Internal CLI
Built-in CLI tool for testing, demonstration, and interactive conversations with AbstractCore. This is AbstractCore's internal testing CLI, not to be confused with external agentic CLIs.
🚀 Quick Start
Start the internal CLI with any provider:
# Start the internal CLI
python -m abstractcore.utils.cli --provider ollama --model qwen3-coder:30b
# Or with any provider
python -m abstractcore.utils.cli --provider openai --model gpt-4o-mini
python -m abstractcore.utils.cli --provider anthropic --model claude-3-5-haiku-latest
# With streaming enabled
python -m abstractcore.utils.cli --provider ollama --model qwen3-coder:30b --stream💡 Note: Looking for external agentic CLI integration (Codex, Gemini CLI, Crush)? See the Server Documentation for complete setup guides.
⚡ Available Commands
/compact
Compress chat history while preserving context
/compact
/facts
Extract facts from conversation as structured data
/facts [file]
/judge
Evaluate conversation quality using LLM-as-a-judge
/judge
/stream
Toggle real-time streaming of responses
/stream
/history
Show conversation history without truncation
/history [n]
/system
Manage system prompt that guides AI behavior
/system [prompt]
📎 Media Handling
The internal CLI supports universal media handling with the @filename syntax for attaching images, documents, and data files.
@filename Syntax
Attach any file type using the @filename syntax in your prompt:
# Image analysis
python -m abstractcore.utils.cli --prompt "What's in this image? @photo.jpg"
# PDF document analysis
python -m abstractcore.utils.cli --prompt "Summarize @report.pdf"
# Office documents
python -m abstractcore.utils.cli --prompt "Extract key points from @presentation.pptx"
python -m abstractcore.utils.cli --prompt "Analyze the data in @spreadsheet.xlsx"
# Data files
python -m abstractcore.utils.cli --prompt "What patterns are in @sales_data.csv"
# Multiple files
python -m abstractcore.utils.cli --prompt "Compare @chart.png and @data.csv"Supported File Types
- Images: PNG, JPEG, GIF, WEBP, BMP, TIFF
- Documents: PDF, DOCX, XLSX, PPTX
- Data/Text: CSV, TSV, TXT, MD, JSON, XML
How It Works
The system automatically:
- Detects file type from extension
- Applies appropriate processor (Image, PDF, Office, Text)
- Formats content for the selected provider's API
- Handles vision fallback for text-only models when needed
📚 Command Details
🗜️ /compact - Chat History Compaction
Compacts your chat history using the fast local gemma3:1b model to create a summary while preserving recent exchanges.
# In the CLI
/compact
# Output example:
🗜️ Compacting chat history...
Before: 15 messages (~450 tokens)
Using gemma3:1b for compaction...
✅ Compaction completed in 3.2s
After: 5 messages (~280 tokens)
Structure:
1. ⚙️ System prompt
2. 📚 Conversation summary (1,200 chars)
3. 👤 How do I handle errors in Python?
4. 🤖 You can use try/except blocks...
5. 👤 What about logging?When to use:
- Long conversations that slow down responses
- When you want to preserve context but reduce memory usage
- Before switching to a different topic
⚖️ /judge - Conversation Quality Evaluation
Evaluates the quality and interest of the current discussion using LLM-as-a-judge. This demonstrates objective assessment capabilities with constructive feedback.
# In the CLI
/judge
# Output example:
⚖️ Evaluating conversation quality...
Analyzing 450 characters of conversation...
✅ Evaluation completed in 35.2s
📊 Overall Discussion Quality: 4/5
📈 Quality Dimensions:
Clarity : 5/5
Coherence : 4/5
Actionability: 3/5
Relevance : 5/5
Completeness : 4/5
Soundness : 4/5
Simplicity : 5/5
✅ Conversation Strengths:
• The discussion maintains clear focus and addresses questions directly
• Explanations are well-structured and easy to follow
• Technical concepts are explained in accessible language
💡 Suggestions for Better Discussions:
• Include more specific examples to illustrate key points
• Add actionable next steps or recommendations where appropriate
• Consider exploring alternative approaches or edge casesWhen to use:
- Get objective feedback on discussion quality
- Understand strengths and areas for improvement
- Demonstrate LLM-as-a-judge evaluation capabilities
- Practice critical assessment and constructive feedback
🔍 /facts - Extract Structured Facts
Extracts facts from your conversation history as simple triples (subject-predicate-object). Display them in chat or save as structured JSON-LD.
