Embeddings

Embeddings are dense numeric vectors that represent text, images, or other objects so that semantic similarity approximates distance in vector space. A model maps "refund policy" and "money back guarantee" to nearby points even without shared keywords.

Typical flow: choose an embedding model (OpenAI, Cohere, open-source sentence transformers), chunk source content, batch-encode offline or on write, store vectors with metadata (source id, tenant, ACL), and query by encoding the user question and finding nearest neighbors.

| Concern | Implication | |---------|-------------| | Model version | Re-embed entire corpus on model change | | Dimension | Storage and index memory scale linearly | | Language/domain | General models may miss niche vocabulary | | Cost | API calls vs self-hosted GPU |

On interviews: explain embeddings as learned semantic coordinates—not magic; describe when keyword search still wins (SKUs, exact codes); mention normalization and distance metric choice.

Common pitfalls: tiny chunks with no context; stale embeddings after content edits; mixing models in one index; ignoring metadata filters for multi-tenant data; assuming embeddings replace evaluation.

The trade-off is semantic recall for paraphrases and concepts versus model cost, reindex burden, opacity of relevance, and the need for hybrid keyword + vector retrieval in production.

Checklist:

• Define embedding as fixed-size semantic vector.
• Name chunking, model choice, and re-embed triggers.
• Contrast with inverted-index keyword search.
• State distance metric and normalization.
• Tie vectors to metadata and access control.