Bedrock AI Agent

The Bedrock AI Agent provides conversational analytics with natural language queries and AI-powered recommendations.

Building the knowledge base

The knowledge base stores remediation playbooks as vector embeddings for semantic search.

Aurora PostgreSQL with pgvector

Cluster configuration: * Engine: Aurora PostgreSQL 15.4 * Instance: db.r6g.large (Graviton3) * Scaling: Serverless v2 (0.5-2 ACUs) * Multi-AZ: Enabled * Encryption: KMS at rest, TLS in transit

Vector table schema:

CREATE EXTENSION vector;
CREATE SCHEMA bedrock_integration;
CREATE TABLE bedrock_integration.bedrock_kb (
    id UUID PRIMARY KEY,
    embedding vector(1536),
    chunks TEXT,
    metadata JSONB
);
CREATE INDEX ON bedrock_integration.bedrock_kb
USING hnsw (embedding vector_cosine_ops);

Knowledge base content

Battery remediation playbook: * Supplier batch defects * Thermal management issues * BMS failures * Charging problems * Success rates and timelines

Customer churn analysis playbook: * Churn risk factors * Retention strategies by segment * Proactive engagement tactics * Win-back campaigns * ROI calculations

Analytics metadata guide: * Data schema and tables * KPI calculations * Metric thresholds * Query examples * Dashboard interpretation

Document format

Documents stored as markdown for rich context:

## Supplier Batch Defect Issues

### Early Battery Health Degradation

**Symptoms:**
- Battery health < 80% within 12 months
- Q2 2024 production batches
- Premium segment concentration

**Severity:** Critical

**Root Cause:**
Supplier defect in thermal management affecting 12,000 vehicles

**Recommended Action:**
1. Proactive battery replacement
2. Customer notification within 48 hours
3. Priority scheduling
4. Loaner vehicle provision

**Success Metrics:**
- Success Rate: 94%
- Resolution Time: 5 days
- Revenue Recovery: 87%
- Satisfaction: 4.2/5.0

Knowledge base ingestion

To populate the knowledge base, documents are uploaded to S3 and an ingestion job is started:

# Upload playbook documents
aws s3 sync source/knowledge-base/ \
  s3://cx360-knowledge-base-<ACCOUNT-ID>/docs/

# Start ingestion job
aws bedrock-agent start-ingestion-job \
  --knowledge-base-id <KB_ID> \
  --data-source-id <DS_ID>

Chunking strategy

Document chunking configuration: * 500 tokens per chunk * 50 token overlap * Keep related info together * Use markdown structure

These parameters balance retrieval precision with context completeness. A 500-token chunk is large enough to include the full context of a remediation step (symptom, root cause, recommended action, success metrics) without splitting related content across chunks. The 50-token overlap ensures that sentences near chunk boundaries appear in adjacent chunks, preventing retrieval gaps when a query matches content that spans a boundary.

Building action groups

Action groups enable the agent to query real-time customer data.

Action group 1: Query customer health

Purpose: Retrieve health scores, churn risk, lifetime value

Lambda function:

def query_customer_health(params):
    min_score = params.get('min_score', '0')
    max_score = params.get('max_score', '100')

    query = f"""
    SELECT customer_id, health_score, churn_probability
    FROM {GLUE_DATABASE}.customer_health
    WHERE health_score BETWEEN {min_score} AND {max_score}
    ORDER BY health_score ASC
    LIMIT 100
    """

    results = execute_athena_query(query)
    return {'customers': results}

Action group 2: Query at-risk revenue

Purpose: Calculate revenue at risk by segment

def query_at_risk_customers(params):
    segment = params.get('segment', 'all')

    query = f"""
    SELECT customer_type, at_risk_revenue, at_risk_customers
    FROM {GLUE_DATABASE}.at_risk_revenue_view
    WHERE month_date = (SELECT MAX(month_date)
                        FROM at_risk_revenue_view)
    """

    results = execute_athena_query(query)
    return {'segments': results}

Action group 3: Query sentiment trends

Purpose: Analyze sentiment and NPS trends

def query_customer_trends(params):
    months = params.get('months', '12')

    query = f"""
    SELECT month_label, nps_score, health_score
    FROM {GLUE_DATABASE}.kpi_trends
    ORDER BY month_date DESC
    LIMIT {months}
    """

    results = execute_athena_query(query)
    return {'trends': results}

Configuring the Bedrock agent

The agent orchestrates knowledge base retrieval and action group execution.

