Modify Agent Prompts

Modify Agent Prompts and Configuration

Agent Architecture

review-item-processor/
├── agent.py               # Prompt generators + agent execution
├── model_config.py        # Model registry with capabilities
├── tools/
│   ├── factory.py         # Dynamic tool creation
│   ├── knowledge_base.py  # Bedrock KB integration
│   ├── code_interpreter.py
│   └── mcp_tool.py        # MCP tool support
└── index.py               # Lambda handler entry point

Common Modifications

1. Modifying Document Review Prompts

File: agent.py

Key functions:

• get_document_review_prompt() - Main entry point
• _get_document_review_prompt_with_citations() - With citations
• _get_document_review_prompt_legacy() - Without citations

JSON Schema:

{
  "result": "pass" | "fail",
  "confidence": <number 0-1>,
  "explanation": "<detailed reasoning>",
  "shortExplanation": "<max 80 chars>",
  "pageNumber": <integer from 1>,
  "citations": ["<quoted text>", ...]
}

2. Modifying Image Review Prompts

File: agent.py, function get_image_review_prompt()

JSON Schema (Nova models add boundingBoxes):

{
  "result": "pass" | "fail",
  "confidence": <number 0-1>,
  "explanation": "<detailed reasoning>",
  "shortExplanation": "<max 80 chars>",
  "usedImageIndexes": [<list of indexes>]
}

3. Model Configuration

Environment Variables (set in CDK):

• DOCUMENT_PROCESSING_MODEL_ID, IMAGE_REVIEW_MODEL_ID, BEDROCK_REGION
• ENABLE_CITATIONS, ENABLE_CODE_INTERPRETER

Change models via CDK:

cdk deploy -c rapid.documentProcessingModelId="global.anthropic.claude-opus-4-5-20251101-v1:0"

Add new model to model_config.py in _get_model_configs():

"model-id": ModelConfig(
    model_id="model-id",
    supports_document_block=True,
    supports_citation=True,
    supports_caching=True,
    input_per_1k=0.XXX,
    output_per_1k=0.XXX,
)

4. Adding New Environment Variables

1. Define in cdk/lib/parameter-schema.ts 2. Pass to Lambda in cdk/lib/constructs/agent.ts 3. Read in agent.py with os.environ.get()

5. Adding New Tools

For detailed tool creation guide, see references/TOOL-CREATION.md.

Quick Reference

Troubleshooting

After Modification

1. Run /build-and-format if Python code changed 2. Run /deploy-cdk-stack if CDK changes made 3. Test with sample documents and monitor CloudWatch logs

Agent Tool Creation Guide

Reference for adding new tools to the review-item-processor agent.

Step 1: Create Tool Implementation

Create new file in review-item-processor/tools/:

# tools/my_new_tool.py
from strands.types.tools import AgentTool

def create_my_new_tool(config: dict) -> AgentTool:
    """Create custom tool with configuration"""

    def tool_function(param1: str, param2: int) -> str:
        """
        Tool description shown to the agent.

        Args:
            param1: Description of parameter 1
            param2: Description of parameter 2

        Returns:
            Result description
        """
        result = perform_action(param1, param2)
        return result

    return AgentTool.from_function(
        tool_function,
        name="my_new_tool",
        description="What this tool does"
    )

Step 2: Register Tool in Factory

Edit tools/factory.py:

from tools.my_new_tool import create_my_new_tool

def create_custom_tools(tool_config: Dict[str, Any]) -> List[AgentTool]:
    tools = []
    if tool_config.get("myNewTool"):
        tools.append(create_my_new_tool(tool_config["myNewTool"]))
    return tools

Step 3: Update Tool Configuration Schema

If tool needs configuration from database/event, update:

• Database schema (backend Prisma schema)
• API validation (backend routes)
• Frontend types and UI

Step 4: Add Tool Usage Instructions to Prompts

Edit prompt generators in agent.py:

# In _build_tool_usage_section() or directly in prompt
tool_instruction = """
## WHEN TO USE MY_NEW_TOOL
- WHEN you need to do X -> USE my_new_tool
- WHEN you need to verify Y -> USE my_new_tool
"""

Tool Configuration Structure

tool_configuration = {
    "knowledgeBase": [
        {
            "knowledgeBaseId": "KB123",
            "name": "Building Regulations",
            "dataSourceIds": ["DS456"]
        }
    ],
    "codeInterpreter": True,
    "mcpConfig": {
        "servers": {
            "web-search": {
                "command": "npx",
                "args": ["-y", "@modelcontextprotocol/server-web-search"]
            }
        }
    }
}

Available Tools

• file_read - Read document content
• image_reader - Read image content
• knowledge_base_query - Query Bedrock Knowledge Bases
• code_interpreter - Execute Python code
• MCP tools - Dynamic tools from MCP servers
• Custom tools - Your own tool implementations