PolyAI connected SurrealDB to Agent Studio for 30 millisecond RAG under the customer's control
PolyAI builds voice AI for customer service, the assistants that answer the phone for enterprises. Agent Studio is where those assistants are built, and SurrealDB is an external retrieval provider for it, running inside the customer's own infrastructure.
Challenge
PolyAI's enterprise customers wanted their assistants to answer from their own knowledge bases, inside Agent Studio, rather than migrating sensitive content into a vendor's ecosystem. For regulated organisations, where knowledge is stored and queried is a requirement, not a preference.
Spoken dialogue sets a hard latency budget. Response times have to stay well below the 1 to 1.5 second threshold at which a pause becomes perceptible, and every retrieval hop eats into it.
PolyAI needed to show that an external knowledge base could power retrieval and reasoning without compromising conversational quality, before customers would build on it.
Solution
PolyAI connected SurrealDB as an external RAG provider through Agent Studio's function-calling runtime, which calls SurrealDB for vector similarity search as the assistant reasons.
The integration required no architecture change, no specialised plugin and no data migration. SurrealDB slotted into the existing pipeline, running entirely within the organisation's own infrastructure so the customer keeps ownership and control of its knowledge.
SurrealDB holds documents, relational, graph and vector data in one system, so the same deployment is ready for graph retrieval and new data modalities as Agent Studio evolves.
Results
Near-zero added latency
Vector lookups from SurrealDB run at about 30 milliseconds, matching PolyAI's internal stack at about 31.
Full data ownership
Knowledge stays under enterprise control, in place, with nothing migrated into PolyAI's environment.
An unchanged architecture
The integration needed no plugin, migration or redesign.
A future-ready foundation
Graph and new data modalities are available on the engine already deployed.








