A content-generation app needs to find useful passages across Google Drive and Confluence, even when a user’s wording differs from the documents. Vectorization turns text into numeric embeddings for semantic search: an embedding model places similar meanings close together and different meanings farther apart. The resulting representations can be stored and searched to retrieve relevant content. section 7.2.2
Similar wording, nearby representations
“I adore my dog,” “I love my dog,” and “I like my dog” cluster together in an embedding space. In a document-search trial, pairs of Confluence pages judged similar by different models were inspected by eye; MiniLM gave the most convincing results among the options tested. That quick check helped choose a model for a baseline, not establish that retrieval was fully reliable. section 7.2.2
Treat vectorization as one decision in a retrieval pipeline:
- Choose what to represent. Document lengths vary, so split content into retrievable chunks; sentence, semantic, or hierarchical chunking may suit different document types. section 7.2.2 section 7.2.4
- Compare embedding models against representative content, then store the resulting embeddings in a database for search. Open-source models offer a route to lower cost and more customization; fine-tuning can wait until the end-to-end pipeline works. section 7.2.2
- Refine retrieval when results miss the mark. Adjust chunk size and embedding models, and consider metadata filters or lexical search for more precise matches. section 7.2.4 section 7.4
In retrieval-augmented generation, retrieved documents provide context for a language model’s response. A good-looking cluster of embeddings is therefore only an early signal: assess whether retrieved context is relevant and whether generated answers are grounded and relevant to the question. section 7.2.2 section 7.4
