The latest post in the Enterprise Document Intelligence series explains how a Typed Answer Contract can stop hallucination in Retrieval-Augmented Generation (RAG). By forcing the model to fill a rigid, typed schema, the pipeline limits the model’s creative freedom. The result is an answer that must come from the retrieved passages, not from memory exactly.

The article maps the four bricks of an enterprise RAG system: document parsing turns PDFs into structured tables; question parsing converts user strings into typed parsed questions; retrieval filters documents to passages containing the answer; and generation produces the final answer. Each brick pre‑shapes the input, shrinking the room for hallucination before the model runs.

At the heart of the contract is a typed schema built with Pydantic v2. Each answer type—Amount, Date Value, Table Value—has a dedicated class, and every field is paired with evidence spans that point to the source Ama. The model’s output is parsed by OpenAI’s Responses API client, enforcing the schema at decoding time and flagging any deviation before it reaches downstream code.

By requiring the model to emit only typed, evidence‑anchored fields, the contract eliminates the need for post‑generation filtering. Developers can extend the schema with custom indicators—such as jurisdiction flags or table confidence—without altering the core pipeline. The result is a deterministic, document‑grounded answer that can be directly consumed by downstream services.