For business
What it is
Each passage is turned into a list of numbers that captures its meaning, so similar ideas can be found even when the words differ.
Why it matters
“Wants to stop working early” can match “retire at 60”. That is what makes search over meetings feel intelligent.
Risk without it
Re-embedding everything on every change gets expensive fast, and switching models without a plan breaks search.
For engineers
How it works
Chunks are embedded in batches; results are cached by content hash, so unchanged text is never paid for twice. The model name and version are stored with every vector, so an index is only ever rebuilt deliberately.
Typical stack
- OpenAI / Cohere / Bedrock embeddings
- batch APIs
- content-hash cache
Pitfalls & trade-offs
- Never mix vectors from two models in one index.
- Batch for throughput and back off on rate limits.
- Clean text the same way at index time and query time.
In this example
In
→ c02 income
Out
[0.021, -0.113, 0.087, … ] cache: miss → stored
Fictional sample · not a client system