GDPR retrieval architecture with citation integrity for enterprise LLM deployment — Reference Architecture

Reference architecture · published 2026-10-16 · regulated enterprise · GDPR · layout diagram_first · implementation executed in CI before publish

The control that resolves this is a Query-Time ACL Retrieval with Citation Integrity, evaluated before the request reaches a model provider rather than audited afterwards. In regulated enterprise the binding constraint is GDPR, and the question it actually asks is what is the lawful basis, and can you actually erase. The shape below holds p95 ≤ 600ms over REST over HTTPS, and ships with a standard-library implementation that runs as written.

Reference architecture

Query-Time ACL Retrieval with Citation Integrity — the request path, with the decision points that must exist before the first model call:

flowchart LR
    Q[Query + live entitlements] --> SRCH[Search index]
    SRCH --> ACL{ACL match at query time?}
    ACL -->|no| DROP[Drop silently from results]
    ACL -->|yes| SUP{Superseded?}
    SUP -->|yes| DROP
    SUP -->|no| FRESH[Attach version + staleness]
    FRESH --> ANS[Answer + citations]
    ANS --> VER{Cited doc contains the claim?}
    VER -->|no| STRIP[Refuse the citation]
    VER -->|yes| OUT[Answer with verified citations]

The parts you have to build

1. Query-time ACL — Entitlements evaluated per query against live rights, never baked in.
2. Supersession links — A replaced document is unreachable, not merely lower-ranked.
3. Freshness SLO — Index age per document, with staleness reported on the hit.
4. Citation containment — A cited document must contain a claim-bearing span of the answer.
5. Provenance on the hit — Document id and version travel with the retrieved text.
6. Revocation path — An entitlement change takes effect without a reindex.

Implementation parameter matrix

Every value below is a parameter you can hold a vendor or an internal team to. Adjectives are not acceptance criteria.

The control gap

Permissions were applied when documents were indexed, the index is refreshed nightly, and citations are whatever the model names. So a revoked user still gets content, a superseded policy is quoted as current, and a citation points at a document that does not contain the claim.

What GDPR specifically requires

GDPR asks two things an audit log does not answer: under what lawful basis was this processed, and can you erase it on request. Erasure is the harder one — a prompt embedded into a vector index and a fine-tune has propagated beyond the row you deleted.

The artifact that satisfies it is a lineage graph from source record to every derived artifact. The way teams most often fail it: an erasure request that cannot be honoured because the data survives in embeddings, caches and eval sets.

Working implementation — Query-time ACL retrieval with citation containment

Standard-library Python, no external dependencies, runs as written. Executed by the publish gate: if it raises, this page does not ship. Wire it behind REST over HTTPS; the failure mode it guards against is a retried POST creating a duplicate record. The output you hand the assessor is the data-subject lineage and erasure report.

"""Retrieval with query-time ACLs, freshness SLO and citation integrity.

Three failures, one surface. Permissions applied at INDEX time go stale the
moment access changes, so a revoked user keeps seeing content through the
model. A stale index answers confidently from a superseded document. And a
citation that does not actually contain the claim is worse than no citation,
because it manufactures confidence.
"""
from dataclasses import dataclass, field


@dataclass
class Doc:
    doc_id: str
    text: str
    version: int
    indexed_at: int
    acl: set = field(default_factory=set)
    superseded_by: str | None = None


class Index:
    def __init__(self, freshness_sue_s: int = 3600):
        self.docs: dict[str, Doc] = {}
        self.freshness_sue_s = freshness_sue_s

    def put(self, doc: Doc) -> None:
        self.docs[doc.doc_id] = doc

    def stale(self, now: int) -> list[str]:
        return sorted(d.doc_id for d in self.docs.values()
                      if now - d.indexed_at > self.freshness_sue_s)

    def search(self, term: str, user_acl: set, now: int) -> list[dict]:
        """ACL is applied HERE, at query time, against live entitlements —
        not baked into the index where it goes stale on the first revocation."""
        hits = []
        for d in self.docs.values():
            if term.lower() not in d.text.lower():
                continue
            if d.superseded_by:
                continue
            if d.acl and not (d.acl & user_acl):
                continue
            hits.append({"doc_id": d.doc_id, "version": d.version,
                         "text": d.text,
                         "stale": now - d.indexed_at > self.freshness_sue_s})
        return sorted(hits, key=lambda h: h["doc_id"])


