Agent Settings - PLai Framework Documentation

Agent settings are organized into four tabs — Agent, Model, Datasources, and Tools — each covering a distinct part of your agent’s setup, from personality and instructions to model behavior, knowledge sources, and tool access.

Overview

Agent settings allow you to customize every aspect of your AI agent’s behavior, appearance, and capabilities. Each setting plays a crucial role in determining how your agent interacts with users, processes information, and executes tasks.

Agent Tab

Identity, behavior, and personality settings

Model Tab

LLM provider, model selection, and advanced features

Datasources Tab

Knowledge base configuration and retrieval settings

Tools Tab

Tool integration and execution parameters

Agent Configuration

The Agent tab contains settings related to your agent’s identity, appearance, and core behavior.

Avatar

Upload a custom image to represent your agent in chat interfaces.

Supported Formats: .jpg, .jpeg, .png
Recommended Size: 512x512 pixels for optimal display

Use Cases:

Brand consistency across customer-facing agents
Visual differentiation between multiple agents
Enhanced user experience with personalized avatars

Name

The human-readable identifier for your agent.

string

required

Display name that appears throughout the interface

Best Practices:

Keep it clear and descriptive (e.g., “Customer Support Bot”, “Data Analyst”)
Avoid special characters that might cause display issues
Consider your branding and user-facing context

Example:

Marketing Content Assistant

Description

A brief summary of your agent’s purpose and capabilities.

string

required

Explains what the agent does and when to use it

Best Practices:

Be specific about the agent’s specialty
Mention key capabilities or domains
Keep it concise (1-2 sentences)

Example:

This agent is an expert at creating engaging marketing content, 
including blog posts, social media updates, and email campaigns.

Name Slug

A unique, URL-friendly identifier for your agent.

string

Lowercase identifier with underscores (minimum 5 characters)

Format Requirements:

Lowercase letters and numbers only
Use underscores to separate words
Minimum 5 characters
Must match pattern: ^[a-z0-9]+(?:_[a-z0-9]+)*$

Examples:

✅ customer_support_agent
✅ data_analysis_bot
✅ marketing_assistant_v2

❌ CustomerSupport (uppercase not allowed)
❌ agent-name (hyphens not allowed)
❌ bot (too short)

Core Agent

Availability: This setting is only visible for Core Agents Organization accounts.

boolean

default:"false"

Designates this agent as a core system agent

Core agents have special privileges and are typically used for system-level operations.

Prompt (System Instructions)

The system prompt defines your agent’s personality, behavior guidelines, and response style.

string

required

Instructions that guide the agent’s behavior and responses

Components of a Good Prompt:

Role Definition: Who the agent is
Capabilities: What it can do
Constraints: What it should/shouldn’t do
Style Guidelines: How it should communicate
Context: Additional relevant information

Basic Example
Advanced Example
Markdown Editor

You are a helpful customer support assistant for an e-commerce platform.

Your role is to:
- Answer questions about products, orders, and policies
- Help troubleshoot common issues
- Guide users through the return process

Always be friendly, professional, and empathetic. If you don't know 
something, admit it and offer to connect the user with a human agent.

You are an AI Data Analyst specializing in business intelligence and 
predictive analytics.

## Your Capabilities:
- Analyze datasets and identify trends
- Create visualizations and reports
- Provide actionable business insights
- Run SQL queries on connected databases

## Communication Style:
- Be precise and data-driven
- Use clear, jargon-free language
- Always cite your sources
- Provide confidence levels for predictions

## Constraints:
- Never make up data or statistics
- Always validate data quality before analysis
- Respect data privacy and security guidelines
- Escalate sensitive queries to human analysts

## Context:
You have access to the company's sales database (2020-present) and 
customer demographics data. Focus on actionable insights that drive 
business decisions.

Important: Changes to the prompt can significantly alter your agent’s behavior. Always test thoroughly after updates.

