AI Integrator now lets each account choose the embedding model behind memory, chat history and file search, with models from Alibaba, Google, Amazon, OpenAI and Voyage AI.
Where to change it
On the web, open Settings > Memory and find Embedding model. Each model in the list shows its provider, how much text it reads at once, and a ZDR or No ZDR badge for whether a zero data retention route is available.
| Model | Provider |
|---|---|
| Qwen3 Embedding 4B (default) | Alibaba |
| text-embedding-005 | |
| Titan Text Embeddings V2 | Amazon |
| text-embedding-3-small | OpenAI |
| voyage-4-large, voyage-4, voyage-4-lite | Voyage AI |
| voyage-3.5, voyage-3.5-lite | Voyage AI |
What the embedding model is used for
- Memory. Finding the saved facts that match the message you just sent.
- Chat history. With Chat History Memory on, finished replies are split into short passages and indexed, so an old answer can inform a new one. Incognito chats are never indexed.
- Files. Searching the Files library by meaning when a chat looks for a document.
Search combines the embedding match with ordinary keyword matching, so exact names and numbers are still found.
Switching models
Choosing a different model asks "Switch embedding model?" and explains that every memory, history and file passage is re-embedded with the new one. Switch and re-embed starts the work, and the page shows "Re-embedding… this can take a minute." When it finishes, it reports how many memories, history passages and file passages were re-embedded. Search keeps working once the re-embed is done.
Every model's output is stored at the same size, 1,024 numbers per passage, so all of them fit the same index.
Privacy
Embedding requests go through the same model gateway as chat, and they carry your Zero Data Retention and No Training settings, including during a re-embed. With Zero Data Retention on, pick a model with a ZDR badge.
The vectors are stored with your account in AI Integrator's database, next to the text they were made from. Wipe in the Chat history section deletes the whole chat history index.
Which model to pick
The default, Qwen3 Embedding 4B, suits most accounts. Larger models such as voyage-4-large are designed for long and technical text. Before switching, check the model's badge against your privacy settings.
