Skip to content

Text Generation

Text Generation is Bro's primary local Gemma chat and text-task interface.

Supported tasks

The Task Type selectbox exposes the same canonical task list consumed by the instruction builder:

  • Chat
  • Analysis
  • Reasoning
  • Coding
  • Writing
  • Editing
  • Translation
  • Summarization
  • Extraction
  • Classification
  • Comparison
  • Structured Output

This avoids maintaining a UI subset that disagrees with the execution path.

🧭 Task Preset

One row contains:

Control Purpose
Task Type Selects the instruction branch.
Task Detail Concise, Standard, or Detailed guidance.
Task Focus Accuracy, Balanced, or Creativity focus.

The bottom 🔄 Reset resets all Task Preset values.

🧩 Reasoning Controls

  • Reasoning Depth
  • Answer Only
  • Use Self-Check
  • Prefer Deterministic Reasoning

These are prompt-strategy controls; they are not presented as a separate native Gemma thinking API.

🧾 Coding Controls

One row contains:

  • Code Language
  • Coding Task
  • Include Comments
  • Use Editor Format
  • Emit Fenced Code

Coding tasks

  • Generate
  • Complete
  • Refactor
  • Debug
  • Review
  • Explain
  • Optimize
  • Convert
  • Test
  • Document
  • Design

Coding languages

Bro uses a bounded language selector including Python, C, C++, C#, Java, JavaScript, TypeScript, SQL, VBA, PowerShell, Bash, HTML, CSS, Markdown, JSON, YAML, and Other.

✍️ Writing Controls

  • Writing Task
  • Tone
  • Audience
  • Length

These settings apply only to writing/editing instruction construction.

🌐 Translation Controls

Translation has its own expander rather than sharing Coding Controls.

  • Source Language
  • Target Language
  • Translation Mode
  • Preserve Formatting

Translation modes are:

  • Natural
  • Literal
  • Formal
  • Technical
  • Localization

Language arguments are bounded selectbox values. The source selector includes Auto Detect.

🏷️ Classification Controls

  • Classification Type
  • Return Confidence
  • Allow Unknown
  • Explain Classification

Classification types include Binary, Multi-Class, Multi-Label, Sentiment, Intent, Topic, and Relevance.

↔️ Response Controls

Response Controls describe the output contract rather than the sampling algorithm:

  • Response Format
  • Response Language
  • Response Length
  • Include Headings

Supported response formats are:

  • Plain Text
  • Markdown
  • Bullet List
  • Numbered List
  • Markdown Table
  • JSON
  • XML
  • YAML
  • CSV
  • Code

JSON requests can use llama.cpp's JSON response-format constraint.

🎚️ Inference Settings

Inference Settings map to generation/runtime arguments.

Row 1

  • Temperature
  • Top-P
  • Top-K
  • Repeat Penalty
  • Repeat Window

Row 2

  • Presence Penalty
  • Frequency Penalty
  • Random Seed
  • Max Tokens

The single bottom Reset resets both rows.

🎛️ Context Controls

  • Context Window
  • Use Conversation History
  • Use Document Context
  • Use Semantic Context
  • Use Grounding

These settings affect context/message construction, not response formatting.

⚙️ Runtime Settings

  • CPU Threads
  • Batch Size
  • Micro Batch Size

🖥️ System Instructions

System Instructions provide:

  • Category selectbox;
  • category-filtered template selectbox;
  • ID-backed template loading;
  • editable system-instruction text;
  • clear/reset;
  • XML ↔ Markdown conversion;
  • Apply Preset;
  • Effective Prompt Preview.

Template captions are display values; database identity is the integer Prompts.ID.

Generation pipeline

task controls
   -> build_task_instruction_block()
   -> build_chat_messages()
   -> get_runtime_llm()
   -> create_chat_completion()
   -> stream/render response