Initial Message (Intro Message)

The first message your agent sends when a new conversation starts.

string

Greeting message displayed at the start of conversations

Best Practices:

Be welcoming and friendly
Set expectations for what the agent can do
Keep it concise
Consider adding a call-to-action

Examples:

Simple
Informative
With Options

Hi, how can I help you today?

Hello! I'm your marketing content assistant. I can help you create 
blog posts, social media content, and email campaigns. What would 
you like to work on today?

Welcome! I'm here to help you with:
• Product questions
• Order tracking
• Returns and refunds

What can I assist you with?

Model Configuration

The Model tab contains all settings related to the language model powering your agent.

Language Model Selection

Choose the AI model provider and specific model for your agent.

string

required

AI provider (OpenAI, Anthropic, Google, etc.)

string

required

Specific model within the selected provider

Available Providers:

OpenAI: GPT-4o, GPT-4.1, GPT-5
Anthropic: Claude Sonnet, Claude Opus, Claude Haiku
Google: Gemini Pro models, Gemini Flash models
OpenRouter: Open-source models (Llama, Mixtral, DeepSeek, Gemma 4, Kimi K2/3)
RouteLLM: Intelligent model routing

Selection Guide

Use Case	Recommended Model	Reason
Complex reasoning	GPT o1/o3 / Claude 4.x Opus	Best accuracy
Fast responses	GPT-4.1 mini / Gemini 2.5 Flash	Low latency
Cost optimization	RouteLLM	Automatic routing
Long context	Claude 4.x Sonnet / Gemini 3 Pro	Large context window
Structured output	GPT-4.x models / Gemini 3 Pro	Native support

RouteLLM Configuration

What is RouteLLM?
RouteLLM intelligently routes requests between a “strong” model (for complex tasks) and a “weak” model (for simple tasks) based on query complexity, optimizing cost and performance.

When RouteLLM is selected as the provider, additional configuration options become available:

Threshold

number

default:"0.11"

Sensitivity for routing between strong and weak models (0.01 - 0.30)

How it Works:

Lower threshold (0.01 - 0.10): Routes more queries to the strong model (higher quality, higher cost)
Higher threshold (0.15 - 0.30): Routes more queries to the weak model (lower cost, faster)
Default (0.11): Balanced approach

Example Scenarios:

Threshold	Strong Model Usage	Best For
0.05	~70%	Critical applications requiring high accuracy
0.11	~50%	Balanced cost and quality
0.20	~30%	Cost-sensitive applications

Strong Model Config

Configuration for the high-performance model used for complex queries.

string

Provider for the strong model

string

Specific strong model

number

Creativity level for strong model (0.0 - 2.0)

number

Maximum response length for strong model

Weak Model Config

Configuration for the efficient model used for simple queries.

string

Provider for the weak model

string

Specific weak model

number

Creativity level for weak model (0.0 - 2.0)

number

Maximum response length for weak model

Enable Streaming

boolean

default:"true"

Stream responses in real-time as they’re generated

Benefits:

✅ Better user experience with progressive display
✅ Reduced perceived latency
✅ More interactive feel
✅ Users can start reading immediately

Limitations:

⚠️ Not supported by all models
⚠️ Cannot be used with Structured Output
⚠️ May not work well with certain integrations

Streaming and Structured Output are mutually exclusive. Enabling one will automatically disable the other.

When to Disable Streaming:

API integrations requiring complete responses
Batch processing scenarios
When using Structured Output
WebSocket limitations in your application

Enable Language Detection

boolean

default:"false"

Automatically detect and respond in the user’s language

How it Works:

Agent analyzes the user’s first message
Detects the language automatically
Responds in the same language throughout the conversation

Benefits:

✅ Seamless multi-language support
✅ No manual configuration needed
✅ Better global user experience
✅ Works with all models

Example:

User: "¿Cuál es el estado de mi pedido?"
Agent: "Déjame verificar el estado de tu pedido..."

User: "What's the status of my order?"
Agent: "Let me check the status of your order..."

Best Practice: Enable this for public-facing agents serving international audiences.

Enable Citations

boolean

default:"false"

Include source citations in agent responses

This feature is currently in development and may not be fully functional in all scenarios.

