AI System Design Language

Option C is the professional architecture description. It uses precise technical vocabulary:

• Embedding vector — the numeric representation of the query for similarity search
• Semantic similarity search — finds documents by meaning, not keyword match
• Vector store / pgvector — names the specific technology (important for concrete conversations)
• Top-k — a standard parameter meaning "return the k most similar results"
• Context window injection — explains how retrieval output is fed to the LLM
• Grounding context — the industry term for retrieved documents used to anchor the response
• Pre-training knowledge — contrasts RAG with purely parametric (in-weights) knowledge

Standard RAG vocabulary for system design discussions:
— Chunking strategy (fixed-size, semantic, recursive) — how documents are split
— Embedding model (e.g. text-embedding-ada-002) — converts text to vectors
— Reranker — a second pass model that re-orders retrieved chunks by relevance
— Hybrid search — combining vector similarity with keyword search (BM25)

Option B correctly defines both terms and their relationship.

Vector database: a specialised database for storing and querying high-dimensional numeric vectors (embeddings). It supports approximate nearest-neighbour (ANN) search — finding the vectors most similar to a query vector. Examples:
• Pinecone — managed, cloud-native
• Weaviate — open-source, graph-capable
• Qdrant — open-source, Rust-based
• pgvector — Postgres extension (simplest to add if already using Postgres)
• Chroma — lightweight, developer-friendly

Inference pipeline (in the context of a production AI feature): the sequence of steps that runs when a user makes a request:
1. Query preprocessing
2. Embedding generation
3. Vector store retrieval
4. Context assembly (injecting chunks into the prompt)
5. LLM API call (the "inference" step properly)
6. Response post-processing
7. Response delivery

The colleague is proposing to add a vector database to step 3 — the retrieval step — which is a core RAG architecture decision.