Need a Knowledge Base Retrieval Tool to Provide Supplemental Grounding #44
Description
Knowledge Base Retrieval Tool
Summary
A new tool that lets agents retrieve grounding context from a curated knowledge base of semantic terms, taxonomies, and domain definitions. Agents invoke it at query time to pull authoritative chunks, improving terminology consistency and factual accuracy.
No foundation analyzer code — this is purely retrieval.
Components
CDK Stack
Defines the full deployment unit. Creates the Lambda function, wires IAM permissions for Bedrock (embeddings) and OpenSearch Serverless (vector search), and registers the tool with the AgentCore Gateway. Single stack, single cdk deploy.
Lambda Function
Entry point for tool invocations. Receives a query from the gateway, calls the embedder, queries the vector index, filters by score threshold, and returns ranked chunks with metadata. ARM64, Node 20, ~512MB memory, 15s timeout.
Config
Centralizes tunable defaults: tool name/description, embedding model ID (amazon.titan-embed-text-v2:0), index name, top-k (default 5), score threshold (default 0.55), Lambda sizing. All overridable via stack props.
Embedder
Thin client around Bedrock InvokeModel. Takes a query string, returns a normalized vector. Uses Titan Embed Text v2 at 256dimensions.
Retriever
Handles OpenSearch Serverless kNN queries with SigV4-signed requests. Supports optional metadata filters (e.g. { "domain": "finance", "taxonomy": "GL" }). Returns content, score, metadata, and source reference per hit.
Gateway Registration
Registers knowledge_base_retrieve as a tool in the AgentCore Gateway with its input schema and Lambda ARN. Agents see the tool name, description, and schema — standard tool-use contract.
Tool Schema (what agents see)
| Parameter | Type | Required | Description |
|---|---|---|---|
| query | string | yes | Natural-language query to search the KB with |
| top_k | number | no | Override number of results (default 5) |
| filter | object | no | Metadata key-value filter (e.g. domain, taxonomy) |
Response Shape
Each invocation returns a list of results, each containing:
- content — the retrieved text chunk
- score — similarity score
- metadata — domain, taxonomy, source doc, chunk ID, etc.
- source — reference back to the originating document
Infrastructure Dependencies
- OpenSearch Serverless collection (vector index, pre-existing)
- Bedrock access for Titan Embed Text v2
- AgentCore Gateway (pre-existing, receives tool registration)