When enabled, the agent will attempt to cite sources when providing information from datasources or external knowledge.

Structured Output

boolean

default:"false"

Force responses to follow a predefined JSON schema

Requirements:

✅ Model must support structured outputs (GPT-4, Gemini Pro)
✅ JSON Schema must be defined
❌ Cannot be used with RouteLLM
❌ Cannot be used with Streaming

Important Limitations:

Not available with RouteLLM provider
Streaming must be disabled
Requires compatible model

JSON Schema

string

JSON Schema defining the structure of agent responses

When Structured Output is enabled, you must define a JSON Schema that specifies the exact format of the agent’s responses.

Contact Extraction
Sentiment Analysis
Product Catalog
Schema Generator

{
  "type": "object",
  "properties": {
    "name": {
      "type": "string",
      "description": "Full name of the person"
    },
    "email": {
      "type": "string",
      "format": "email",
      "description": "Email address"
    },
    "phone": {
      "type": "string",
      "description": "Phone number"
    },
    "company": {
      "type": "string",
      "description": "Company name"
    }
  },
  "required": ["name", "email"],
  "additionalProperties": false
}

Use Case: Extracting contact information from resumes, emails, or forms.

{
  "type": "object",
  "properties": {
    "sentiment": {
      "type": "string",
      "enum": ["positive", "negative", "neutral"],
      "description": "Overall sentiment"
    },
    "confidence": {
      "type": "number",
      "minimum": 0,
      "maximum": 1,
      "description": "Confidence score"
    },
    "key_phrases": {
      "type": "array",
      "items": {
        "type": "string"
      },
      "description": "Important phrases from the text"
    },
    "summary": {
      "type": "string",
      "description": "Brief summary of the sentiment"
    }
  },
  "required": ["sentiment", "confidence"]
}

Use Case: Analyzing customer feedback, reviews, or social media posts.

{
  "type": "object",
  "properties": {
    "product_name": {
      "type": "string"
    },
    "category": {
      "type": "string",
      "enum": ["electronics", "clothing", "food", "books", "other"]
    },
    "price": {
      "type": "object",
      "properties": {
        "amount": {
          "type": "number",
          "minimum": 0
        },
        "currency": {
          "type": "string",
          "default": "USD"
        }
      },
      "required": ["amount", "currency"]
    },
    "features": {
      "type": "array",
      "items": {
        "type": "string"
      }
    },
    "in_stock": {
      "type": "boolean"
    }
  },
  "required": ["product_name", "category", "price"]
}

Use Case: Extracting structured product information from descriptions or catalogs.

Common Use Cases:

Data extraction and transformation
API integration with strict requirements
Database population from unstructured text
Form filling automation
Consistent report generation

Temperature

number

Controls randomness and creativity in responses (0.0 - 2.0)

Temperature Scale:

Value	Behavior	Best For
0.0 - 0.3	Deterministic, focused	Factual Q&A, data extraction, structured tasks
0.4 - 0.7	Balanced	General conversation, customer support
0.8 - 1.2	Creative, varied	Content creation, brainstorming
1.3 - 2.0	Highly creative	Creative writing, diverse ideas

Examples:

Low (0.2)
Medium (0.7)
High (1.5)

Prompt: “Describe a sunset”Response:

A sunset is the daily event when the sun descends below the horizon. 
The sky typically displays colors ranging from orange to red due to 
the scattering of light through the atmosphere.

✅ Consistent, factual, predictable

Prompt: “Describe a sunset”Response:

The sunset painted the sky in warm hues of orange and pink as the 
sun slowly dipped below the horizon. The clouds caught the fading 
light, creating a beautiful display that marked the end of the day.

✅ Balanced, natural, appropriate variation

Prompt: “Describe a sunset”Response:

Like a masterful painter's final brushstroke, the sun melted into 
the horizon, bleeding crimson and gold across the canvas of the sky. 
Clouds became silken ribbons of fire, dancing in the twilight's 
embrace, while shadows stretched long and lazy across the earth.

