AI Retrieval-Augmented Generation Platform

Topology Notes

• ·PostgreSQL serves as both the vector store (pgvector extension) and the relational metadata store. Embeddings are stored in a vector column alongside document metadata (tenant_id, document_type, created_at, status). Hybrid queries combine ANN similarity with SQL WHERE predicates for metadata filtering.
• ·Redis provides a semantic cache keyed by query embedding hash (or approximate nearest cache key via cosine similarity threshold). Cache hits avoid both the embedding API call and the pgvector ANN scan. Cache TTL must account for document corpus changes: a document addition should invalidate or expire relevant cache entries.
• ·Kafka topic receives document ingestion events (document_id, raw_content). Embedding consumer reads from the topic, calls the embedding model API (e.g., OpenAI text-embedding, Cohere, or local model), and writes the resulting vector to the PostgreSQL vector column. The consumer must be idempotent: duplicate processing must not create duplicate embeddings.
• ·The embedding model is fixed at index creation time. Changing the embedding model requires rebuilding the entire vector index and invalidating all cached query embeddings. Model upgrades are a planned maintenance event, not a routine deployment.
• ·Vector index type selection is a permanent decision per table: IVFFlat for lower memory usage with periodic rebuilds, HNSW for better recall with higher memory footprint. Both require explicit creation after data is loaded: do not create the index on an empty table, as the index build quality depends on the data distribution.