← all topics

🧮 Embeddings & Vector DBs

Embedding models, chunking, hybrid retrieval, vector store choice (Pinecone, Qdrant, Weaviate, pgvector), reranking.

6 videos · embeddingsvector-searchrecsysmultimodalvector-dbpostgres

The workflow

flowchart LR
    A[Documents] --> B[Chunking<br/>strategy & overlap]
    B --> C[Embedding model<br/>OpenAI / BGE / Cohere]
    C --> D[Vector store<br/>Pinecone / Qdrant / pgvector]
    D --> E[Hybrid search<br/>dense + sparse]
    E --> F[Filter + rerank]
    F --> G[Top-k results]

Embedding choice is rarely the bottleneck; chunking and reranking are.

Key takeaways

Recsys is moving from hash-based IDs to learned multimodal semantic IDs that share representational space with LLMs.
You probably don't need a dedicated vector DB — PG Vector inside the Postgres you already run gets you 90% of the way with RLS and joins.
Treating embedding space as a steerable canvas with discoverable semantic directions unlocks UI patterns prompts alone can't deliver.
Co-locate embeddings and business data in MongoDB documents so RAG can blend transactional context with vector search via one query language.
Production AI search now means domain-specific embeddings, structured-filter hybrid retrieval, and agentic query decomposition — not one-shot vector search.
Treat vendor vector-search benchmarks as marketing — only trust automated, reproducible, recall-anchored benchmarks across realistic read-write workloads.