Bedrock Knowledge Base Metadata: Filter Retrieval by Language and Repo

A monorepo Knowledge Base without metadata filters will happily retrieve Python ETL snippets while you’re editing a TypeScript Lambda. Attach language, repository, service, and owner metadata at ingest, then pass retrieval filters from the IDE context so Bedrock Retrieve only searches the slice that matters.

⚡ TL;DR: Emit sidecar metadata JSON (or inline attributes) per chunk at sync time; map filterable fields in the KB; call Retrieve/RetrieveAndGenerate with vectorSearchConfiguration.filter; derive filters from open files / git root in the client. See Bedrock Retrieval Filters (tenants) and KB Sync webhooks.

Metadata at ingest

// s3://code-kb/payments/src/getOrder.ts.metadata.json
{
  "metadataAttributes": {
    "repo": "acme/backend",
    "service": "payments",
    "language": "typescript",
    "owner": "payments-platform",
    "path": "services/payments/src/getOrder.ts",
    "visibility": "internal"
  }
}

# sync_chunk.py — write chunk + metadata for Bedrock KB S3 data source
def write_chunk(bucket: str, key: str, text: str, meta: dict):
    s3.put_object(Bucket=bucket, Key=key, Body=text.encode("utf-8"))
    s3.put_object(
        Bucket=bucket,
        Key=f"{key}.metadata.json",
        Body=json.dumps({"metadataAttributes": meta}).encode("utf-8"),
        ContentType="application/json",
    )

Derive language from extension + package.json / pyproject.toml presence; derive service from path conventions (services/<name>/). Fail the sync job if required keys are missing—unfiltered docs are a production bug.

Filter at Retrieve time

import boto3
bedrock_agent = boto3.client("bedrock-agent-runtime")

def retrieve_for_ide(query: str, *, repo: str, language: str | None, service: str | None):
    and_all = [{"equals": {"key": "repo", "value": repo}}]
    if language:
        and_all.append({"equals": {"key": "language", "value": language}})
    if service:
        and_all.append({"equals": {"key": "service", "value": service}})

    return bedrock_agent.retrieve(
        knowledgeBaseId=KB_ID,
        retrievalQuery={"text": query},
        retrievalConfiguration={
            "vectorSearchConfiguration": {
                "numberOfResults": 8,
                "filter": {"andAll": and_all},
            }
        },
    )

Default filters from the active editor (languageId, nearest service root). Do not use global search as the default for “where is X implemented?” inside a 40-service monorepo.

Client context wiring

function filtersFromEditor(ctx: IdeContext) {
  return {
    repo: ctx.gitRemoteSlug,           // acme/backend
    language: mapVscodeLang(ctx.languageId),
    service: detectService(ctx.filePath), // services/payments/...
  };
}

Offer an explicit “search all services” toggle for cross-cutting questions (auth, platform libs). Log which filter mode was used for eval slices.

Pitfalls

Closing checklist

Dos
– Require repo, language, service metadata on every chunk
– Pass andAll filters from IDE context by default
– Canonicalize enum values at sync time
– Provide an explicit broad-search mode
– Include filter mode in RAG eval dimensions

Donts
– Do not ingest without sidecars “to move faster”
– Do not rely on path substrings in the query text instead of filters
– Do not mix tenant/customer data without tenant isolation filters
– Do not fail open (ignore filter errors)
– Do not forget to re-sync metadata after service renames

Related reading

• Bedrock Retrieval Filters: Tenant Isolation
• Bedrock Knowledge Base Sync
• Multi-Index RAG Design
• Query Rewriting for Code RAG

Last updated on September 11, 2026