Common use cases for VDBs (Recommendation, Chatbots)

Recommendation

In recommendation, each item (product, video, article) is embedded; you query with the user’s history (e.g. average of recently viewed vectors) or a single item and return the nearest neighbors as suggestions. “Users who liked this also liked…” and “similar items” are both nearest-neighbor queries over item embeddings. You can add filters (e.g. category, in-stock) so results are relevant and actionable. The same primitive—fast ANN over embeddings—powers this and the other use cases below.

Item-to-item recommendation uses one item’s vector as the query. User-to-item can use an aggregate of the user’s interaction vectors (e.g. mean or weighted sum) as the query vector. Cold start for new users or items can be handled with fallbacks (e.g. popular items, random) or with side-information embeddings.

Semantic search

Semantic search finds content by meaning. You embed the corpus (e.g. document chunks) and the query; the vector DB returns the nearest chunks. Use it for site search, knowledge bases, and “find docs like this.” Often combined with hybrid search (vector + keyword) and re-ranking for better relevance.

Chunking strategy and embedding model choice directly affect result quality. For long documents, splitting into overlapping passages and embedding each passage gives finer-grained retrieval than embedding the whole document once.

RAG and chatbots

In chatbots and RAG, you embed chunks of knowledge (docs, FAQs, past conversations), then for each user question you run semantic search to fetch the most relevant chunks and pass them as context to the LLM. That’s the RAG stack: retrieve with a vector DB, then generate with an LLM so answers are grounded in your data. Chatbots may also use the VDB for long-term memory (store and retrieve past interactions by similarity).

RAG quality depends on retrieval recall (are the right chunks in the top k?) and on context length limits of the LLM. Tuning k, chunk size, and re-ranking helps. For multi-turn chatbots, including recent conversation turns in the query or storing turn embeddings improves context relevance.

Other use cases

Other uses include anomaly detection (e.g. embed log lines or transactions; flag points far from “normal” clusters), long-term memory for agents, and LLM response caching (cache prompts/context by embedding and reuse cached responses for similar queries). Deduplication: embed items and find near-duplicates by small distance. The common thread is: turn data into vectors, store them in a vector database, and query by similarity.