Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/TrinaxCode/TrinaxAI/llms.txt

Use this file to discover all available pages before exploring further.

TrinaxAI runs voice and vision capabilities entirely on your device. Speech recognition uses the browser’s Web Speech API, text-to-speech uses browser TTS, and image analysis runs through a local qwen2.5vl model via Ollama. No audio, no images, and no transcripts are ever sent to a cloud API.

​
Voice Mode

Voice mode lets you speak your queries and hear responses read aloud — a natural, hands-free conversation with your local AI.

​
How It Works

1

Speech Recognition

TrinaxAI uses the browser’s Web Speech API (SpeechRecognition) for speech-to-text. The browser streams audio to its built-in recognition engine (on Chrome/Edge, this is on-device when offline or via Google’s speech service when online — see the privacy note below). The recognised text is inserted into the chat input field.
2

Response Synthesis

When a response is received, TrinaxAI uses the browser’s Web Speech Synthesis API (SpeechSynthesisUtterance) to read it aloud. Responses are split at sentence boundaries so playback begins as soon as the first sentence is available, without waiting for the full reply.
3

Interrupt Support

You can interrupt TrinaxAI mid-sentence. Speaking or pressing the voice button while audio is playing cancels the current utterance and starts listening immediately. This keeps conversation flow natural.
The Web Speech API’s speech-to-text engine may use an online service (Chrome uses Google’s servers when connected). Text-to-speech synthesis is always local to your browser. If full offline voice is required, use the PWA in a browser that supports on-device speech recognition (e.g., Firefox with a local engine).

​
Activating Voice Mode in the PWA

  1. Open TrinaxAI at https://localhost:3334
  2. Click the microphone icon (🎤) in the chat input bar
  3. Grant microphone permission when prompted — this permission is remembered for the origin
  4. Speak your question; the transcript appears in the input field as you talk
  5. TrinaxAI sends the message automatically when you pause, or press Enter to send immediately
  6. Toggle the speaker icon (🔊) to enable or disable text-to-speech responses
Voice mode pairs well with the Ollama engine for quick conversational exchanges. For code questions against your indexed projects, switch to the RAG engine — cited answers are also read aloud.

​
Vision: Image Analysis

TrinaxAI can analyse images and screenshots you attach to the chat. The entire analysis runs locally on your machine through a qwen2.5vl vision-language model served by Ollama.

​
Vision Models

Two model sizes are available, configured as Vite build-time environment variables:
VariableDefaultUse Case
VITE_TRINAXAI_VISION_MODELqwen2.5vl:3bDefault — fast, good quality, runs on 8 GB RAM
VITE_TRINAXAI_VISION_QUALITY_MODELqwen2.5vl:7bQuality mode — better detail analysis, needs 16 GB+
The PWA reads VITE_TRINAXAI_VISION_QUALITY_MODEL when you enable Quality vision mode in Settings. Both are standard Ollama pull targets.

​
Attaching an Image

1

Open the Attachment Menu

Click the paperclip / image icon (📎) in the chat input bar.
2

Choose an Image

Select any image file from your device (JPEG, PNG, WebP, GIF). You can also paste an image directly from the clipboard.
3

Ask Your Question

Type your question about the image — or leave the text field empty to get a general description. Examples: “What does this error message mean?”, “Describe the UI layout”, “What’s wrong with this chart?”
4

Receive a Local Analysis

TrinaxAI sends the image and your question to qwen2.5vl:3b (or :7b in quality mode) via Ollama’s vision API. The response streams back like any other chat message. No image data leaves your machine.

​
Image Preprocessing

Before sending large images to the vision model, TrinaxAI preprocesses them in the browser to prevent out-of-memory (OOM) errors on the local Ollama process:
  • Images are downscaled if their dimensions exceed a safe threshold
  • Files are re-encoded as JPEG at reduced quality when the file size is large
  • The processed image is base64-encoded and sent in the Ollama API payload
This keeps memory usage predictable even when analysing high-resolution screenshots.

​
Skipping Vision During Install

If your machine doesn’t have enough RAM for a vision model, skip the download during installation: 
./install.sh --no-vision
Vision analysis will be unavailable until you pull a vision model manually: 
ollama pull qwen2.5vl:3b

​
Privacy Guarantee

Everything stays local. Vision requests are sent to Ollama at http://localhost:11434 — the same endpoint used for text models. No image, screenshot, or analysis result is transmitted to any external server.

​
Configuration Reference

VariableDefaultDescription
VITE_TRINAXAI_VISION_MODELqwen2.5vl:3bOllama model for standard vision analysis
VITE_TRINAXAI_VISION_QUALITY_MODELqwen2.5vl:7bOllama model for high-quality vision analysis

Build docs developers (and LLMs) love

Get started for freeTalk to us