Add FERPAMetadataRetriever component for regulated higher-education RAG

[ashutoshrana]

Summary

This PR adds a FERPAMetadataRetriever component to haystack/components/retrievers/ that enforces FERPA (Family Educational Rights and Privacy Act) identity boundaries in retrieval-augmented generation pipelines for higher-education deployments.

Problem

Higher-education teams building RAG systems on Haystack need to ensure that student records do not cross identity or institutional boundaries during retrieval. The naive approach — retrieval followed by post-filtering — is insufficient for regulated environments: unauthorized documents enter the candidate pool and reach the scoring layer before being discarded, leaving a gap if the post-filter has a defect.

The correct approach is to apply the identity constraint at the document store query (pre-filter), so unauthorized documents never enter the retrieval pipeline regardless of their semantic similarity to the query. Haystack's FilterRetriever provides the mechanism; this component wraps it with FERPA-specific semantics, validation, and documentation.

What this adds

haystack/components/retrievers/ferpa_metadata_retriever.py

• FERPAMetadataRetriever — a @component-decorated retriever that accepts student_id and institution_id (at init or run time) and builds a compound AND filter over those metadata fields before querying the document store
• Configurable field names (student_id_field, institution_id_field) for deployments using non-standard naming
• run() and run_async() both implemented
• to_dict() / from_dict() serialization following existing patterns
• Pipeline-compatible: identity can be injected at run time from an upstream authentication component

test/components/retrievers/test_ferpa_metadata_retriever.py

16 tests covering:

• Initialization with and without identity values
• Custom metadata field names
• Serialization round-trip
• Filter structure (AND operator, meta. field prefix)
• Cross-student isolation (Alice's query cannot return Bob's records)
• Cross-institution isolation (univ-east query cannot return univ-west records)
• Runtime identity override (run-time values take precedence over init-time)
• Missing-identity ValueError for both fields
• Pipeline integration with init-time and run-time identity

Design notes

The component enforces identity at query time rather than post-retrieval because:

1. Documents never enter the candidate set — no semantic ranking of unauthorized content occurs
2. The behavior is deterministic — a cross-identity result cannot appear regardless of embedding similarity
3. The approach is consistent with how student information systems already implement field-level access control

The component addresses the retrieval-layer isolation requirement only. Callers remain responsible for session authentication, document category authorization (e.g., excluding counseling notes), and 34 CFR § 99.32 disclosure logging.

FERPA context

FERPA (20 U.S.C. § 1232g; 34 CFR Part 99) governs access to education records in US higher-education institutions. Its requirements are directly analogous to HIPAA's minimum-necessary standard for healthcare records and GLBA's safeguards rule for financial records — the same retrieval-layer pattern applies to all three.

Checklist

• Component follows @component decorator pattern used throughout haystack/components/retrievers/
• run() and run_async() both implemented with @component.output_types
• to_dict() / from_dict() serialization implemented
• Exported from haystack/components/retrievers/__init__.py via _import_structure
• 16 tests including cross-student/institution isolation and pipeline integration
• SPDX license headers present

[vercel]

@ashutoshrana is attempting to deploy a commit to the deepset Team on Vercel.

A member of the Team first needs to authorize it.

[CLAassistant]

All committers have signed the CLA.

[ashutoshrana]

Closing for now — will reopen after internal review.