AI Commons

Bot Configuration Quick Reference

Quick reference guide for configuring AI Commons bots. For detailed technical documentation, see the Create Bot Technical Specification.

February 2025 Update: This guide has been updated with 31 audit corrections. See the Changelog for details on all changes.

Basic Configuration

Field Required Best Practice
Title ✓ Yes Use descriptive, unique titles. Whitespace-only titles are not accepted.
Description ✗ No Always provide. Explain purpose, audience, and scope.
Instructions ✗ No Required for intended behavior. Define role, scope, tone, and format.

Instructions Templates

Support Bot Template

You are a helpful support assistant for [ORGANIZATION/PRODUCT].

Your role:
- Answer questions about [TOPIC]
- Provide clear, concise explanations
- Direct users to official resources when needed

Response style:
- Professional and friendly
- Use bullet points for multi-step processes
- Include relevant links or references

Scope:
- IN SCOPE: [List topics]
- OUT OF SCOPE: [List topics]

When asked about out-of-scope topics:
"I'm specifically designed to help with [SCOPE]. For questions about [OUT_OF_SCOPE],
please contact [RESOURCE]."

Source priority:
- Use configured knowledge base as primary source
- Only use Internet Search for current/external information not in knowledge base
- Always cite sources when providing information

Classification Bot Template

You are a classification assistant that categorizes [INPUT_TYPE].

Your task:
1. Read the provided [INPUT_TYPE]
2. Classify it into ONE of these categories:
   - Category A: [Description]
   - Category B: [Description]
   - Category C: [Description]
   - Category D: [Description]

Output format:
Category: [Selected Category]
Confidence: [High/Medium/Low]
Reasoning: [Brief explanation of why this category was chosen]

Classification rules:
- Choose the MOST SPECIFIC applicable category
- If multiple categories apply, choose based on [PRIORITY_RULE]
- If none apply, use category "Unclassified" and explain why

Research Assistant Template

You are a research assistant helping with [DOMAIN] research.

Your capabilities:
- Search configured knowledge sources for relevant information
- Use Internet Search for current external research
- Synthesize information from multiple sources
- Provide citations for all claims

Response approach:
1. Search internal knowledge base first
2. If additional context needed, use Internet Search
3. Synthesize findings into clear summaries
4. Always provide source citations

Output format:
**Summary:** [Brief overview of findings]

**Key Points:**
- [Point 1 with citation]
- [Point 2 with citation]
- [Point 3 with citation]

**Sources:**
[List all sources used]

Quality standards:
- Base answers on verifiable sources
- Distinguish between facts and interpretations
- Acknowledge uncertainty when sources conflict
- Never fabricate citations or data

Generation Settings Quick Reference

Use Case Temperature Top P Max Tokens
Factual Q&A 0.1 - 0.3 0.999 500 - 1000
Code Generation 0.2 - 0.4 0.999 1000 - 2000
General Conversation 0.5 - 0.7 0.999 1000 - 1500
Creative Writing 0.8 - 1.0 0.999 2000 - 4096
Classification 0.1 - 0.2 0.999 100 - 300
Summarization 0.3 - 0.5 0.999 500 - 1000

Model Selection Guide

High-Volume Simple Tasks

Recommended: Amazon Nova Micro, Claude Haiku

Use for: Classification, simple Q&A, routing, basic support

Cost: Lowest

General Purpose

Recommended: Claude Sonnet, Amazon Nova Lite

Use for: Standard conversations, moderate complexity analysis

Cost: Medium

Complex Reasoning

Recommended: Claude Opus, Claude Sonnet

Use for: Deep analysis, complex problem-solving, research

Cost: Higher

Multimodal

Recommended: Claude 4.6 models, Nova Pro

Use for: Image analysis, document processing, video understanding

Cost: Varies by model

Multilingual

Recommended: Qwen3 32B

Use for: Non-English content, translation, multilingual support

Cost: Medium

Code Generation

Recommended: Claude Sonnet, GPT-OSS 120B

Use for: Code writing, debugging, technical documentation

Cost: Medium to high

Knowledge Base Configuration

Chunking Strategy Selection

Strategy Best For Considerations
Fixed-Size Uniform content, technical docs Simple, predictable, may split concepts
Semantic Articles, reports, narratives Preserves context, variable chunk sizes
Hierarchical Complex docs, nested information Best accuracy, more processing

Retrieval Settings

Content Type Max Results Search Mode
Short FAQs 3-5 Hybrid
Long Documents 5-8 Semantic
Technical Docs 5-10 Hybrid
Policies/Procedures 3-5 Hybrid