def verify_citations(answer: str, citations: list, index: Index) -> dict:
    """A citation must contain a claim-bearing span of the answer. This is a
    containment check, not a similarity score — a loose threshold here is how
    plausible-but-unsupported citations get through."""
    supported, unsupported = [], []
    sentences = [s.strip() for s in answer.split(".") if s.strip()]
    for cite in citations:
        doc = index.docs.get(cite)
        if not doc:
            unsupported.append({"cite": cite, "reason": "cited doc not in index"})
            continue
        if any(s.lower() in doc.text.lower() for s in sentences):
            supported.append(cite)
        else:
            unsupported.append({"cite": cite,
                                "reason": "no cited span appears in the document"})
    return {"supported": supported, "unsupported": unsupported,
            "trustworthy": not unsupported}


if __name__ == "__main__":
    idx = Index(freshness_sue_s=3600)
    idx.put(Doc("d1", "Refunds are processed within 14 days", 7,
                indexed_at=0, acl={"finance"}))
    idx.put(Doc("d2", "Refunds are processed within 30 days", 3,
                indexed_at=0, acl={"finance"}, superseded_by="d1"))
    idx.put(Doc("d3", "Public holiday schedule", 1, indexed_at=0))

    # Entitlement is checked at query time.
    allowed = idx.search("refunds", user_acl={"finance"}, now=10)
    assert [h["doc_id"] for h in allowed] == ["d1"], allowed
    denied = idx.search("refunds", user_acl={"marketing"}, now=10)
    assert denied == [], denied

    # Revocation takes effect immediately, with no reindex.
    idx.docs["d1"].acl = {"finance-restricted"}
    assert idx.search("refunds", user_acl={"finance"}, now=10) == []

    # Superseded documents never surface.
    assert all(h["doc_id"] != "d2" for h in
               idx.search("refunds", user_acl={"finance-restricted"}, now=10))

    # Staleness is reported rather than silently served as current.
    assert idx.stale(now=10) == []
    assert set(idx.stale(now=10_000)) == {"d1", "d2", "d3"}
    hits = idx.search("refunds", user_acl={"finance-restricted"}, now=10_000)
    assert hits and hits[0]["stale"] is True, hits

    good = verify_citations("Refunds are processed within 14 days.", ["d1"], idx)
    assert good["trustworthy"], good
    bad = verify_citations("Refunds are processed within 2 days.", ["d1"], idx)
    assert not bad["trustworthy"], bad
    missing = verify_citations("Anything.", ["d99"], idx)
    assert missing["unsupported"][0]["reason"] == "cited doc not in index"

    print("Query-time ACL retrieval + citation integrity: OK")
    print(f"  stale docs at t=10000: {len(idx.stale(10_000))}, "
          f"citation trustworthy good={good['trustworthy']} bad={bad['trustworthy']}")

The GDPR control, as code

Article 17 erasure propagated into derived artifacts. This is the artifact an assessor asks for under GDPR specifically — it is not the same control as the one above, and a system that implements one does not thereby implement the other. It executes under the same publish gate.

"""Article 17 erasure that propagates into derived artifacts.

The erasure request everyone implements is a DELETE on the source row. The one
that fails an audit is the derived copy: embeddings in a vector index, a cached
retrieval chunk, a fine-tuning corpus, a prompt log. Article 17(2) expects
controllers to take reasonable steps to inform processors of the erasure, and a
vector built from personal data is still personal data.

So erasure is modelled as a lineage walk with a receipt per derived artifact,
and the request is not complete until every downstream store confirms.
"""
from datetime import datetime, timezone


class Store:
    def __init__(self, name: str, kind: str):
        self.name, self.kind = name, kind
        self.rows: dict[str, dict] = {}
        self.erased: list[str] = []

    def put(self, key: str, subject_id: str, payload: str) -> None:
        self.rows[key] = {"subject_id": subject_id, "payload": payload}

    def erase_subject(self, subject_id: str) -> int:
        doomed = [k for k, v in self.rows.items() if v["subject_id"] == subject_id]
        for k in doomed:
            del self.rows[k]
            self.erased.append(k)
        return len(doomed)

    def holds(self, subject_id: str) -> int:
        return sum(1 for v in self.rows.values() if v["subject_id"] == subject_id)


class ErasureOrchestrator:
    """Lineage-aware erasure. A store that is not registered cannot be erased,
    so registration is the control — an unregistered derived store is exactly
    how personal data survives a 'completed' erasure."""

