--- title: "CORRA AI" description: "CORRA is the AI assistant built into Corridor. Ask it to write object definitions, explain job output, trace lineage, and diagnose errors." ---
TL;DR
CORRA is the AI assistant built into Corridor. It writes starter code for object definitions, answers questions about your job output and reports, traces what feeds an object and what depends on it, and explains why a run failed. It works inside your existing permissions and only sees objects you already have access to.
## What is it? CORRA is an AI assistant that runs inside the platform. Rather than being a general-purpose chatbot bolted onto the side, it understands Corridor's own concepts: registered objects, their versions, job output, lineage, and the approval lifecycle. That context is what makes it useful. When you ask "what feeds this feature?", CORRA reads the saved lineage for that exact object version rather than guessing from names. When you ask it to write a feature definition, it already knows the inputs available to that object and the return format the platform expects. CORRA is available in two forms: - **In the platform**, as an assistant you can open from the pages you already work on. - **Through [Corridor MCP](/technology/ai/)**, which lets an external agent connect to Corridor and answer the same kinds of questions from your own tooling. ## What can it help with? ### Write object definitions Ask CORRA to generate the Python or PySpark for a data element, feature, model, or global function. Because it knows the object's context, including the inputs available and what the definition needs to return, you do not have to paste that setup into the prompt yourself. This is most useful when moving existing analytical code onto the platform, where the logic is already settled and the work is reshaping it into the required format. ### Ask questions about job output After a job runs, ask questions about the results in plain language instead of exporting the data first. CORRA can query the output directly and answer questions like which segment carries the highest default rate, or how a score distribution shifted between two runs. It can also read the reports attached to a job, so questions about PSI, CSI, and performance metrics can be answered from the run itself. ### Trace dependencies Ask what feeds an object, or what would be affected if you changed it. CORRA traces the saved upstream inputs and downstream usage for a specific object version. This is the kind of question that is slow to answer by hand, because it means opening each object in turn and checking its inputs. Asking directly turns it into a single question. ### Diagnose errors When a run fails, CORRA can read the error trace and explain what went wrong. It covers both logic errors in a definition and infrastructure-level failures such as memory exhaustion or a lost worker, and it suggests what to change. ### Find and review objects Ask CORRA to find objects by description rather than exact name, including when you are not sure of the spelling. It can also summarize an object's saved configuration, which is a faster way to review something than opening every tab of its detail page. ## What it can and cannot see CORRA runs as you. Every question it answers is scoped to the objects your account can already read, and it cannot surface anything you would not be able to open yourself. Because CORRA works from saved platform metadata rather than from a general knowledge base, its answers reflect what is actually registered in your workspace. ## Getting good answers A few habits make a noticeable difference: - **Be specific about the object.** Naming the object and version removes ambiguity when several objects have similar names. - **Ask one question at a time.** Compound questions tend to get compound answers, and it is harder to see which part is wrong. - **Check generated code before saving it.** CORRA produces a starting point that reflects the object's context, not a finished, reviewed definition. It still goes through the same testing and approval path as anything you write yourself. - **Give it the error text.** When diagnosing a failure, the full trace is far more useful than a summary of it. :::caution Generated code and AI answers are subject to the same governance as any other change. An object still has to be tested and approved before it can be used in production, and CORRA does not shortcut that path. ::: ## What's Next - [AI Policy and Data Handling](/user-guide/corra-ai/ai-policy/): what CORRA does with your data, and what you stay responsible for - [Understand Job Output](/user-guide/how-tos/understand-job-output/): read the results CORRA can query for you - [Run a Simulation](/user-guide/how-tos/run-a-simulation/): execute an object's logic across a sample - [AI Architecture](/technology/ai/): how Corridor MCP works, and how to connect an external agent