Vercel AI Chatbot

Routes chat requests through Vercel AI Gateway to multiple LLM providers (OpenAI, Anthropic, Google, etc.) with automatic provider selection and fallback logic. Implements server-side streaming via Next.js API routes that pipe model responses directly to the client using ReadableStream, enabling real-time token-by-token display without buffering entire responses. The /api/chat route integrates @ai-sdk/gateway for provider abstraction and @ai-sdk/react's useChat hook for client-side stream consumption.

Stores all chat messages, conversations, and metadata in PostgreSQL using Drizzle ORM for type-safe queries. The data layer (lib/db/queries.ts) provides functions like saveMessage(), getChatById(), and deleteChat() that handle CRUD operations with automatic timestamp tracking and user association. Messages are persisted after each API call, enabling chat resumption across sessions and browser refreshes without losing context.

Displays a sidebar with the user's chat history, organized by recency or custom folders. The sidebar includes search functionality to filter chats by title or content, and quick actions to delete, rename, or archive chats. Chat list is fetched from PostgreSQL via getChatsByUserId() and cached in React state with optimistic updates. The sidebar is responsive and collapses on mobile via a toggle button.

Implements light/dark theme switching via Tailwind CSS dark mode class toggling and React Context for theme state persistence. The root layout (app/layout.tsx) provides a ThemeProvider that reads the user's preference from localStorage or system settings, and applies the 'dark' class to the HTML element. All UI components use Tailwind's dark: prefix for dark mode styles, and the theme toggle button updates the context and localStorage.

Provides inline actions on each message: copy to clipboard, regenerate AI response, delete message, or vote. These actions are implemented as buttons in the Message component that trigger API calls or client-side functions. Regenerate calls the /api/chat route with the same context but excluding the message being regenerated, forcing the model to produce a new response. Delete removes the message from the database and UI optimistically.

Implements dual authentication paths using NextAuth 5.0 with OAuth providers (GitHub, Google) and email/password registration. Guest users get temporary session tokens without account creation; registered users have persistent identities tied to PostgreSQL user records. Authentication middleware (middleware.ts) protects routes and injects userId into request context, enabling per-user chat isolation and rate limiting. Session state flows through next-auth/react hooks (useSession) to UI components.

Implements schema-based function calling where the AI model can invoke predefined tools (getWeather, createDocument, getSuggestions) by returning structured tool_use messages. The chat route parses tool calls, executes corresponding handler functions, and appends results back to the message stream. Tools are defined in lib/ai/tools.ts with JSON schemas that the model understands, enabling multi-turn conversations where the AI can fetch real-time data or trigger side effects without user intervention.

Stores in-flight streaming responses in Redis with a TTL, enabling clients to resume incomplete message streams if the connection drops. When a stream is interrupted, the client sends the last received token offset, and the server retrieves the cached stream from Redis and resumes from that point. This is implemented in the /api/chat route using redis.get/set with keys like 'stream:{chatId}:{messageId}' and automatic cleanup via TTL expiration.

Accepts file uploads (images, documents) via the MultimodalInput component, stores them in Vercel Blob storage, and passes file references to the AI model as part of the message context. The /api/upload route handles multipart form data, generates signed URLs, and returns file metadata. Vision-capable models (GPT-4V, Claude 3) can then analyze images or extract text from documents, with results embedded in the conversation stream.

Allows the AI to create and edit code artifacts, documents, or structured content that are stored separately from chat messages. Artifacts are managed via the Document and Artifact System with versioning support — each edit creates a new version with a timestamp and diff tracking. The UI renders artifacts in a dedicated panel with syntax highlighting (for code) or rich text editing, and users can fork, copy, or export artifacts. Artifact state is managed via React Context and persisted to PostgreSQL.

Enforces per-user rate limits on API calls (chat, file uploads, tool execution) based on user tier (free, pro, enterprise). Rate limit state is tracked in PostgreSQL with counters for requests per minute/hour/day, and middleware checks limits before processing requests. Entitlements determine feature access (e.g., free users get 10 chats/day, pro users unlimited). Exceeded limits return 429 Too Many Requests with retry-after headers.

Allows users to mark chats as private (only visible to owner) or public (shareable via URL). Public chats generate a unique share token that can be embedded in URLs, and unauthenticated users can view public chats in read-only mode. Visibility is stored in the chats table with a visibility enum (private/public) and optional share_token. The chat route checks visibility before returning messages, enforcing access control at the API level.

Allows users to upvote/downvote individual messages (AI responses or user messages) to collect feedback on response quality. Vote data is stored in a votes table with message_id, user_id, and vote_type (up/down), and the /api/votes route handles vote creation/updates. Vote counts are aggregated and displayed alongside messages in the UI. This feedback is used for model evaluation and fine-tuning.