Mira, Aurora & Multimodal AI
Relevant source files
The following files were used as context for generating this wiki page:
Mira and Aurora represent the core of the Tuturuuu AI ecosystem, designed to transform the platform from a collection of passive tools into a proactive, intelligent operating system. Mira serves as the empathetic conversational interface—a "JARVIS" for modern life—while Aurora functions as the contextual engine or "nervous system" that interconnects various workspace data points.
Sources: README.md:105-110, apps/web/src/app/[locale]/(marketing)/careers/page.tsx:455-475, apps/meet/messages/en.json:115-135
The AI Ecosystem Architecture
The ecosystem is structured into specialized agents and engines that collaborate to automate administrative work and eliminate digital friction.
Core AI Components
| Component | Role | Description |
|---|---|---|
| Mira | Soul & Voice | A proactive AI companion that plans, reasons, and acts on the user's behalf. |
| Aurora | Nervous System | A context graph engine linking emails, tasks, files, and events into a unified knowledge graph. |
| Crystal | Bridge to Humanity | A multimodal interface enabling real-time collaboration via voice, video, and screen sharing. |
| Rewise | Collective Mind | An aggregator of leading AI models (OpenAI, Gemini, Anthropic) for best-in-class intelligence. |
| Nova | Conscience & Forge | A prompt-engineering and alignment platform ensuring AI reasoning is safe and effective. |
Sources: apps/meet/messages/en.json:115-135, apps/web/src/app/[locale]/(marketing)/careers/page.tsx:455-495, README.md:105-112
Intelligence Architecture Flow
The following diagram illustrates how user interactions flow through the aggregator (Rewise) and leverage the contextual intelligence of Aurora to provide Mira's conversational responses.
This architecture ensures that Mira is not just a chatbot but a context-aware assistant capable of cross-app integration.
Sources: apps/web/src/app/[locale]/(marketing)/careers/page.tsx:455-495, README.md:105-112
Mira: Proactive Conversational AI
Mira is defined as an internal AI product operating on the Tuturuuu platform, customized and engineered by the founder, Võ Hoàng Phúc. Mira follows a strict set of behavioral guidelines to ensure a professional and helpful user experience.
System Instructions and Guidelines
Mira's logic is governed by a comprehensive systemInstruction string used across different AI provider routes. Key mandates include:
- Tone: Always polite, respectful, and professional.
- Language Localization: Mira must respond in the same language used by the user in the most recent messages.
- Content Structure: Responses are formatted using Markdown, often structured as essays or blog posts to enhance readability.
- Proactive Engagement: Mira provides three helpful follow-up prompts at the end of every response, predicting what the user might ask next.
Sources: packages/ai/src/chat/google-vertex/route.ts:155-185, packages/ai/src/chat/google/route.ts:310-340
Interactive Learning Features
Mira supports specialized interactive components through custom tags that the frontend renders as UI elements:
- Quizzes: Enclosed in
@<QUIZ>and</QUIZ>tags, including<QUESTION>and<OPTION isCorrect>components. - Flashcards: Enclosed in
@<FLASHCARD>and</FLASHCARD>tags, containing<QUESTION>and<ANSWER>parts. - Follow-ups: Predictive prompts enclosed in
@<FOLLOWUP>tags to maintain conversational flow.
Sources: packages/ai/src/chat/google-vertex/route.ts:186-215, packages/ai/src/chat/anthropic/route.ts:180-210
Aurora: The Context Graph Engine
Aurora acts as the intelligence layer that transforms flat data into a "nervous system." It connects disparate entities within a workspace to provide Mira with the context needed for proactive assistance.
Knowledge Synthesis
Aurora's primary function is to link:
- Communications: Emails and chat messages.
- Planning: Tasks, initiatives, and calendar events.
- Resources: Files and documents.
This creates a "Data Moat" and enables "Contextual Intelligence," allowing Mira to answer complex queries like "What's on my agenda tomorrow?" by looking across all integrated Tuturuuu applications.
Sources: apps/web/src/app/[locale]/(marketing)/careers/page.tsx:465-475, apps/meet/messages/en.json:115-135
Multimodal AI & File Processing
The platform supports multimodal inputs, allowing users to upload files and images directly into conversations with Mira.
File Handling Logic
The system includes logic to retrieve and process files associated with a chat thread. The getAllChatFiles function fetches resources from Supabase storage, while processMessagesWithFiles attaches this content to the model messages.
Supported Media Types
The implementation distinguishes between different media types to ensure they are passed to the AI models correctly:
- Images: Processes types starting with
image/asImagePart. - Text/JSON: Reads content as strings and encodes them into
FilePart. - Binary (PDFs, etc.): Converts
ArrayBufferdata intoUint8ArrayforFilePart.
Sources: packages/ai/src/chat/google/route.ts:50-80, packages/ai/src/chat/google/route.ts:135-165
Technical Implementation Details
The AI routes are implemented using the Vercel AI SDK and are hosted on the edge for performance.
Request Handling Flow
Each AI route (Google, Anthropic, OpenAI) follows a similar execution pattern:
- Authentication: Verifies the user session via Supabase.
- Chat Resolution: Identifies the
chatIdor retrieves the most recent chat for the user. - Message Conversion: Transforms
UIMessagearrays intoModelMessageformats. - Validation: Checks message content length against
MAX_CHAT_MESSAGE_LENGTH. - Credit Management: (Specifically in the Google route) Checks and deducts AI credits based on token usage.
- Streaming: Uses
streamTextwith thesmoothStreamexperimental transform for a fluid UI experience.
Sources: packages/ai/src/chat/google-vertex/route.ts:25-75, packages/ai/src/chat/google/route.ts:180-250, packages/ai/src/chat/anthropic/route.ts:25-85
AI Credit and Token Management
The platform enforces a usage-based economy for AI features. The system pre-flights requests to ensure sufficient credits exist and caps maxOutputTokens to prevent over-expenditure. On completion, the usage (input, output, and reasoning tokens) is recorded and credits are deducted from the workspace.
Sources: packages/ai/src/chat/google/route.ts:285-305, packages/ai/src/chat/google/route.ts:345-385
Summary
Mira and Aurora form a cohesive AI ecosystem within Tuturuuu. Mira provides the human-centric interface, following structured guidelines to provide interactive and helpful responses. Aurora powers this interface by serving as a context engine that unifies workspace data. Supported by a multimodal infrastructure that handles complex file types and a robust multi-model routing system (Rewise), the platform provides an integrated "Age of Partners" experience where AI acts as a proactive collaborator.