Avrana is not just another chatbot. It is a configurable, modular AI platform designed to be your augmented.self — a tool that molds itself to your specific needs, your knowledge domain, your preferred language, and your infrastructure choices. Whether you want a fully offline voice assistant, a document-grounded Q&A system for your company, or a multilingual knowledge companion, Avrana provides the foundation.

What Makes Avrana Different?

Specialized Understanding Modules (Not Just a Raw LLM)

Most AI assistants send your question directly to a large language model and hope for the best. Avrana takes a smarter approach: your query is first routed to a specialized Understanding module — a microservice that owns a specific domain of knowledge.

Each Understanding enriches your question with context, relevant data, and domain-specific instructions before it ever reaches the LLM. The result is a response that is more accurate, more relevant, and grounded in the information that matters to you.

Currently available Understanding modules include:

• Polymath — A general-purpose knowledge system covering geography, history, science, culture, society, and economics. Built on a broad encyclopedic database, it can answer factual questions, explain processes, identify locations and notable figures, and clarify concepts across disciplines.
• Librarian — A custom document Retrieval-Augmented Generation (RAG) system. Point it at your own document collection, and it will search, retrieve, and synthesize answers from your private knowledge base. Ideal for corporate knowledge management, research archives, and
  internal documentation.

And the architecture is fully extensible — you can create your own
Understanding module for any domain, and Avrana’s Description Router will automatically learn when to use it.

True Multi-Modal Interaction

Avrana speaks, listens, and reads.

Mode	Input	Output
Voice	Speak naturally; Avrana transcribes your speech in real time	Avrana responds with synthesized speech, in your chosen voice
Text	Type your question in the web UI	Avrana replies with formatted text, including Markdown rendering
Both	Seamlessly switch between voice and text mid-conversation	Both text and speech are delivered simultaneously

The audio-visualizer in the web UI gives real-time feedback, howing when Avrana is listening, thinking, and speaking.

Offline-First Philosophy

Avrana is designed to run without internet access. Every component in the pipeline — speech recognition, language understanding, text generation, and speech synthesis — can be replaced with a local model. This means:

• Zero API costs after initial setup
• Complete data privacy — everything stays on your hardware
• No dependency on external services — your AI works even when the
  internet is down
• Full control over model selection, performance tuning, and upgrades

Prefer cloud accuracy for some tasks and local privacy for others? Avrana’s configuration lets you mix and match — for example, use a local LLM with a cloud ASR service, or vice versa.

The Pipeline: From Voice to Answer

Avrana processes your input through a carefully orchestrated pipeline:

Every step is configurable, replaceable, and optional. You decide which capabilities you need.

What Can You Build With Avrana?

Corporate Knowledge Assistant

Feed Avrana your company’s documentation, policies, and procedures through the Librarian module. Employees can ask questions in natural language and get instant, cited answers sourced from your internal documents.

Example: “What is our travel reimbursement policy for international flights?” → Librarian finds the relevant policy document → LLM summarizes it in a clear, conversational answer.

Educational & Research Companion

With Polymath’s broad knowledge base, Avrana functions as a tireless tutor and research assistant. It can explain complex concepts, provide historical context, clarify scientific principles, and help with language-related tasks.

Domain-Specific Expert Systems

Create custom Understanding modules for your industry — healthcare, legal, finance, logistics, or any specialized field. The modular architecture lets you package domain expertise into independently deployable services that Avrana’s router automatically selects when relevant.

Multilingual Voice Assistant

Avrana’s interface supports English, German, and Hungarian out of the box, with an architecture that makes adding new languages straightforward. Combined with multilingual ASR and LLM models, you can build a voice assistant that serves users in their native language.

Fully Offline Private Assistant

For privacy-conscious environments, classified settings, or remote locations without reliable internet, Avrana can run entirely on local hardware with open-source models. No data ever leaves your network.

