Checkpoint frontend UI/UX overhaul and ingestion embed benchmark work

apps/ai-service/rag/manifest.py

@@ -0,0 +1,62 @@
+"""F-05: refuse to become ready on a corpus/model manifest mismatch.
+
+Two different embedding models can produce vectors of the same
+dimensionality; Qdrant returns plausible-looking but meaningless nearest
+neighbours with no error at query time — a stale or wrong collection is
+otherwise invisible until a clinician notices the answers are subtly off.
+The ingestion loader already writes a sidecar manifest recording what a
+collection was built from (`ingestion/ingestion/load/manifest.py`); this is
+the query-time half — compare it against the configured query embedder
+*before* serving anything, not after a bad answer is reported.
+"""
+from __future__ import annotations
+
+MANIFEST_POINT_ID = "00000000-0000-5000-8000-000000000001"
+
+
+class ManifestMismatch(RuntimeError):
+    """The configured query embedder does not match what the collection was
+    built from. Raised at startup so the service refuses to become ready
+    rather than search with mismatched vectors."""
+
+
+def manifest_collection(name: str) -> str:
+    return f"{name}__manifest"
+
+
+def check_manifest(
+    payload: dict | None,
+    collection: str,
+    expected_model_id: str,
+    expected_dimensions: int,
+) -> None:
+    """Raises `ManifestMismatch` unless `payload` (the manifest sidecar
+    point's payload, or `None` if the sidecar/point is missing entirely)
+    matches the configured query embedder.
+
+    A collection with no manifest at all is refused for the same reason a
+    mismatched one is: nothing can be said about what it was built from, and
+    "probably fine" is not a load-bearing claim for a medical formulary.
+    """
+    if payload is None:
+        raise ManifestMismatch(
+            f"{collection!r} has no corpus manifest "
+            f"({manifest_collection(collection)!r}) — refusing to query an "
+            "unattested corpus."
+        )
+    mismatches = []
+    if payload.get("model_id") != expected_model_id:
+        mismatches.append(
+            f"model_id: corpus={payload.get('model_id')!r} "
+            f"query_embedder={expected_model_id!r}"
+        )
+    if payload.get("dimensions") != expected_dimensions:
+        mismatches.append(
+            f"dimensions: corpus={payload.get('dimensions')!r} "
+            f"query_embedder={expected_dimensions!r}"
+        )
+    if mismatches:
+        raise ManifestMismatch(
+            f"{collection!r}'s corpus manifest does not match the configured "
+            "query embedder — " + "; ".join(mismatches)
+        )