Billing-Data Knowledge Agent
A retrieval agent over Oracle Utilities CCS billing systems that answers from curated documentation first and only reaches for generated SQL behind a flag.
- Role
- Sole author
- Period
- 2026
- Stack
- AWS Bedrock · RetrieveAndGenerate · Fargate · S3 · Oracle CCS · TypeScript
Problem
Answering a question about a utility account meant knowing which of several tenant systems held the answer, which schema it used, and which of its account types applied. That knowledge lived with a handful of people, and the questions arrived constantly.
A language model over the data is the obvious idea and the dangerous one. On billing records the failure mode is not a bad sentence, it is a confident wrong number, or a claim that something does not exist when the query simply missed it.
Constraints
The corpus describes real customers, including person identifiers such as SSN fields, so the schema had to be modelled explicitly rather than left for the model to infer. Tenants must never see each other's data, and an answer that cannot be grounded has to say so instead of guessing.
Decisions
- Generated SQL is the fallback, not the default.
- The prompt forbids claiming absence - the model may not say a thing does not exist merely because a lookup returned nothing.
- Tenant routing is explicit, scoped per tenant family, so a question can only reach the data it is entitled to.
Architecture
Curated first. Questions route through Bedrock's RetrieveAndGenerate over a curated corpus - an account-type catalog covering collective, secondary, usage, registration and service states - and answers cite the source they came from, with a badge marking whether that source is authoritative for the environment being asked about.
The SQL path sits behind an environment kill-switch and a separate grantable permission - flip the switch and the agent runs curated-only - and a source selector lets the asker choose documentation, live data, or automatic.
The guardrails are the product. The prompt is guarded against inventing terms, confusing environments, or mislabelling a count column.
The corpus does not go stale: a scheduled Fargate task refreshes it, with a least-privilege task role for the bucket, the knowledge base and the secrets it reads.
Outcome
- Answers arrive with citations and an authoritative-source badge, so a reader can tell grounded fact from generated text.
- The generated-SQL path sits behind an environment kill-switch and its own grantable permission, so any deployment can run curated-only with one flag.
- The agent refuses to claim something is absent - the failure mode that quietly misleads is the one it is built to avoid.