✅ Creative, poetic, highly varied

Some models don’t support temperature adjustments. The interface will display an alert if temperature control is unavailable for your selected model.

Max Output Tokens

number

Maximum length of the agent’s responses (in tokens)

What are Tokens?

Tokens are pieces of words used by language models
Roughly 1 token ≈ 4 characters or ≈ 0.75 words
Both input and output count toward limits

Common Settings:

Tokens	Words (approx)	Best For
256	~192	Short answers, chatbots
512	~384	Standard responses
1024	~768	Detailed explanations
2048	~1536	Long-form content
4096+	~3072+	Articles, reports

Cost Considerations:

Higher token limits = higher costs per request
Unused tokens still count toward limits
Balance between thoroughness and cost

Setting this too low may cause responses to be cut off mid-sentence. Monitor your agent’s behavior and adjust accordingly.

Max Steps (Tool Execution)

number

default:"1"

Maximum number of tool executions per agent response (1 - 128)

What are Steps? Each time your agent uses a tool, it counts as one step. Multiple steps allow the agent to:

Chain multiple tool calls together
Iterate on results
Execute complex multi-step workflows

Recommended Values:

Steps	Use Case	Example
1-3	Simple tool use	Single database query, one API call
4-10	Multi-step workflows	Search → Analyze → Summarize
11-25	Complex automation	Data gathering → Processing → Formatting → Output
26+	Advanced workflows	Complex research or data pipelines

Example Workflow (Max Steps: 5):

Step 1: Search knowledge base for product info
Step 2: Query inventory database for stock levels
Step 3: Check pricing API for current prices
Step 4: Calculate shipping costs
Step 5: Format response with all information

Higher step limits give your agent more autonomy but can increase response time and costs. Start conservative and increase as needed.

Datasources Configuration

The Datasources tab configures how your agent retrieves and uses information from connected knowledge bases.

Vector Top K

number

default:"10"

Number of most relevant document chunks to retrieve from vector search

How Vector Search Works:

User query is converted to a vector embedding
Vector database finds similar document chunks
Top K most similar chunks are retrieved
Agent uses these chunks to formulate response

Choosing the Right Value:

Value	Retrieval Scope	Best For
3-5	Narrow, focused	Precise questions with clear answers
6-10	Balanced	General purpose Q&A
11-20	Broad	Complex questions requiring multiple sources
20+	Comprehensive	Research, detailed analysis

Trade-offs:

Lower K: Faster, more focused, may miss relevant context
Higher K: More comprehensive, slower, may include irrelevant info

Enable Rerank

boolean

default:"false"

Use a reranking model to improve relevance of retrieved documents

What is Reranking? After initial vector search, a specialized reranking model re-evaluates the retrieved documents and reorders them by true relevance to the query. Benefits:

✅ Improved answer accuracy
✅ Better handling of complex queries
✅ Reduced hallucinations
✅ More relevant context for the agent

Reranking adds latency (~100-300ms) and additional costs. Use when accuracy is more important than speed.

Rerank Top K

number

Number of top documents to keep after reranking

This setting only appears when “Enable Rerank” is turned on.

Typical Configuration:

Vector Top K: 20 (cast a wide net)
Rerank Top K: 5 (keep only the best)

This approach retrieves many candidates but only uses the most relevant after reranking. Example:

Vector Search retrieves 20 documents → 
Reranking model scores all 20 →
Keep top 5 most relevant →
Agent uses these 5 for response

Rerank Threshold

number

default:"0.2"

Minimum relevance score required to include a document (0.01 - 1.0)

This setting only appears when “Enable Rerank” is turned on.