    def __init__(self):
        self.stores: list[Store] = []
        self.lineage: dict[str, list[str]] = {}

    def register(self, store: Store, derived_from: str | None = None) -> Store:
        self.stores.append(store)
        if derived_from:
            self.lineage.setdefault(derived_from, []).append(store.name)
        return store

    def execute(self, subject_id: str) -> dict:
        receipts = []
        for store in self.stores:
            held = store.holds(subject_id)
            removed = store.erase_subject(subject_id)
            receipts.append({"store": store.name, "kind": store.kind,
                             "held": held, "erased": removed})
        residual = {s.name: s.holds(subject_id) for s in self.stores
                    if s.holds(subject_id)}
        return {
            "subject_id": subject_id,
            "completed_at": datetime.now(timezone.utc).isoformat(timespec="seconds"),
            "receipts": receipts,
            "residual": residual,
            "complete": not residual,
            "derived_stores_covered": sorted(
                s.name for s in self.stores if s.kind != "source"),
        }


if __name__ == "__main__":
    orch = ErasureOrchestrator()
    src = orch.register(Store("crm_postgres", "source"))
    vec = orch.register(Store("vector_index", "embedding"), derived_from="crm_postgres")
    cache = orch.register(Store("retrieval_cache", "cache"), derived_from="vector_index")
    logs = orch.register(Store("prompt_log", "log"), derived_from="crm_postgres")

    src.put("c-1", "subject-77", "Jana Soor, contract renewal notes")
    vec.put("v-1", "subject-77", "[0.11, -0.4, ...]")
    cache.put("k-1", "subject-77", "chunk: contract renewal notes")
    logs.put("l-1", "subject-77", "prompt containing the same notes")
    src.put("c-2", "subject-90", "unrelated subject")

    result = orch.execute("subject-77")
    assert result["complete"], result["residual"]
    assert len(result["derived_stores_covered"]) == 3, result
    assert sum(r["erased"] for r in result["receipts"]) == 4, result
    assert src.holds("subject-90") == 1, "erasure must not over-reach"

    # An unregistered derived store is the failure this design exists to expose.
    orphan = Store("analytics_extract", "warehouse")
    orphan.put("a-1", "subject-77", "same personal data, different copy")
    assert orphan.holds("subject-77") == 1
    assert orphan.name not in {s.name for s in orch.stores}

    print("GDPR Article 17 lineage erasure: OK")
    print(f"  stores erased {len(result['receipts'])}, complete={result['complete']}")
    print(f"  UNREGISTERED store still holding data: {orphan.name} "
          f"({orphan.holds('subject-77')} row) — registration is the control")

Controls and acceptance parameters

Each row carries an acceptance parameter rather than an adjective, because "low latency" cannot be signed off. Retention here is bound by per the data class retention schedule, enforced automatically.

Questions to answer before approval

• A data subject requests erasure. Trace exactly what gets deleted, including embeddings and cached completions.
  Every derived artifact carries the source record id, so erasure traverses the lineage graph. The completion report lists each store touched and the rows removed.
• A user's access is revoked at 09:00. When does the model stop retrieving that content for them?
  On their next query. Entitlements are evaluated at query time against live rights, so revocation does not wait for a reindex — which is the whole reason index-time ACLs fail.
• What stops a superseded policy document being quoted as current?
  Supersession links. A replaced document is unreachable in search rather than ranked lower, so it cannot surface as a near-miss when the replacement is less similar.
• How do you know a citation actually supports the sentence it is attached to?
  A containment check: a claim-bearing span of the answer must appear in the cited document. A citation that fails is refused, because an unsupported citation manufactures confidence.
• How does a reader know whether the retrieved content is current?
  Every hit carries its document version and a staleness flag derived from the freshness SLO, so content older than the agreed window is visibly old rather than silently current.
• Can a restricted document influence results through similarity without being shown?
  No. Restricted documents are filtered during the scan, before ranking, so they never contribute to the candidate set a user may not read.

Adjacent problems

• Part of enterprise ai compliance architecture.
• This pattern also governs GDPR data governance framework for enterprise LLM deployment.
• GDPR prompt injection defence architecture for enterprise LLM deployment covers the adjacent case.
• This pattern also governs GDPR vendor evaluation criteria for enterprise AI implementation.
• GDPR provider failover and degradation architecture for enterprise LLM covers the adjacent case.
• If this is the blocker in front of you: enterprise AI infrastructure review.