SmartZees
A shop assistant who carries the whole warehouse to the counter, every single time
SmartZees is a conversational commerce platform: you say what you need in plain language and it does the rest. Groceries, home services, community management. I built the backend, the retrieval layer and the infrastructure all three run on.
- Role
- Backend, RAG and infrastructure
- Timeline
- 2026, ongoing
- Agents shipped
- Three, all in production
- Live
- smartzees.com
Every message to the grocery agent re-read all 25,631 product rows and 206 MB of stored vectors out of MySQL, then threw them away. Replies took six to fifteen seconds. Against the production database it was twenty to twenty-three, and at one point the remote server simply dropped the connection mid-query under the weight of it.
A conversation cannot survive that. If the assistant takes fifteen seconds to answer "do you have milk", nobody has a second conversation with it.
One platform, three fronts
The three agents share a retrieval core, an intent layer and a deployment pattern. What differs is the domain, the data store and what "done" means for the user.
SmartMarket
Order groceries by talking. Searches a 25,631 item catalog, keeps the cart, and speaks its replies out loud.
- MySQL
- Voice STT/TTS
SmartService
Describe what has gone wrong and it finds the trade, shows the price, and books a real appointment slot.
- Postgres 18
- pgvector
SmartCommunity
Answers residents from their own community's documents only, issues parking passes, serves forms on request.
- Scoped RAG
- QR passes
Learn the catalog once in the morning
I replaced the per-query database read with an index built once at startup: the vectors stream into a pre-allocated float32 matrix, the metadata rides alongside as interned strings to keep the memory flat, and a rebuild swaps in atomically under a lock so no request ever sees a half-built index.
The build runs on a background thread, so the health endpoint keeps answering in a fraction of a second while a restart is still warming up. The original database path stayed in the codebase behind a single environment flag, which made it both the instant revert and the honest "before" arm of the benchmark.
One detail made bit-identical results possible: every stored vector was already L2 normalised, so the existing dot product was true cosine similarity. I sampled four hundred of them to confirm it rather than assume it.
Measured, on 25,631 products
| Measure | Before | After | Change |
|---|---|---|---|
| Catalog search, dev | 6,456 ms | 8.2 ms | 787× |
| Search on the live database | 20.1 to 23.4 s | 7 to 110 ms | 196× to 3,339× |
| Database rows read per message | 25,701 | 2 | 12,850× |
| Result parity against the old path | baseline | 10 of 10 identical | no drift |
Voice moved from the browser's speech API to server side speech recognition and synthesis in the same milestone, which took the product out of Chrome only and gave the assistant one consistent voice. A small cache on repeated phrases cut a repeated voice turn from 10.0 to 6.9 seconds.