New serverless pattern - bedrock-semantic-cache-s3vectors-sam

[manishh-13]

New serverless pattern: bedrock-semantic-cache-s3vectors-sam

Serverless semantic cache for Amazon Bedrock using AWS Lambda and Amazon S3 Vectors.

An AWS Lambda function sits in front of Amazon Bedrock and returns a cached answer whenever an incoming prompt is semantically similar to one already answered, skipping the LLM call on repeats and near-repeats.

How it works

• Lambda embeds the prompt with Amazon Titan Text Embeddings v2 (1024-dim).
• Queries an Amazon S3 Vectors index for the nearest prior prompt (cosine similarity); the cached answer is stored in the vector metadata, so no separate database is needed.
• HIT (similarity >= configurable threshold, within TTL, current epoch, passes negation-parity guard): returns the stored answer in ms at $0 LLM cost.
• MISS: calls the Bedrock text model, stores embedding + answer + metadata, returns the fresh result.

Correctness features

• Tunable similarity threshold (precision vs. hit rate).
• Freshness TTL to bound staleness.
• One-call force-invalidation via an epoch counter in SSM Parameter Store (no scan/delete).
• Negation-parity guard (prevents "is X" vs "is NOT X" false hits).
• Top-K retrieval so a stale near-neighbour never blocks a valid hit.

Notes

• Fully serverless: S3 Vectors scales to zero; Lambda is stateless (cache survives cold starts/redeploys).
• Access via IAM-signed Lambda function URL + optional app-level API key. Least-privilege IAM.
• Model-agnostic: the answer model is a drop-in for any Amazon Bedrock model (Amazon Nova, Anthropic Claude, Meta Llama, Mistral) via the converse API; the embedding model is configurable as long as the index dimension matches.
• Complements (does not replace) Bedrock native prompt caching, which is exact-prefix based.

Includes README.md, template.yaml (AWS SAM), example-pattern.json, and src/ (Python). Language: Python. Level: 300. Framework: AWS SAM.

Pattern-submission issue: #3263