# Display facts in chat
/facts
# Output example:
🔍 Extracting facts from conversation history...
Processing 415 characters of conversation...
✅ Fact extraction completed in 4.5s
📋 Facts extracted from conversation:
==================================================
1. OpenAI creates GPT-4
2. Microsoft Copilot uses GPT-4
3. Google develops TensorFlow
==================================================
📊 Found 3 entities and 2 relationships
# Save to JSON-LD file
/facts myconversation
# Output example:
🔍 Extracting facts from conversation history...
Processing 415 characters of conversation...
✅ Fact extraction completed in 3.3s
💾 Facts saved to myconversation.jsonld
📊 Saved 3 entities and 2 relationships as JSON-LDWhen to use:
- Extract key information discussed in the conversation
- Create structured knowledge from chat sessions
- Document facts for later reference or analysis
- Generate semantic data for knowledge graphs
🏷️ /tooltag - Test Tool Call Tag Rewriting
Test tool call tag rewriting with custom tags to verify compatibility with different agentic CLIs.
# Test LLaMA3 format (Crush CLI)
/tooltag '' ' '
# Test XML format (Gemini CLI)
/tooltag '' ' '
# Test Qwen3 format (Codex CLI)
/tooltag '<|tool_call|>' ''
# Test custom format
/tooltag '' ' 'When to use:
- Test compatibility with specific agentic CLIs
- Verify tool call tag rewriting works correctly
- Debug tool call format issues
- Compare different tag formats side by side
💡 Usage Examples
Complete Workflow Example
# Start CLI with a model that supports auto-compaction
python -m abstractcore.utils.cli --provider ollama --model gemma3:1b
# Have a long conversation...
👤 You: What is Python?
🤖 Assistant: Python is a high-level programming language...
👤 You: What about data types?
🤖 Assistant: Python has several built-in data types...
# Check history
👤 You: /history 2
# Change system prompt for specific task
👤 You: /system You are an expert Python tutor focused on data science.
# Compact when needed
👤 You: /compact
# Extract facts from the conversation
👤 You: /facts
# Continue conversation with preserved context
👤 You: Can you explain more about functions?
🤖 Assistant: [Responds as Python tutor, refers to previous context]Controlling Tool Usage
If the AI is calling tools unnecessarily, you can control this behavior:
# Check current system prompt to understand behavior
👤 You: /system
# Update to prevent unnecessary tool calls
👤 You: /system You are a helpful AI assistant. Only use tools when explicitly requested by the user or when necessary to answer a specific question. Do not demonstrate tool capabilities unprompted.
# Now the AI will be less likely to call tools unnecessarily
👤 You: Who are you?
🤖 Assistant: I am an AI assistant. [No tool calls]✅ Best Practices
Regular History Checks
Use /history n to review recent context and understand conversation flow.
Strategic Compaction
Use /compact when conversation becomes slow, you've covered multiple topics, or before switching contexts.
System Prompt Control
Use /system to adapt AI behavior for different tasks and prevent unwanted tool calls.
Context Verification
After compaction, use /history to verify preserved context and summary quality.
🔧 Technical Details
- Compaction Model: Uses
gemma3:1bfor fast, local processing - Preservation: Keeps last 6 messages (3 exchanges) by default
- Fallback: If
gemma3:1bunavailable, uses current provider - Token Estimation: Provides rough token count estimates
- Message Types: Distinguishes between system, user, and assistant messages
- Error Handling: Gracefully handles various scenarios with helpful messages