Vector Search

Why it matters

Users and analysts ask questions in natural language; keyword search misses relevant chunks phrased differently. Vector search powers RAG retrieval and internal discovery across large knowledge corpora.

How it works

Text chunks embed into high-dimensional vectors. Queries embed the same way. Nearest neighbors in cosine or dot-product distance retrieve candidates; optional reranking improves precision. HNSW indexes make search fast at scale.

Common mistakes

• Embedding tiny chunks with no contextual headers.
• Never re-embedding after source documents change.
• Confusing vector similarity with factual correctness.
• Returning too many chunks and blowing token budgets.

Best practices

• Chunk with semantic boundaries and metadata titles.
• Cap retrieval count and rerank for precision.
• Monitor retrieval misses with logged failed analyst queries.
• Keep embedding model version consistent per corpus.

Learn Domains perspective

FAQ

Vector search vs keyword search?

Keywords match tokens; vectors match meaning, use both hybrid for best recall.

Which embedding model does Learn Domains use?

Learn Domains handles embedding and retrieval for you, you focus on keeping Knowledge Base sources accurate.

Can I export my embeddings?

Vectors power in-product retrieval for your org, export your source documents from the Knowledge Base instead.

Next steps

1. 1Add two Knowledge Base sources with distinct topics.
2. 2Query the AI Analyst using paraphrased wording.
3. 3Confirm retrieval pulls the correct chunk.