Changelog

Agentic Lead Capture: The AI can now autonomously detect buying intent mid-conversation and trigger a lead capture action — collecting name, email, and phone naturally without a hard-coded form.

Fixed a critical build failure caused by a missing TypeScript interface property (agentic_lead_capture_enabled) on the bot customization page.

Resolved an AttributeError in the backend semantic caching layer (CachedBot.system_prompt and CachedWidgetSettings) that crashed session initialization for returning visitors.

Added missing page_url and visitor_id attributes to ChatMessageRequest validation schemas, fixing 422 Unprocessable Entity errors during background database syncing.

Launched the full Next.js Admin Dashboard for bot configuration, analytics, conversation inbox, and lead management.

Stripe webhook integration for subscription management — Free and Pro plan enforcement, SubscriptionStatus tracking, and automated message quota management.

API rate limiting via slowapi and per-organization monthly message quota enforcement with automatic resets at the start of each billing cycle.

Dynamic loader.js script: the embed snippet now only references the Next.js frontend URL. The backend API is never exposed in client websites, making infrastructure migrations completely transparent to embedded widgets.

Streaming server-sent events (SSE) for real-time typewriter-effect responses from the LLM — no polling, no full-page reloads.

Full visual widget customization: primary colors, background colors, border radius, avatar image, widget position, and starter questions — all configurable from the dashboard.

Full RAG (Retrieval-Augmented Generation) pipeline: document chunking, vector embedding, and semantic similarity search against per-organization isolated knowledge bases.

Four knowledge source types supported: plain text, FAQs (structured Q&A), file uploads (PDF/Word), and automated website crawling with scheduled re-crawls.

Background workers powered by arq + Redis for async embedding tasks and scheduled web crawl jobs — knowledge indexing never blocks the main API.

Semantic caching layer: frequently asked questions are cached at the vector level, returning instant responses for similar queries without hitting the LLM.

Initialized FastAPI backend and Next.js 15 (App Router) frontend.

PostgreSQL database schema with Alembic migrations. Core models: Organization, User, Bot, Conversation, Message, KnowledgeSource.

Multi-tenant architecture: isolated knowledge bases, bot settings, and usage limits per organization from day one.

JWT-based authentication with hashed passwords (passlib). Organization-level team membership and role management.