Customization Options

Avrana is built to be yours. Here is what you can tailor:

Pipeline Components

Component	What You Can Choose
ASR (Speech-to-Text)	Cloud: Whisper API, Azure Whisper
LLM (Text Generation)	Any OpenAI-compatible endpoint: GPT-4, Azure OpenAI, Ollama (Gemma, DeepSeek, Llama, etc.), OpenRouter
TTS (Text-to-Speech)	Google TTS, OpenAI TTS, Piper TTS (local, via Wyoming protocol)
Embeddings	Any SentenceTransformer-compatible model for Understanding services
Rerankers	Cross-encoder models for routing and search result refinement

Understanding Services

You can enable, disable, and configure Understanding modules independently. Each module has its own:

• Domain description — a plain-text file that tells the Description Router
  what kind of queries it handles best
• System prompt — custom instructions for the LLM when responding in that
  domain
• Search configuration — embedding model, reranker, vector database
  settings, number of results
• Document sources — for RAG-based modules like Librarian, simply point
  to a folder of documents

User Interface

The web UI is built with HTML, CSS, and JavaScript and can be customized extensively:

• Themes — Light and dark mode with CSS custom properties. The theme
  toggle persists user preference and respects system settings.
• Localization — Full UI translation in en, hu, and de. Add new
  languages by creating a JSON translation file.
• Feature toggles — Show or hide the recording button, enable or disable
  the comment popup, host under a URL sub-path.
• Visualizer — A real-time audio visualization that makes the interaction
  feel alive.

Behavior & Personality

Avrana’s behavior is shaped by system prompts — plain Markdown files that you can edit without touching any code. You can customize:

• The main assistant persona and response style
• Formal vs. informal communication modes
• Query rewriting instructions
• Per-Understanding behavior guidelines

These prompts are loaded at startup and can be overridden per deployment
environment by placing custom versions in a config/ directory.

Conversation Style

Avrana supports formal and informal communication styles, configurable per session. The UI allows users to switch between styles, and the system prompts adapt accordingly.

Language & Localization

Avrana takes language support seriously. The platform offers:

• UI translations in English, German, and Hungarian
• ASR language hints — guide transcription to the target language for
  better accuracy
• TTS language detection — automatically selects the appropriate voice
  based on the response language
• Understanding modules in multiple languages — Polymath’s embedding
  models (e.g., multilingual-e5-large) are designed for cross-lingual
  understanding
• Message catalog — server-side messages (status notifications, error
  messages) are localized via gettext and can be translated for any locale

Architecture at a Glance

The core communicates with Understanding modules via gRPC — a high-performance, streaming-friendly protocol. This means:

• Each Understanding can be developed, deployed, and scaled independently
• Adding a new domain doesn’t require modifying the core
• Streaming responses enable real-time status updates (“Searching
  documents…”, “Generating answer…”)

Who Is Avrana For?

• Developers & system integrators who want a customizable AI platform they can adapt to any domain
• Businesses that need a private, on-premise knowledge assistant for their internal documentation
• Privacy-conscious users who want the power of AI without sending data to third-party APIs
• Hobbyists & makers who enjoy tinkering with models, prompts, and configurations to build their ideal assistant
• Organizations operating in environments with limited or no internet connectivity

Summary

Avrana is a modular, configurable, multi-modal AI platform that turns speech and text into intelligent, domain-grounded conversations. Its Understanding microservice architecture sets it apart from monolithic chatbots, delivering more relevant and accurate answers by routing queries through specialized knowledge modules before they reach the LLM.

Key highlights:

• Modular by design — swap, add, or remove any component
• Offline-capable — every stage can run local models
• Multi-lingual — UI and understanding in multiple languages
• Extensible — create custom Understanding modules for any domain
• Customizable — from UI themes to system prompts, everything is yours to shape
• Private — full control over where your data goes and which models process it
• Real-time — voice interaction with streaming audio and text responses