Overview

vector = embed(json.dumps(record))

• Measure what you are embedding: count tokens and work out how much of a chunk is syntax rather than content.
• Separate embedded text from metadata: decide which fields users search for in words and which belong in a filterable column.
• Spot silent truncation: recognise when a wrapper pushes a record past an encoder’s input limit.
• Model pipeline cost: separate one-off indexing from the recurring cost of chunks pasted into prompts.
• Evaluate retrieval quality: use recall@1, recall@3 and mean reciprocal rank, and judge which differences matter.
• Recognise the exceptions: identify where structure belongs in the embedded text, as with code and log data.

• Developers building RAG systems who want to reduce token spend without changing models.
• Data scientists and ML engineers new to embeddings and vector databases.
• Engineers whose retrieval pipeline costs more, or performs worse, than expected.
• Students who want practice measuring a system rather than trusting its performance.

• Basic Python knowledge. Familiarity with embeddings helps but is not required.
• No API key and no vector database. The lab runs end to end offline with a small local model.