Agent Tools Configuration

Internet Search

When to Enable

  • Need current information
  • External research required
  • Supplementing knowledge base

When to Disable

  • Internal-only information
  • Strict source control needed
  • Policy/compliance bots

Bedrock Agent

When to Enable

  • Need to execute actions
  • Complex orchestration
  • AWS service integration
  • Custom business logic

Security Notes

  • Verify IAM permissions
  • Review agent actions
  • Test thoroughly
  • Monitor usage

Guardrails Configuration

Recommended Guardrail Settings by Bot Type

Bot Type Content Filters Grounding Check Relevance Check
Public-Facing Medium-High ✓ Enabled ✓ Enabled
Internal Support Low-Medium ✓ Enabled ✓ Enabled
Research Assistant Low Optional ✓ Enabled
Policy/Compliance Medium ✓ Enabled (High) ✓ Enabled
Creative Assistant Low-Medium ✗ Disabled Optional
Cost Impact: Guardrails add processing time and cost. Configure based on actual risk profile rather than enabling all options by default.

Common Issues & Solutions

Bot gives incorrect answers

  • ✓ Check knowledge base contains correct information
  • ✓ Increase max retrieved results (try 8-10)
  • ✓ Review chunking strategy (try semantic or hierarchical)
  • ✓ Enable grounding check to prevent unsupported claims
  • ✓ Review Instructions for clarity and conflicts

Bot response is too slow

  • ✓ Switch to faster model (Nova Micro, Haiku)
  • ✓ Reduce max retrieved results
  • ✓ Reduce max_tokens setting
  • ✓ Disable unnecessary guardrails
  • ✓ Optimize knowledge base size

Bot goes off-topic

  • ✓ Add explicit scope to Instructions
  • ✓ Enable relevance check guardrail
  • ✓ Define out-of-scope handling in Instructions
  • ✓ Reduce temperature (try 0.3-0.5)

Bot can't find information

  • ✓ Verify documents uploaded correctly
  • ✓ Check parsing method appropriate for doc type
  • ✓ Try different chunking strategy
  • ✓ Increase max retrieved results
  • ✓ Switch search mode to Hybrid

Bot hallucinates sources

  • ✓ Enable grounding check (High sensitivity)
  • ✓ Lower temperature (0.1-0.3)
  • ✓ Add explicit instruction: "Never invent citations"
  • ✓ Enable citation display to make sources visible

Tool use not working

  • ✓ Verify model supports tool use
  • ✓ Check tool is properly configured (IDs, credentials)
  • ✓ Review Instructions for tool usage guidance
  • ✓ Test with simpler query requiring tool

Pre-Deployment Testing Checklist

Before deploying to production, verify all items below:

Configuration Validation

  • ☐ Title is descriptive and unique
  • ☐ Description accurately reflects bot purpose
  • ☐ Instructions define scope and behavior clearly
  • ☐ Knowledge base contains all required documents
  • ☐ Appropriate model(s) activated for use case
  • ☐ Generation settings appropriate for task type
  • ☐ Guardrails configured for risk level

Functional Testing

  • ☐ Bot answers in-scope questions correctly
  • ☐ Bot appropriately handles out-of-scope questions
  • ☐ Citations/sources are accurate (if enabled)
  • ☐ Response tone matches intended style
  • ☐ Response format follows Instructions
  • ☐ Tool use works correctly (if enabled)
  • ☐ Internet Search provides relevant results (if enabled)
  • ☐ Bedrock Agent integration works (if configured)

Quality Assurance

  • ☐ Tested with 20+ representative queries
  • ☐ No hallucinated information or sources
  • ☐ Response latency acceptable for use case
  • ☐ Multiple models tested (if multiple activated)
  • ☐ Edge cases handled appropriately
  • ☐ Error handling tested (malformed queries, etc.)

Security & Compliance

  • ☐ No sensitive information in Name/Description
  • ☐ Guardrails appropriate for exposure level
  • ☐ Knowledge base contains only approved content
  • ☐ Tool permissions reviewed and appropriate
  • ☐ Citation behavior matches policy requirements

Additional Resources

Technical Specification

Detailed technical documentation covering all configuration options and behaviors.

Available Models

Complete model reference with capabilities, context windows, and use cases.

Bot API Reference

API documentation for programmatic bot interaction and conversation management.

Code Examples

Production-ready code samples in Python, JavaScript, and cURL.