Chapter 15: Thinking Support (AI Chat)

Chapter 15: Thinking Support (AI Chat)

ThinkNavi’s core AI chat feature. Access by clicking “Thinking Support” in the sidebar.

15.1 Screen Layout

Left Sidebar:

• “+ New Chat” button
• Search bar (search conversations)
• Conversation history list (newest first)
• Personal Model status display
• Language switcher

Main Area:

• Chat message display area
• Input area at the bottom

15.2 Sending Messages

Basic Operations:

1. Enter a message in the text input at the bottom
2. Press Enter to send (Shift+Enter for new line)
3. Or click the send button (paper airplane icon)
4. AI streams its response (text appears in real-time)

Input Field Buttons:

15.3 Conversation Management

Conversation history is listed in the sidebar.

Starting a New Conversation

Click the “+ New Chat” button. A new empty conversation is created and the main area clears.

Switching Conversations

Click a conversation title in the sidebar to switch to it.

Renaming Conversations

1. Hover over a conversation item
2. A pencil (edit) icon appears
3. Click to make the title editable
4. Enter a new title and confirm with the checkmark
5. Cancel with the × mark

Keeping Conversations (Favorites)

• Click the bookmark icon on a conversation item
• Kept conversations are pinned at the top of the list

Deleting Conversations

1. Click the trash icon on a conversation item
2. Confirmation dialog “Delete conversation?” appears
3. Select “Delete” to remove the conversation and all messages

Searching Conversations

Enter text in the sidebar search box to search by:

• Conversation title
• Message content

Conversations matching the input are filtered and displayed.

Importing Conversations

1. Click the upload icon in the sidebar
2. Select a JSON conversation file
3. Conversation data is imported

15.4 Message Actions

Hover over a message to see action icons at the top-right (or bottom-right).

Keep (Bookmark)

Kept messages are used as input data for Personal Model learning. Keeping important insights, ideas, and conclusions improves Personal Model quality.

Tagging

1. Click the tag icon
2. A popover appears
3. Enter labels separated by commas (e.g., “ideas, decision-making, important”)
4. Press Enter to confirm

Tags are saved as Personal Model metadata and used for message search.

Adding Notes

1. Click the note icon
2. A note input field appears
3. Record your own observations or annotations about the message
4. Press Enter to confirm

Notes are saved with the message and can be reviewed later.

Editing and Regenerating Messages

1. Click the pencil icon on your sent message
2. Message content becomes editable
3. Edit and click the send button
4. All AI responses after that message are regenerated

Note: Conversation history after the edited message will be overwritten.

Code Block Actions

When AI responses contain code:

Python Code Execution:

• Click “Run” to execute code in the browser
• Standard output (print results) appears below the code
• Errors display error messages
• Basic libraries like NumPy are available (Pyodide-supported packages)

15.5 Voice Input/Output

Voice Input (Speech-to-Text)

1. Press and hold the microphone icon (push-to-talk)
2. The microphone icon flashes red, indicating recording
3. Release when done speaking
4. Audio is automatically converted to text via OpenAI Whisper API
5. Converted text appears in the input field
6. Edit text as needed and send

Notes:

• HTTPS access required
• Browser microphone permission required (dialog appears on first access)
• To change microphone permission later, check microphone settings in browser site settings

Voice Output (Text-to-Speech)

Manual Playback:

• Click the speaker icon on an AI response message
• That message is read aloud in the selected voice

Auto Playback:

• When messages are sent via voice input, AI responses are automatically read aloud
• Auto playback does not occur for keyboard-typed messages

Changing Voice: Change the voice type in Settings (Alloy, Echo, Fable, Onyx, Nova, Shimmer)

15.6 Working with Images

Sending Images (Vision)

Have AI analyze and explain images.

Method 1: From the image icon

1. Click the image icon next to the input field
2. Select an image from the file dialog
3. Image preview appears
4. Send with a text message

Method 2: From clipboard

1. Copy an image in another application (Ctrl+C / Cmd+C)
2. Paste in ThinkNavi’s chat input (Ctrl+V / Cmd+V)
3. Image preview appears
4. Send with a text message

Supported Formats: JPEG, PNG, GIF, WebP

Image Generation

Have AI generate images.

Trigger Keywords: Including keywords like the following in messages automatically switches AI to image generation mode:

• “draw”, “illustrate”, “generate image”
• “create an illustration”, “make a picture”

Results: Image appears in the chat. Click to enlarge.

Provider Settings: Select image generation provider (Google Gemini / OpenAI) in Settings.

15.7 Personal Model

Personal Model is your personal concept structure model automatically built from chat history. It functions as a long-term “associative memory” that improves AI response quality.

Checking Personal Model Status

The “Personal Model” section in the sidebar shows:

• Status badge: Not trained / Training / Trained / Training failed
• Message count: Number of kept messages
• Node count: Number of nodes in the built model
• Training step: Current step name (during processing)

Starting Training

Prerequisites: At least 5 “kept” messages required. Messages must be 10+ characters.

Steps:

1. Click the “Start Training” button in the “Personal Model” sidebar section
2. The following steps execute automatically in sequence:

1. Progress bar and status badge show each step’s progress
2. After all steps complete, Personal Model is activated

Personal Model Effects

After training, the following changes occur in chat:

Memory Observer: For each question, an auxiliary LLM (gpt-4o-mini) automatically determines “Does this question need past knowledge?”

Context Injection: When past knowledge is deemed relevant:

1. Keywords are automatically extracted
2. Database search of past chat history
3. Vector search of Personal Model (semantically similar nodes)
4. Search results are injected into the system prompt

Example:

• User: “I want to develop the idea we discussed before about XX”
• Memory Observer: Determines “past knowledge needed”
• Search: Extracts related content from past kept messages
• AI Response: Presents new development ideas building on the past idea

Growing Your Training Data

Tips to improve Personal Model accuracy:

• Actively keep messages containing important insights and ideas
• Use tags to classify messages (improves search accuracy)
• Add notes to supplement context
• Periodically re-train (new kept messages will be reflected)

15.8 Thinking Mode Details

Thinking modes are switched in Settings. Here, each mode’s internal behavior is explained in detail.

Rule Mode Behavior

1. Keyword matching on the user’s message
2. Matched thinking technique hints added to system prompt
3. AI structures responses following the hints

System Prompt Base Instruction:

“You are an assistant that supports the user’s thinking. Read the user’s intent and context carefully, respond in natural language without being overly technical.”

Observer Mode Behavior

1. Auxiliary LLM (gpt-4o-mini) analyzes the user’s message
2. Returns the optimal thinking technique as JSON (e.g., {"technique": "analogy", "reason": "..."})
3. Selected technique’s hints added to system prompt
4. Main LLM responds following the hints

Observer Mode Advantages: More context-aware technique selection than Rule mode

Disadvantages: Slightly slower response start due to auxiliary LLM call (~1-2 seconds additional)

15.9 Switching AI Providers

Multiple AI providers can be used interchangeably.

Streaming Differences: OpenAI uses native streaming (character-by-character display), other providers use SSE (Server-Sent Events) wrapping (chunk-by-chunk display).

15.10 Troubleshooting