Overview
An Agent refers to a natural language processing entity that utilizes an LLM-powered orchestration mechanism to execute multi-space traversal workflows tailored to user’s inquiries. Each agent is configured with a set of input spaces (along with specific instructions for using those spaces) and parameters for the completion LLM.An Agent plans how to leverage multiple embedding spaces to derive enterprise insights for user queries.
How They Work
Agents combine information from multiple embedding spaces to generate enterprise insights, leveraging the LLM for enhanced reasoning. This process involves:- Query Understanding: Parse and understand user queries
- Space Selection: Choose relevant embedding spaces
- Multi-Space Traversal: Query across multiple spaces
- Information Synthesis: Combine results from different spaces
- Response Generation: Generate insights using LLM
Creating
Step 1: Basic Configuration
- Navigate to Agents in your data room
- Click “Create Agent”
- Provide a name and description
- Select the agent type
Step 2: Select Spaces
Choose the embedding spaces the agent will use:- Select one or more spaces
- Configure space-specific instructions
- Define how each space should be used
- Set space priorities
Step 3: Configure LLM
Set up the language model:- Choose LLM provider (OpenAI, Anthropic, etc.)
- Select model (GPT-4, Claude, etc.)
- Configure model parameters
- Set temperature and other settings
Step 4: Add Instructions
Provide instructions for the agent:- Define agent behavior
- Specify how to use spaces
- Set response format preferences
- Add domain-specific guidance
Step 5: Activate
- Review configuration
- Test with sample queries
- Activate the agent
- Monitor performance
Configuration
Space Configuration
For each space, you can specify:- Usage Instructions: How to interpret space results
- Priority: Which spaces to prioritize
- Filters: Constraints on space queries
- Context: Additional context for space usage
LLM Configuration
Configure the language model:- Provider: OpenAI, Anthropic, etc.
- Model: Specific model version
- Temperature: Creativity vs. consistency
- Max Tokens: Response length limit
- System Prompt: Base instructions
Query Handling
Configure how queries are processed:- Query Parsing: How to interpret queries
- Intent Recognition: Understanding user intent
- Query Expansion: Enhancing queries
- Context Management: Maintaining conversation context
Types
Single-Space
Agents that use a single embedding space:- Simpler configuration
- Focused on specific domain
- Faster response times
Multi-Space
Agents that traverse multiple spaces:- More comprehensive insights
- Cross-domain analysis
- Complex query handling
Specialized
Agents configured for specific use cases:- Financial analysis
- Customer insights
- Product recommendations
- Document search
Using
Query Interface
Interact with agents through:- Web Interface: Chat interface in dashboard
- API: Programmatic access
- MCP Integration: Model Context Protocol
- Custom Integrations: Framework-specific
Query Examples
Financial Analysis:Best Practices
Space Selection
- Choose spaces relevant to your use case
- Balance coverage and performance
- Test different space combinations
Instruction Design
- Be specific about agent behavior
- Provide clear examples
- Include domain knowledge
- Test and refine instructions
LLM Configuration
- Choose appropriate models
- Balance cost and performance
- Tune temperature for your use case
- Monitor token usage
Performance Optimization
- Cache frequent queries
- Optimize space queries
- Monitor response times
- Scale as needed
Monitoring
Metrics
Track agent performance:- Query Count: Number of queries processed
- Response Time: Average response time
- Success Rate: Percentage of successful queries
- User Satisfaction: Feedback and ratings
Logs
Review agent logs:- Query history
- Space usage
- LLM interactions
- Error logs