Quill keeps AI agents beside production SQL—not inside it

RavenDB’s Quill launch makes a specific architectural bet: production AI should sit beside an operational SQL database rather than run its retrieval workload directly against the system of record. The product supports PostgreSQL, SQL Server and MySQL, and uses change data capture (CDC) to maintain a continuously updated AI-facing layer while leaving the original database authoritative. (RavenDB)

The important boundary is the copy

Quill does not ask an enterprise to expose an entire database to an agent. Administrators select the tables, columns and rows that enter the layer; material that is not synchronized never reaches the agent. RavenDB says each agent can then receive a narrower scope for its assigned actions. The practical effect is two gates: an ingestion boundary around the copied data, followed by an agent-specific authorization boundary. (RavenDB)

That design also separates AI traffic from production query traffic. RavenDB says Quill reads database changes passively from CDC logs, avoiding table scans and vector indexing on the live system. Retrieval, embeddings and vector search run in the adjacent layer instead. If the source connection drops, reads can continue from the last synchronized state, while write-back actions queue until connectivity returns. (RavenDB)

This is not classic text-to-SQL. Quill translates selected relational records into a document model proposed by an automated schema analyzer and reviewed by an operator. Agents then use that AI-ready representation for retrieval and action workflows. That trade replaces direct SQL generation over the full live schema with a governed, eventually synchronized context surface. (RavenDB)

Price data churn before you deploy

Quill’s commercial model makes the architecture measurable. Reads and queries are free, but each ingested, updated or deleted document consumes a Write Request Unit; agent messages and embedding generation also count as writes. The Starter tier lists 1 million WRUs for $499 per month, while Pro lists 8 million for $1,799 per month. A 90-day free tier includes unlimited WRUs, and on-premises or air-gapped deployment is reserved for custom enterprise pricing. (RavenDB pricing)

That means the key capacity question is not only query volume. Teams must estimate the initial row-to-document import, ongoing CDC churn, embedding writes and agent-message volume. High-churn tables can be excluded, but doing so also narrows what agents know.

For production teams, Quill’s useful idea is the boundary—not the vendor’s aggressive deployment timeline. Copy the minimum approved context, keep AI indexing away from the transactional workload, and make every write-back path explicit. Whether Quill itself fits will depend on synchronization lag, authorization fidelity and the cost of maintaining that copy under real data churn.