A production AI platform on Amazon Bedrock: RAG pipelines, agent orchestration, guardrails, multi-tenant cost tracking, and the observability needed to run AI in production.
4–8 weeks
$25k–$55k
Amazon Bedrock provides a unified API for accessing multiple foundation models — Anthropic Claude, Amazon Nova, Meta Llama, Cohere, Mistral, and others — without the data leaving your AWS account. Key enterprise properties:
Privacy by default: model invocations are processed within your AWS account infrastructure. Using Bedrock's VPC endpoint (PrivateLink), all traffic between your application and the model endpoint stays within the AWS network — no public internet egress. AWS does not use your prompts or completions to train or improve foundation models.
Compliance: Bedrock is covered by the AWS Business Associate Agreement (BAA), making it eligible for HIPAA workloads. Covered under AWS SOC 2, ISO 27001, and PCI DSS certifications. Available in AWS GovCloud (US) for government workloads.
AWS-native integration: IAM for access control (no separate API keys), CloudTrail for audit logging of every model invocation, CloudWatch for cost and performance metrics, and cost allocation tags for per-team/per-tenant tracking.
Models on Bedrock now arrive and retire on a published schedule — Claude Sonnet 4 reached end of life on October 14, 2026, and the Claude 3.x family, Opus 4.1, most Llama and Mistral models and Claude Sonnet 4.5 all end between November 2026 and April 2027. So we choose models on your own evaluation set, keep model IDs in configuration, and record each model's end-of-life commitment the day it is adopted. Our starting points as of October 2026:
Claude Sonnet 5.5: the default for customer-facing assistants, coding, document analysis and agents. On Bedrock since September 28, 2026, with a 1M-token context window, 128K output tokens and adaptive reasoning whose effort you set per use case.
Claude Opus 5.5: the hardest reasoning and long-running agent work, where quality matters more than cost per request.
Claude Haiku 5.5: high-volume classification, extraction and routing, where speed and cost dominate.
Amazon Nova 2 Lite: a low-cost AWS-built option worth including in every evaluation for simple, high-volume tasks.
OpenAI models on Bedrock: GPT-6 and GPT-5.x models and the open-weight gpt-oss family are now available through Bedrock, so choosing an OpenAI model no longer means leaving AWS's security and billing boundary.
Open-weight models: for new work, prefer the current generation — Gemma 4, Qwen3, Mistral Large 3, GLM 5 — over Llama 3.x/4, Mixtral and DeepSeek-R1, which are Legacy and end on March 30, 2027.
Retrieval Augmented Generation (RAG) allows LLMs to answer questions using your private documents and data. The architecture:
Document ingestion pipeline: documents (PDFs, Word, HTML, Confluence pages, S3 files) are ingested, chunked, embedded using an embedding model, and stored in a vector store.
Chunking strategy: fixed-size chunking (simple, predictable) vs semantic chunking (breaks on sentence/paragraph boundaries, better coherence) vs hierarchical chunking (parent chunks for retrieval context, child chunks for precise matching). For most enterprise document QA, semantic chunking with 512–1024 token chunks performs well.
Vector stores: Bedrock Knowledge Bases can manage the vector store — Amazon S3 Vectors (the lowest-cost option, generally available since December 2025), OpenSearch Serverless when you need hybrid keyword and vector search, or Aurora PostgreSQL with pgvector. For teams needing more control, self-managed pgvector on RDS or a dedicated vector database (Pinecone, Weaviate) are alternatives.
Retrieval: at query time, the user's question is embedded and the top-K most similar chunks are retrieved from the vector store. These chunks are injected into the LLM's context window along with the user's question.
Generation: the LLM generates an answer grounded in the retrieved context. Bedrock Guardrails' grounding check can detect when the model's response goes beyond what the retrieved context supports.
The original Bedrock Agents service — now called Agents Classic — entered maintenance on June 30, 2026: existing agents keep running, but new customers cannot adopt it and it gets no new features. New agent work on AWS runs on Amazon Bedrock AgentCore, generally available since October 2025, with agents written in any framework:
Runtime: isolated sessions per user, executions of up to eight hours, Agent-to-Agent protocol.
Gateway: turns your APIs and Lambda functions into MCP tools, and connects to existing MCP servers, with IAM or OAuth.
Identity and Policy: agents act on behalf of users with vaulted tokens, and every tool call is checked against rules before it runs.
Memory and Observability: managed short- and long-term memory, and end-to-end traces in CloudWatch or your OpenTelemetry tool of choice.
If you run agents on Agents Classic today, plan a move to AgentCore while it is a choice rather than a deadline.
Bedrock Guardrails apply content safety policies at the platform level — before responses reach your application code:
Content filters: block harmful content categories (hate speech, violence, sexual content) with configurable sensitivity levels.
Denied topics: custom topic restrictions specific to your use case (e.g., "do not provide specific investment advice," "do not discuss competitor products").
Word filters: block specific words or phrases.
PII redaction: automatically detect and redact personally identifiable information in both prompts and responses — configurable per data type (email addresses, phone numbers, SSNs, credit card numbers).
Grounding check: detect when model responses include information not supported by the provided context — critical for RAG applications where hallucination is a risk.
For SaaS applications serving multiple tenants on a shared Bedrock platform:
Per-tenant knowledge bases: each tenant has a separate Bedrock Knowledge Base with their own vector store. Strongest isolation, highest cost.
Shared knowledge base with metadata filtering: all tenants share a knowledge base, but documents are tagged with tenant IDs and retrieval queries include a metadata filter. More cost-efficient but requires careful filtering to prevent cross-tenant data leakage.
Per-tenant cost tracking: use Bedrock's cost allocation tags and CloudWatch metrics to track token usage per tenant. Build a token budget system that alerts when a tenant approaches their limit.
CloudWatch metrics: Bedrock emits metrics for invocations, input/output tokens, latency, and error rates per model. Create dashboards showing cost per model, latency percentiles, and error rates.
CloudTrail: every Bedrock API call is logged — who invoked which model, when, from which service/role. Critical for compliance audits.
Model evaluation: Bedrock Evaluations allows you to run systematic quality assessments of your RAG pipeline — relevance, groundedness, completeness, and coherence — using both automated metrics and human raters.
Prompt versioning: Bedrock Prompt Management provides versioning and A/B testing for prompts — track which prompt version produces better results, roll back to previous versions, and measure the impact of prompt changes.
Bedrock VPC endpoint (PrivateLink) configuration
RAG pipeline: ingestion pipeline, vector store (Knowledge Bases or self-managed), retrieval + generation
Agents on Amazon Bedrock AgentCore, with tools through Gateway and policy on every tool call
Guardrails configuration aligned to your content policy
Streaming API + WebSocket delivery layer for real-time response streaming
Per-tenant cost tracking and budget alerting
Bedrock Evaluations pipeline for ongoing RAG quality measurement
Runbooks: model updates, guardrail tuning, Knowledge Base re-indexing
Tell us about your current environment and goals. We'll review the details and respond with a proposed engagement structure and timeline.
Typical response time: 1–2 business days.
Prices and timelines are estimates for typical engagements, not offers. Scope, price and terms are set in a written agreement after we understand your situation. Terms of use.