Skip to content

IVR Groundedness Validator Using Citations #775

@psschwei

Description

@psschwei

Build a groundedness validator that determines whether a RAG agent's response is grounded in the provided context documents. The validator uses a citation-based approach, leveraging the Citation Generation LoRA adapter to link spans of the response back to spans in the source passages.

The validator pipeline should have four stages:

  1. Citation Generation -- Use the citation generation intrinsic to produce citations mapping response spans to supporting passages in the source documents.
  2. Citation Necessity -- Classify each response span as needing or not needing a citation (e.g., conversational filler like "I'm happy to help" does not need one).
  3. Citation Support -- For spans that need citations, evaluate whether the generated citations fully support, partially support, or do not support the claim.
  4. Groundedness Output -- Produce a final boolean (grounded / not grounded) and a reason string explaining any unsupported spans. The reason should be usable by a downstream repair strategy.

The multi-stage pipeline should be encapsulated behind a simple interface so that integrating it into an existing agent requires minimal code, ideally one or two lines to add groundedness validation to a flow.

The validator should be tested end-to-end within an IVR (Introspective Validation and Repair) loop, not just in isolation.

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions