From dec7c34ce9a9ab9ca677591769322386b7f520d9 Mon Sep 17 00:00:00 2001
From: kaminimangal <128952805+kaminimangal@users.noreply.github.com>
Date: Mon, 10 Aug 2026 19:17:47 +0530
Subject: [PATCH] Surface dropped_citations in /ask API response and CLI

---
 tests/test_api.py | 20 ++++++++++++++++++++
 vaultrag/cli.py   |  4 ++++
 vaultrag/main.py  |  2 ++
 3 files changed, 26 insertions(+)

diff --git a/tests/test_api.py b/tests/test_api.py
index 3f70e17..f25d074 100644
--- a/tests/test_api.py
+++ b/tests/test_api.py
@@ -119,6 +119,26 @@ async def test_ask_returns_only_permitted_documents(client):
     cited = {c["doc_id"] for c in body["citations"]}
     assert "ceo-private" not in cited, "LEAK: CEO notes surfaced over the API"
 
+async def test_dropped_citations_are_surfaced_in_ask_response(client):
+    """The model cites source [9] when only one source (index 1) was actually given.
+
+    The dropped citation should be visible in the API response, not silently discarded.
+    """
+    await _seed(client)
+    from vaultrag.generate import FakeLLM
+    client._transport.app.state.llm = FakeLLM(
+        '{"answer": "the bonus is 10%", "cited": [1, 9], "conflict": false}'
+    )
+    r = await client.post(
+        "/ask",
+        json={"question": "quarterly bonus payout policy"},
+        headers={"X-User-Id": "alice"},
+    )
+    assert r.status_code == 200
+    body = r.json()
+    assert body["answered"] is True
+    assert body["dropped_citations"] == [9]
+
 
 async def test_user_with_no_access_gets_a_refusal_not_a_leak(client):
     """bob is in sales. Neither seeded document is his, so he should be told nothing was found,
diff --git a/vaultrag/cli.py b/vaultrag/cli.py
index e02a6b1..863c00c 100644
--- a/vaultrag/cli.py
+++ b/vaultrag/cli.py
@@ -91,6 +91,10 @@ async def _ask(args) -> int:
             console.print(f"[yellow]stale[/] {s.doc_id} last updated {s.age_days}d ago (owner: {s.owner})")
 
         answer = generate(llm, args.question, hits)
+
+        if answer.dropped_citations:
+            console.print(f"[yellow]unverified citations dropped[/] {answer.dropped_citations}")
+
         console.print(f"\n[bold]{answer.text}[/]")
         if answer.citations:
             console.print("[dim]sources: " + ", ".join(c.doc_id for c in answer.citations) + "[/]")
diff --git a/vaultrag/main.py b/vaultrag/main.py
index 707df80..65b71ac 100644
--- a/vaultrag/main.py
+++ b/vaultrag/main.py
@@ -49,6 +49,7 @@ class AskResponse(BaseModel):
     citations: list[CitationOut] = []
     conflict: bool = False
     refusal_reason: str | None = None
+    dropped_citations: list[int] = []
     query_id: int
 
 
@@ -161,6 +162,7 @@ async def ask(req: AskRequest, user_id: str = Depends(current_user)) -> AskRespo
         answered=answer.answered,
         conflict=answer.conflict,
         refusal_reason=answer.refusal_reason,
+        dropped_citations=answer.dropped_citations,
         query_id=query_id,
         citations=[
             CitationOut(doc_id=c.doc_id, title=c.title, chunk_id=c.chunk_id, url=c.url, owner=c.owner)