How it Works: Documents with reranking scores below this threshold are filtered out, even if they’re in the Top K. Choosing a Threshold:

Threshold	Strictness	Result
0.01 - 0.1	Lenient	Includes marginal matches
0.15 - 0.3	Moderate	Balanced filtering (recommended)
0.4 - 0.6	Strict	Only highly relevant documents
0.7 - 1.0	Very Strict	May exclude too many results

Example Scenario:

Rerank Top K: 5
Rerank Threshold: 0.2

Reranked Results:
1. Doc A (score: 0.89) ✅ Included
2. Doc B (score: 0.75) ✅ Included
3. Doc C (score: 0.45) ✅ Included
4. Doc D (score: 0.18) ❌ Excluded (below threshold)
5. Doc E (score: 0.12) ❌ Excluded (below threshold)

Agent receives: 3 documents (only those above threshold)

Multi-Query RAG

boolean

default:"false"

Enable Multi-Query RAG to generate multiple search queries from user input

What is Multi-Query RAG? Instead of searching the datasource using only the user’s original query, Multi-Query RAG automatically generates multiple rephrased versions of the query and searches for all variants. This improves retrieval by covering different ways of expressing the same information. How It Works:

User query: “What’s the password reset process?”
Multi-Query RAG generates:
- Original: “What’s the password reset process?”
- Variant 1: “How do I reset my password?”
- Variant 2: “Password reset instructions”
- Variant 3: “Forgot password procedure”
- Variant 4: “Change account password”
All variants are searched — results are combined and deduplicated
Agent uses the comprehensive, combined results

Benefits:

✅ Better coverage: Finds relevant documents that match any variant
✅ Handles synonyms: Different ways of asking about the same thing
✅ Improves accuracy: More comprehensive context for agent
✅ Natural language: Works with how people actually phrase questions
✅ Reduces missed results: Won’t miss relevant docs just because of wording

Trade-offs:

More API calls: Multiple searches instead of one (5-10x slower)
Higher costs: More queries = more computation
May include noise: Variant queries might retrieve less relevant results

When to Use:

Scenario	Recommendation
FAQ/Customer Support	✅ Recommended - handles varied phrasing
Technical documentation	✅ Recommended - people ask in different ways
Simple lookup (IDs, names)	❌ Skip it - exact match queries don’t need variants
Real-time constraints	❌ Skip it - latency matters more than coverage
Large datasources	⚠️ Use carefully - may increase costs significantly

Configuration Tips: If Multi-Query RAG is enabled, you might want to:

Increase vector_topk (e.g., to 15-20) since variants may retrieve different results
Keep rerank_enabled ON to filter the combined results
Adjust rerank_topk down (e.g., to 3-5) to focus on best matches

Example Setup:

Vector Top K: 15 (wider net for each query variant)
Multi-Query RAG: Enabled
Enable Rerank: Yes (to filter combined results)
Rerank Top K: 5 (keep only the best)
Rerank Threshold: 0.3 (strict filtering)

Result: 3-5 highly relevant documents from comprehensive search

Multi-Query RAG increases latency and costs significantly. Use it when answer quality matters more than speed. For real-time applications, consider disabling it.

Datasources Selection

array

List of datasource IDs that this agent can access

What are Datasources? Datasources are knowledge bases containing documents, files, or data that your agent can search and reference. Each datasource can contain:

PDF documents
Web pages
Text files
Structured data
API integrations

Configuration: Toggle each datasource on/off to control which knowledge bases your agent can access. Best Practices:

✅ Only enable relevant datasources to reduce noise
✅ Separate datasources by topic or domain
✅ Regularly update datasource content
⚠️ Too many datasources can slow retrieval
⚠️ Ensure datasource content is high quality

Example Setup:

Agent Type	Enabled Datasources
Customer Support	Product docs, FAQ, Return policies
Sales Agent	Product catalog, Pricing, Case studies
Technical Support	Technical docs, API reference, Troubleshooting guides
HR Assistant	Company policies, Benefits info, Onboarding materials

Tools Configuration

The Tools tab manages your agent’s ability to execute actions and interact with external systems.

Tool Max Steps

number

default:"1"

Maximum number of tool calls the agent can make per response (1 - 128)

This is the same setting as “Max Steps” in the Model tab, shown here for convenience when configuring tools.

This setting determines how many actions your agent can take in a single interaction. See the Max Steps section in Model Configuration for detailed information.

