    createMessageEvent(
+      EventSource.CUSTOMER,
+      "User",
+      "What's your return policy?"
+    ),
+  ];
+
+  const response3 = await agent.respond({ history: history3, session });
+  console.log("AI:", response3.message);
+  console.log("Route:", response3.session?.currentRoute?.title);
+  // Expected: "We offer a 30-day money-back guarantee..."
+  // Route: "Policy Information"
+
+  // =========================================================================
+  // Example 4: News request (tool execution, but still stateless)
+  // =========================================================================
+  console.log("\n=== EXAMPLE 4: Latest News ===");
+  const history4: Event[] = [
+    createMessageEvent(EventSource.CUSTOMER, "User", "What's new at Acme?"),
+  ];
+
+  const response4 = await agent.respond({ history: history4, session });
+  console.log("AI:", response4.message);
+  console.log("Route:", response4.session?.currentRoute?.title);
+  // Tool fetches news → Updates context → AI responds with news
+  // Route: "Company News"
+  // NO data extraction - tool just enriches context
+
+  // =========================================================================
+  // Example 5: Multi-turn conversation (context maintained)
+  // =========================================================================
+  console.log("\n=== EXAMPLE 5: Multi-turn ===");
+
+  // Turn 1
+  const turn1: Event[] = [
+    createMessageEvent(
+      EventSource.CUSTOMER,
+      "User",
+      "Tell me about the Acme Widget"
+    ),
+  ];
+  const resp1 = await agent.respond({ history: turn1, session });
+  console.log("User: Tell me about the Acme Widget");
+  console.log("AI:", resp1.message);
+
+  // Turn 2 - follow-up question
+  const turn2: Event[] = [
+    ...turn1,
+    createMessageEvent(EventSource.AI_AGENT, "Agent", resp1.message),
+    createMessageEvent(EventSource.CUSTOMER, "User", "How much does it cost?"),
+  ];
+  const resp2 = await agent.respond({
+    history: turn2,
+    session: resp1.session, // Pass previous session
+  });
+  console.log("User: How much does it cost?");
+  console.log("AI:", resp2.message);
+  // AI understands "it" refers to Acme Widget from context
+}
+
+// ==============================================================================
+// KEY TAKEAWAYS FOR Q&A USE CASES
+// ==============================================================================
+
+/*
+ * 1. NO STATE MACHINES REQUIRED
+ *    - Just use the initial state with a chatState description
+ *    - No gather, no skipIf, no state transitions
+ *    - Perfect for stateless question-answering
+ *
+ * 2. ROUTING STILL WORKS
+ *    - Framework routes to most relevant Q&A route
+ *    - Always-on routing handles topic switches
+ *    - Conditions guide the AI to the right knowledge domain
+ *
+ * 3. TOOLS FOR CONTEXT ENRICHMENT
+ *    - Tools fetch additional data (news, search results)
+ *    - Updates context (not extracted data)
+ *    - AI uses enriched context to answer
+ *
+ * 4. KNOWLEDGE IN CONTEXT
+ *    - Store company knowledge in context
+ *    - Use contextProvider for always-fresh data
+ *    - Tools can augment context at runtime
+ *
+ * 5. MIX STATEFUL & STATELESS
+ *    - Have Q&A routes (stateless)
+ *    - AND booking/onboarding routes (stateful)
+ *    - Framework handles both seamlessly
+ *
+ * 6. GUIDELINES FOR BEHAVIOR
+ *    - Use guidelines for general behavioral rules
+ *    - No need for complex state machines for simple policies
+ *    - Keep it simple!
+ *
+ * ARCHITECTURE FOR Q&A:
+ *
+ *   User Question
+ *       ↓
+ *   Routing (score all Q&A routes)
+ *       ↓
+ *   Tool Execution (if needed - fetch data)
+ *       ↓
+ *   Response (AI answers from context)
+ *       ↓
+ *   Done (no state tracking needed)
+ *
+ * vs. STATEFUL FLOWS (booking, onboarding):
+ *
+ *   User Intent
+ *       ↓
+ *   Routing
+ *       ↓
+ *   State Machine (gather data step-by-step)
+ *       ↓
+ *   Tools (validate/enrich extracted data)
+ *       ↓
+ *   Response (continue conversation)
+ *       ↓
+ *   Track Session (extracted data, current state)
+ */
+
+if (require.main === module) {
+  exampleConversations().catch(console.error);
+}