Building a Citation-Backed Policy
Assistant for Lexipol on Amazon Bedrock
Lexipol is a public-safety software company serving law enforcement, fire, EMS, and corrections agencies across the United States. Through its LexOne SaaS platform, Lexipol authors and continuously maintains agency-specific policy manuals grounded in current federal, state, and local legal standards, serving approximately 3,400 law enforcement agencies and 1,300 fire agencies. For the agencies it serves, Lexipol is the system of record for how policy is written, distributed, acknowledged, and operationalized, which makes accuracy and traceability core product requirements rather than nice-to-haves.
The Opportunity
Policy manuals are long, dense, and jurisdiction-specific. When an officer, supervisor, or agency administrator needed to answer a question (what the pursuit policy permits, what the use-of-force reporting requirement is, how a specific procedure has changed since the last revision), the workflow was manual document search. Finding the right passage across a multi-hundred-page manual took minutes, and the burden fell hardest at exactly the moments when speed mattered most.
The obvious remedy, a conversational assistant, carried a risk profile that ruled out a conventional chatbot. In public
safety, an ungrounded or hallucinated answer is not a usability defect. It’s a liability event. Any assistant had to answer only from the specific manual issued to the specific agency asking, cite the source passage for every claim, and never blend one agency’s policy into another’s response. Multi-tenant isolation was a legal requirement, not an architectural preference. If Lexipol shipped an assistant that failed any of these tests, the downside was reputational damage to a trust-dependent business and potential exposure for the agencies relying on it.
Lexipol needed a generative AI capability that was fast enough to change user behavior, grounded strongly enough to be defensible, and isolated strictly enough to be safe across thousands of tenants.
The Solution
Amazon Bedrock as the grounding and inference layer
Amazon Bedrock Knowledge Bases ingest policy manuals issued by Amazon S3 via a custom chunking pipeline implemented in AWS Lambda, preserving the structural boundaries of policy sections so that retrieved passages remain semantically complete. Amazon Titan Text Embeddings v2 generates the embeddings. Retrieval and generation run via the Bedrock RetrieveAndGenerate API using a citation-forced prompt template with streamed responses, so every answer surfaces the underlying policy passage, and the user sees output beginning immediately rather than waiting for full completion. Amazon Bedrock Prompt Management holds prompt configurations under central version control, and a tiered model configuration (fast, standard, and complex tiers mapped to different Anthropic Claude model sizes on Bedrock) is managed through AWS Systems Manager Parameter Store, allowing Lexipol to route simple lookups to lower-latency models and reserve larger models for multi-step reasoning.
Safety, isolation, and multi-turn behavior
A four-layer input and output guardrail system operates in production alongside evaluation of Amazon Bedrock Guardrails as a managed control. Amazon Bedrock AgentCore provides a Memory Store supporting multi-turn conversational agents, with AgentCore session, invocation, latency, and throttle metrics instrumented in production monitoring. Per-manual authorization filtering ties every retrieval result to the requesting user’s identity through Amazon Cognito, so an agency can only ever receive answers grounded in the policy manual issued to it. Bedrock model-invocation and Knowledge Base logging flow to Amazon CloudWatch, Amazon S3, and Amazon Redshift for auditability, and Bedrock Knowledge Base Evaluations run automated RAG quality assessment.
Fully serverless supporting architecture
The solution provisions no fixed-capacity compute, and every layer scales natively with request volume:
- AWS Lambda handles agent execution, orchestration, request authorization, and document chunking, deployed multi-AZ with provisioned and reserved concurrency.
- Amazon API Gateway fronts the service.
- Amazon SQS FIFO manages ordered ingestion work.
- Amazon Aurora PostgreSQL with Amazon RDS Proxy persists chat analytics and application state with pooled connections.
- Amazon DynamoDB scales on demand.
Operations and delivery
Observability spans Amazon CloudWatch and AWS X-Ray tracing across APIs and Lambda functions, with Bedrock invocation counts, p90/p95/p99 latency, time-to-first-token, and token consumption tracked on every request and surfaced through named cost and usage dashboards. Deployments promote through a Dev to Integration to QA to Staging to Production pipeline with per-environment approval gates, with CI authenticating to AWS through GitHub OIDC role assumption rather than long-lived credentials. Dualboot maintains ticketed runbooks for deployment, vector-store migration, and per-datastore disaster recovery.
The Impact
The assistant replaced multi-minute manual policy search with a conversational interface that returns a cited, agency-scoped answer at a warm p50 end-to-end response time of 12-14 seconds.
- Orchestration latency down 73% Dualboot reduced the portion of latency it directly controlled from approximately 6 seconds to approximately 1.6 seconds, measured through CloudWatch spans and AWS X-Ray and cross-validated in Datadog.
- Retrieval infrastructure cost down 57% A vector-store consolidation cut retrieval infrastructure cost by approximately $1,000 per month while preserving the Bedrock Knowledge Base architecture and citation guarantees, verified against AWS billing data.
- Beta expanded to 116 additional agencies tracked through product usage analytics and in-product feedback.
The capability has moved from an add-on feature to the platform’s core interaction model. Agencies now query policy conversationally rather than searching documents, and every answer stays traceable to the passage that produced it.
About the Partner
Dualboot Partners is a nearshore software product development firm of approximately 400 engineers, designers, and product leaders, building and scaling software for venture-backed and private-equity-owned companies across public safety, healthcare, financial services, and industrial sectors. Dualboot embeds full product teams alongside customer organizations rather than delivering discrete projects, and holds AWS Advanced Tier Services Partner status and the AWS Small and Medium Business (SMB) Competency.