proposal-system/docs/adr/0001-bedrock-vector-store-aurora-pgvector.md

ADR 0001 — Bedrock Knowledge Base vector store: Aurora PostgreSQL + pgvector

• Status: Accepted (2026-06-12)
• Decision owner: Adam Moussa
• Scope: proposal-system infrastructure (foundation + compute stacks), v1 PR3

Context

The proposal system's RAG pipeline uses an Amazon Bedrock Knowledge Base (Titan Embed v2) over historical proposals. The original vector store was OpenSearch Serverless (AOSS), which:

• carries a minimum 2-OCU billing floor ($175–350/mo) even at near-zero query volume — the dominant line item of the monthly bill for an internal tool;
• is VPC-only, which forces the NAT gateway and a oss-index-creator bootstrap Lambda that exists purely to pre-create the vector index while an IAM access policy propagates.

The database is a small RDS PostgreSQL instance. The v1 assessment flagged AOSS as over-built for the corpus size and recommended a pgvector store on the existing database.

Decision

Migrate the database from RDS PostgreSQL 15 → Aurora PostgreSQL Serverless v2 and use pgvector as the Bedrock KB vector store.

Aurora is required because Bedrock Knowledge Bases support Aurora PostgreSQL Serverless v2 (with the RDS Data API) as a pgvector store, but not a plain RDS instance. A standard RDS instance cannot back a Bedrock KB, so "pgvector on the existing RDS" was infeasible without the Aurora move.

Implementation:

• Aurora Serverless v2 (min 0.5 / max 4 ACU), enableDataApi: true, defaultDatabaseName: 'proposals'. The cluster also serves the .NET API's application data (one database).
• A bootstrap custom-resource Lambda (aurora-pgvector-init, via the RDS Data API) enables vector, creates the bedrock_integration.bedrock_kb table (vector(1024) for Titan v2 + HNSW cosine + GIN indexes) and a dedicated bedrock_user role. This replaces oss-index-creator — the bootstrap is swapped, not eliminated.
• CfnKnowledgeBase.storageConfiguration → type: 'RDS' with rdsConfiguration (cluster ARN, bedrock_user secret, bedrock_integration.bedrock_kb, field mapping id/embedding/chunks/metadata). KB role IAM swaps aoss:APIAccessAll → scoped rds-data + secret read.
• All AOSS constructs (collection, policies, VPC endpoint, index-creator) are deleted.

Consequences

• Cost: eliminates the AOSS OCU floor (~$175–350/mo). Aurora Serverless v2 at 0.5 ACU min is ~$43/mo and scales toward zero idle — a net reduction.
• Simplification: one data engine (Aurora) instead of RDS + AOSS; fewer constructs. A bootstrap Lambda remains (now for pgvector schema rather than the AOSS index).
• NAT: kept for now — Lambdas and the KB's Data API path still need AWS-service egress. Dropping NAT would require VPC interface endpoints; tracked separately.
• Migration: the RDS→Aurora swap is a CloudFormation replacement. The deployed stacks are test-only with no production data, so this is a clean redeploy.

Alternatives considered

• Keep AOSS: rejected — the cost floor is the single biggest waste for the scale.
• S3 Vectors: viable Bedrock backend, but Aurora unifies app data + vectors and was the owner's preference (also cheaper than AOSS).
• pgvector on the existing RDS instance: infeasible — Bedrock KB does not support a plain RDS instance as a vector store.

References

• Using Aurora PostgreSQL as a Bedrock Knowledge Base
• Bedrock KB vector-store prerequisites