Open Deep Research

If you intend to use TogetherAI's Deepseek model, you must install the corresponding AI SDK provider dependency.

pnpm add @ai-sdk/togetherai

To run the project on your local machine, follow these steps in order:

1. Install Vercel CLI: npm i -g vercel
2. Link project: Run vercel link to link your local instance with your Vercel and GitHub accounts.
3. Pull environment variables: Run vercel env pull to download your environment variables from Vercel.
4. Install dependencies: Use pnpm install.
5. Run migrations: Execute pnpm db:migrate to set up your database.
6. Start development server: Run pnpm dev to start the app at http://localhost:3000/.

The application uses a specific model for reasoning tasks (research analysis, structured outputs, etc.) via the REASONING_MODEL environment variable. If no model is specified, it defaults to o1-mini. If an invalid model is provided, it falls back to o1-mini.

Supported Models

• OpenAI: gpt-4o, o1, o3-mini (These have native JSON schema support).
• TogetherAI: deepseek-ai/DeepSeek-R1 (Requires BYPASS_JSON_VALIDATION=true).

Configuration via .env

To use a non-OpenAI model like DeepSeek, you must also set BYPASS_JSON_VALIDATION=true to allow the model to function without native JSON schema support.

# Example for TogetherAI DeepSeek
REASONING_MODEL=deepseek-ai/DeepSeek-R1
BYPASS_JSON_VALIDATION=true

The default function timeout is 300 seconds (5 minutes). If you are deploying on Vercel's Hobby tier, you must reduce this to 60 seconds to avoid errors. Adjust this using the MAX_DURATION environment variable in your .env file.

MAX_DURATION=60

The project uses Drizzle ORM with a PostgreSQL dialect. The configuration is managed via drizzle.config.ts and relies on environment variables defined in .env.local.

To configure the database connection, ensure the POSTGRES_URL environment variable is set in your .env.local file. The schema is located at ./lib/db/schema.ts and migrations are output to ./lib/db/migrations.

The AI engine uses a Model interface to define available models. You can extend or modify the models array to include different LLMs for standard tasks. Each model requires an id, label, apiIdentifier, and description.

export interface Model {
  id: string;
  label: string;
  apiIdentifier: string;
  description: string;
}

export const models: Array<Model> = [
  {
    id: 'gpt-4o',
    label: 'GPT 4o',
    apiIdentifier: 'gpt-4o',
    description: 'For complex, multi-step tasks',
  },
  // ...
]

The following functions allow for chat lifecycle management:

• saveChat({ id, userId, title }): Creates a new chat record.
• getChatsByUserId({ id }): Retrieves all chats for a specific user, ordered by creation date descending.
• getChatById({ id }): Retrieves a single chat record by its ID.
• deleteChatById({ id }): Deletes a chat and all associated votes and messages.
• updateChatVisiblityById({ chatId, visibility }): Updates a chat's visibility to either 'private' or 'public'.

The system uses documents and suggestions for research data:

• saveDocument({ id, title, kind, content, userId }): Saves a document. kind must be a valid BlockKind.
• getDocumentsById({ id }): Retrieves all documents matching an ID, ordered by creation date ascending.
• getDocumentById({ id }): Retrieves the most recent document for a given ID.
• saveSuggestions({ suggestions }): Bulk inserts an array of Suggestion objects.
• getSuggestionsByDocumentId({ documentId }): Retrieves all suggestions associated with a specific document ID.
• deleteDocumentsByIdAfterTimestamp({ id, timestamp }): Deletes documents and their associated suggestions that were created after the provided timestamp.