Agent configuration

Foundation model: Claude 3.5 Sonnet v2

Agent instructions:

You are a customer analytics expert helping automotive teams
understand customer sentiment, health trends, and churn risk.

Your capabilities:
1. Query customer data using action groups
2. Retrieve remediation playbooks from knowledge base
3. Analyze trends and identify root causes
4. Provide actionable recommendations

Guidelines:
- Always cite data sources
- Include specific metrics
- Reference relevant playbooks
- Prioritize by revenue impact
- Be concise but thorough

Adding knowledge base

aws bedrock-agent associate-agent-knowledge-base \
  --agent-id <AGENT_ID> \
  --knowledge-base-id <KB_ID> \
  --description "Customer remediation playbooks"

Adding action groups

aws bedrock-agent create-agent-action-group \
  --agent-id <AGENT_ID> \
  --action-group-name query-customer-health \
  --action-group-executor '{"lambda": "<LAMBDA_ARN>"}'

Configuring guardrails

PII filtering:

aws bedrock create-guardrail \
  --name customer-360-guardrail \
  --sensitive-information-policy-config '{
    "piiEntitiesConfig": [
      {"type": "EMAIL", "action": "BLOCK"},
      {"type": "PHONE", "action": "BLOCK"},
      {"type": "SSN", "action": "BLOCK"}
    ]
  }'

Using the AI agent

The agent provides conversational analytics through natural language.

Example conversation: Declining sentiment

User: "What's causing declining customer sentiment?"

Agent: Based on analysis of the past 12 months:

**Trend Analysis:**
- NPS declined from 52 to 42 (19% decrease)
- Health scores dropped from 65 to 56 (14% decrease)
- At-risk customers increased 73%

**Root Cause:**
Battery issues increased from 15% to 40% of cases,
primarily affecting premium segment (12,450 customers).

**Revenue Impact:**
$2.3M at risk from premium customers.

**Recommended Actions:**
From Battery Remediation Playbook:

1. **Proactive Battery Replacement**
   - Success Rate: 94%
   - Resolution Time: 5 days
   - Customer Satisfaction: 4.2/5.0

2. **Customer Communication Campaign**
   - Notify within 48 hours
   - Priority scheduling
   - Expected recovery: 87%

[Citations: Battery Remediation Playbook, KPI Trends]

Example conversation: Segment analysis

User: "Which customer segments are at highest risk?"

Agent: Analyzing at-risk revenue by segment:

**Premium Segment** (Highest Risk)
- At-Risk Revenue: $1.8M
- At-Risk Customers: 3,200
- Avg Lifetime Value: $85K
- Primary Issue: Battery degradation

**Standard Segment** (Medium Risk)
- At-Risk Revenue: $450K
- At-Risk Customers: 8,100
- Avg Lifetime Value: $42K
- Primary Issue: Service quality

**Recommendation:**
Focus on premium segment first due to highest
revenue impact and clear remediation path.

[Citations: At-Risk Revenue View, Customer Health]

Optimization and best practices

Agent instruction tuning

Iterative improvement: 1. Test with common questions 2. Review response quality 3. Refine instructions for clarity 4. Add examples for complex scenarios 5. Test edge cases

Instruction patterns:

# Good: Specific and actionable
"When analyzing declining metrics, always:
1. Query the last 12 months
2. Calculate month-over-month changes
3. Identify top 3 factors
4. Cite specific data points"

# Bad: Vague
"Analyze the data and provide insights"

Knowledge base optimization

Document quality: * Use clear section headers * Include specific metrics * Add real-world examples * Update regularly * Remove outdated information

Chunking strategy: * 500 tokens per chunk * 50 token overlap * Keep related info together * Use markdown structure

Cost optimization

Estimated monthly costs (us-east-1 region):

Cost optimization tips:

Aurora Serverless: * Auto-pause after 5 minutes idle * Min ACU: 0.5, Max: 2 * Use read replicas only if needed

Bedrock costs: * Input tokens: ~$0.003 per 1K * Output tokens: ~$0.015 per 1K * Embeddings: ~$0.0001 per 1K * Optimize prompt length

Athena costs: * $5 per TB scanned * Use partitioned tables * Leverage views * Set result limits