AWS Bedrock KB Retrieval

Accepts free-form natural language queries and translates them into semantic search operations against Amazon Bedrock Knowledge Bases using the Bedrock Agents API. The MCP server acts as a bridge that converts client tool calls into RetrieveAndGenerate API invocations, handling query embedding, vector similarity matching, and result ranking through Bedrock's managed retrieval pipeline without requiring clients to manage embedding models or vector indices directly.

Registers Bedrock KB retrieval as a callable tool within the MCP protocol using the tools/list and tools/call message types, enabling LLM clients to discover the retrieval capability and invoke it with structured arguments. The server implements the MCP tool schema with input validation, error handling, and response formatting that conforms to MCP's tool response envelope, allowing seamless integration into agent decision-making loops without custom client code.

Supports querying across multiple Bedrock Knowledge Bases by accepting a knowledge base ID parameter in tool calls, allowing clients to specify which KB to query or implement routing logic. The server maintains a registry of available knowledge bases (discovered via Bedrock API or configuration) and routes each query to the appropriate KB, enabling use cases where different data sources are organized by domain, team, or data classification level.

Extracts and returns source document metadata (document name, location, retrieval confidence score, chunk ID) alongside retrieved content, enabling clients to trace answers back to original sources and assess retrieval quality. The server parses Bedrock KB response envelopes to surface metadata fields that clients can use for citation, audit trails, or relevance filtering, without requiring additional API calls to fetch source information.

Implements MCP-compliant error handling that catches Bedrock API failures (throttling, invalid KB ID, permissions errors) and returns structured error responses with diagnostic information, allowing clients to implement retry logic or fallback strategies. The server distinguishes between transient errors (throttling, temporary service issues) and permanent errors (invalid KB, permission denied) to guide client behavior, and includes error context that helps developers debug integration issues.

Validates incoming MCP tool call parameters (query string length, knowledge base ID format, optional filters) before sending to Bedrock API, preventing malformed requests and reducing unnecessary API calls. The server implements input validation rules (max query length, KB ID pattern matching, filter syntax) and returns validation errors to clients before attempting Bedrock calls, reducing latency and API costs for invalid requests.

Manages server initialization, configuration loading from environment variables or config files, and graceful shutdown. The server implements MCP server initialization protocol (capabilities negotiation, resource listing) and loads Bedrock credentials and KB configuration at startup, enabling deployment in containerized environments (Docker, Lambda, ECS) with standard configuration patterns. Supports environment-based configuration for AWS region, credentials, and KB metadata.

Implements MCP streaming protocol to return large knowledge base results in chunks rather than buffering entire responses, enabling clients to process results incrementally and reducing memory overhead. The server streams document chunks and metadata as they arrive from Bedrock, allowing clients to display results progressively and handle large result sets without loading everything into memory at once.