Urgent.News

What's breaking now, across hundreds of outlets.

AI

OpsPilot AI: Building RAG and Agents Without Giving the LLM Authority

An operator finds context in runbooks and documentation, interprets what is happening, and turns some of that information into tracked work. Finding an instruction and opening an issue seem closely related. The second task, however, crosses a boundary: it produces an effect in another system. That was my starting point for OpsPilot AI . I wanted to search tenant-owned operational knowledge,…

OpsPilot AI is designed to search tenant-owned operational knowledge, answer with references, and prepare a GitLab issue for another person to approve. The AI system treats the generated action as a proposal, separate from the user's permissions. The system separates authenticated principal, tenant knowledge, proposal, approval, and execution into distinct components to prevent confusion between generated information and granted authority.

PostgreSQL persists the lifecycle of these components, ensuring they are not dependent on the process remaining alive. The system defines strict boundaries for the model, restricting it to predefined actions and preventing arbitrary SQL or URL calls. Retrieval must respect the access boundary, meaning that only documents accessible to the authenticated principal can be considered for generating an answer.

The system separates domain contracts from the database, model, and GitLab adapters, allowing for independent testing of policy enforcement without incurring model call costs. Furthermore, the system configures transaction-local tenant context and forced row-level security, ensuring that even if the runtime role can set tenant context, RLS does not protect against arbitrary SQL execution as that role, nor does it provide document ACLs within a tenant.

The Retrieval-Augmented Generation (RAG) process is separated into distinct boundaries that can fail independently, including query, retriever, ranked evidence, generation, citation validation, and answer or abstention. This modular approach allows for easier measurement and improvement of each component without extending control into a universal guarantee about the semantic correctness of the final issue.

Rewritten by urgent.news; may contain errors — check the original.

Read the original at dev.to →

Report a problem with this story

More in AI

More from Wednesday 7 October →