Tools Selection

array

List of tool IDs that this agent can use

What are Tools? Tools extend your agent’s capabilities beyond conversation. They allow the agent to:

Make API calls
Execute code
Query databases
Search the web
Interact with external systems
Call other agents

Available Tool Types:

API Request: Make HTTP requests to external APIs
Web Search (Perplexity AI): Search the internet for current information
External Datasource: Query external data sources
Agent Tool: Call another agent as a tool
MCP Server: Connect to Model Context Protocol servers
BigQuery: Query Google BigQuery databases

Configuration: Toggle each tool on/off to control which actions your agent can perform.

Circular References: If a tool references the current agent (Agent Tool type), it will be disabled to prevent infinite loops.

Model Compatibility: Some models don’t support tool calling. If your selected model doesn’t allow tools, you’ll see an alert in the interface. Consider switching to a tool-capable model like:

Google Models Built-in Tools

Google-specific Feature: This section only applies when using Google models (Gemini) as your agent’s LLM.

When using Google Gemini models, the interface provides built-in tools that are separate from your custom tools list. These are native capabilities that Google models can use automatically. Google Built-in Tools Available: Google models provide access to built-in tools such as:

Google Search: Real-time web search
Location Services: Geographic/location data
Link Preview: Extract information from URLs
Device Info: Access to device capabilities
And other Google-native integrations

How to Configure:

Go to Model Configuration tab
Select a Google model (e.g., gemini-2.0-flash)
In the Tools Configuration tab, you’ll see a “Google Built-in Tools” section
Toggle each built-in tool on/off independently
These work alongside your custom tools

Example: Google Model: gemini-2.0-flash

Google Built-in Tools:
- Google Search: Enabled
- Location Services: Disabled
- Link Preview: Enabled
- Device Info: Disabled
Custom Tools (from Tools Selection):
- API Request: Enabled
- BigQuery: Enabled
Result: Agent can use Google Search, Link Preview, plus custom tools (API, BigQuery, Code)

Best Practices with Google Built-in Tools:

✅ Enable only what you need - Each enabled tool adds context
✅ Test combinations - Built-in + custom tools work best when tested together
✅ Use for search - Google Search is powerful for real-time info
✅ Combine strategically - Link Preview + Google Search is a common pair
⚠️ Monitor costs - Built-in tool usage counts toward your API quota
⚠️ Privacy considerations - Be aware of data sent to Google services

Comparison: Google Built-in vs Custom Tools

Aspect	Google Built-in	Custom Tools
Setup	Built-in, just toggle	Requires configuration
Speed	Native to model, fast	Depends on implementation
Cost	Included in model cost	May have additional costs
Flexibility	Limited to Google offerings	Fully customizable
Maintenance	Managed by Google	You manage credentials/updates
Best For	Web search, links, native features	Custom APIs, internal systems

Best Practices:

✅ Only enable tools the agent actually needs
✅ Test tool behavior thoroughly
✅ Monitor tool usage and errors
✅ Set appropriate max steps for your tools
⚠️ Too many tools can confuse the agent
⚠️ Some tools have rate limits or costs

Example Tool Combinations:

Customer Support
Data Analyst
Research Assistant

Enabled Tools:

API Request (check order status)
External Datasource (knowledge base)
Web Search (product updates)

Max Steps: 3-5Workflow:

User: "Where is my order #12345?"
Step 1: API Request → Query order system
Step 2: External Datasource → Get shipping info
Step 3: Format and respond with details

Enabled Tools:

BigQuery (database queries)
API Request (fetch external data)

Max Steps: 10-15Workflow:

User: "Show me sales trends for Q4"
Step 1: BigQuery → Fetch sales data
Step 2: Agent → Calculate trends
Step 3: Agent → Generate chart (JSON Format)
Step 4: Format and present analysis

Enabled Tools:

Web Search (Perplexity AI)
External Datasource (research papers)

Max Steps: 8-12Workflow:

User: "What are the latest developments in quantum computing?"
Step 1: Web Search → Find recent articles
Step 2: External Datasource → Search academic papers
Step 3: Agent → Extract key points
Step 4-5: Additional searches for specific topics
Step 6: Synthesize and summarize findings

Saving Changes

After configuring your agent settings:

Review your changes in each tab
Click “Update Agent” at the bottom of the settings panel
Wait for confirmation that settings were saved
Test your agent to ensure it behaves as expected

Important: Changes take effect immediately after saving. If your agent is in active use, consider testing changes in a development environment first.

Best Practices Summary

Agent Identity & Behavior

Use clear, descriptive names
Write detailed prompts with specific guidelines
Test initial messages for user engagement
Update prompts iteratively based on performance

Model Selection

Choose models based on your specific use case
Use RouteLLM for cost optimization
Enable streaming for better UX
Only use structured output when necessary
Monitor temperature settings and adjust for consistency

Datasources & Retrieval

Start with moderate Vector Top K (6-10)
Enable reranking for accuracy-critical applications
Only connect relevant datasources
Maintain high-quality datasource content
Monitor retrieval performance

Tools & Capabilities

Only enable necessary tools
Set appropriate max steps
Test tool chains thoroughly
Watch for circular references
Monitor tool usage and costs

Troubleshooting

Agent not responding

Possible Causes:

No LLM model selected
Model doesn’t support required features
Max steps set too low for complex tasks

Solutions:

Check that a model is selected in Model tab
Verify model compatibility with your settings
Increase max steps if using multiple tools

Irrelevant responses from datasources

Possible Causes:

Vector Top K too high
Rerank threshold too low
Poor quality datasource content

Solutions:

Reduce Vector Top K to 5-10
Enable reranking with threshold 0.2-0.3
Review and improve datasource content quality

Tool errors

Possible Causes:

Model doesn’t support tools
Circular agent references
Tool configuration issues

Solutions:

Switch to tool-capable model (GPT-4, Claude, Gemini)
Check for and remove circular references
Verify tool configurations and permissions

Streaming not working

Possible Causes:

Model doesn’t support streaming
Structured output is enabled

Solutions:

Verify model supports streaming
Disable structured output to use streaming

Next Steps

Test Your Agent

Learn how to thoroughly test your agent configuration

Analytics

Monitor your agent’s performance and usage

Tool Configuration

Learn more about configuring and creating tools

Datasources

Manage and optimize your datasources

​Overview

Agent Tab

Model Tab

Datasources Tab

Tools Tab

​Agent Configuration

​Avatar

​Name

​Description

​Name Slug

​Core Agent

​Prompt (System Instructions)

​Initial Message (Intro Message)

​Model Configuration

​Language Model Selection

​RouteLLM Configuration

​Threshold

​Strong Model Config

​Weak Model Config

​Enable Streaming

​Enable Language Detection

​Enable Citations

​Structured Output

​JSON Schema

​Temperature

​Max Output Tokens

​Max Steps (Tool Execution)

​Datasources Configuration

​Vector Top K

​Enable Rerank

​Rerank Top K

​Rerank Threshold

​Multi-Query RAG

​Datasources Selection

​Tools Configuration

​Tool Max Steps

​Tools Selection

​Google Models Built-in Tools

​Saving Changes

​Best Practices Summary

​Troubleshooting

​Next Steps

Test Your Agent

Analytics

Tool Configuration

Datasources

Overview

Agent Configuration

Avatar

Name

Description

Name Slug

Core Agent

Prompt (System Instructions)

Initial Message (Intro Message)

Model Configuration

Language Model Selection

RouteLLM Configuration

Threshold

Strong Model Config

Weak Model Config

Enable Streaming

Enable Language Detection

Enable Citations

Structured Output

JSON Schema

Temperature

Max Output Tokens

Max Steps (Tool Execution)

Datasources Configuration

Vector Top K

Enable Rerank

Rerank Top K

Rerank Threshold

Multi-Query RAG

Datasources Selection

Tools Configuration

Tool Max Steps

Tools Selection

Google Models Built-in Tools

Saving Changes

Best Practices Summary

Troubleshooting

Next Steps