From a6bcf2e916b84275487158074eaa9d3487a8f185 Mon Sep 17 00:00:00 2001
From: Ishaan Jaffer <ishaan@berri.ai>
Date: Wed, 22 Apr 2026 16:51:07 -0700
Subject: [PATCH] add eval_registry: DB ops for evals CRUD and agent attachment

---
 .../proxy/eval_management/eval_registry.py    | 133 ++++++++++++++++++
 1 file changed, 133 insertions(+)
 create mode 100644 litellm/proxy/eval_management/eval_registry.py

diff --git a/litellm/proxy/eval_management/eval_registry.py b/litellm/proxy/eval_management/eval_registry.py
new file mode 100644
index 00000000000..c41440e6220
--- /dev/null
+++ b/litellm/proxy/eval_management/eval_registry.py
@@ -0,0 +1,133 @@
+"""
+DB operations for the Evals feature.
+Mirrors litellm/proxy/guardrails/guardrail_registry.py
+"""
+
+import json
+from datetime import datetime, timezone
+from typing import Any, Dict, List, Optional
+
+
+async def add_eval_to_db(
+    eval_config: Dict[str, Any],
+    prisma_client: Any,
+) -> Dict[str, Any]:
+    criteria = eval_config.get("criteria", [])
+    created = await prisma_client.db.litellm_evalstable.create(
+        data={
+            "eval_name": eval_config["eval_name"],
+            "criteria": json.dumps(criteria),
+            "judge_model": eval_config["judge_model"],
+            "description": eval_config.get("description"),
+            "overall_threshold": eval_config.get("overall_threshold"),
+            "max_iterations": eval_config.get("max_iterations", 1),
+            "created_by": eval_config.get("created_by", ""),
+            "updated_by": eval_config.get("updated_by", ""),
+        }
+    )
+    row = dict(created)
+    if isinstance(row.get("criteria"), str):
+        row["criteria"] = json.loads(row["criteria"])
+    return row
+
+
+async def get_eval_by_id(eval_id: str, prisma_client: Any) -> Optional[Dict[str, Any]]:
+    row = await prisma_client.db.litellm_evalstable.find_unique(
+        where={"eval_id": eval_id}
+    )
+    if row is None:
+        return None
+    result = dict(row)
+    if isinstance(result.get("criteria"), str):
+        result["criteria"] = json.loads(result["criteria"])
+    return result
+
+
+async def list_evals(prisma_client: Any) -> List[Dict[str, Any]]:
+    rows = await prisma_client.db.litellm_evalstable.find_many(
+        order={"created_at": "desc"}
+    )
+    result = []
+    for row in rows:
+        r = dict(row)
+        if isinstance(r.get("criteria"), str):
+            r["criteria"] = json.loads(r["criteria"])
+        result.append(r)
+    return result
+
+
+async def update_eval_in_db(
+    eval_id: str,
+    update: Dict[str, Any],
+    prisma_client: Any,
+) -> Dict[str, Any]:
+    data: Dict[str, Any] = {"version": {"increment": 1}}
+    if "criteria" in update:
+        data["criteria"] = json.dumps(update["criteria"])
+    for field in ("judge_model", "description", "overall_threshold", "max_iterations"):
+        if field in update:
+            data[field] = update[field]
+    if "updated_by" in update:
+        data["updated_by"] = update["updated_by"]
+    updated = await prisma_client.db.litellm_evalstable.update(
+        where={"eval_id": eval_id}, data=data
+    )
+    row = dict(updated)
+    if isinstance(row.get("criteria"), str):
+        row["criteria"] = json.loads(row["criteria"])
+    return row
+
+
+async def delete_eval_from_db(eval_id: str, prisma_client: Any) -> Dict[str, Any]:
+    deleted = await prisma_client.db.litellm_evalstable.delete(
+        where={"eval_id": eval_id}
+    )
+    return dict(deleted)
+
+
+async def attach_eval_to_agent(
+    agent_id: str,
+    eval_id: str,
+    params: Dict[str, Any],
+    prisma_client: Any,
+) -> Dict[str, Any]:
+    created = await prisma_client.db.litellm_agentevalstable.create(
+        data={
+            "agent_id": agent_id,
+            "eval_id": eval_id,
+            "eval_name": params.get("eval_name", ""),
+            "on_failure": params.get("on_failure", "block"),
+            "overall_threshold_override": params.get("overall_threshold_override"),
+            "created_by": params.get("created_by", ""),
+        }
+    )
+    return dict(created)
+
+
+async def detach_eval_from_agent(
+    agent_id: str, eval_id: str, prisma_client: Any
+) -> Dict[str, Any]:
+    deleted = await prisma_client.db.litellm_agentevalstable.delete(
+        where={"agent_id_eval_id": {"agent_id": agent_id, "eval_id": eval_id}}
+    )
+    return dict(deleted)
+
+
+async def get_evals_for_agent(
+    agent_id: str, prisma_client: Any
+) -> List[Dict[str, Any]]:
+    rows = await prisma_client.db.litellm_agentevalstable.find_many(
+        where={"agent_id": agent_id},
+        include={"eval": True},
+    )
+    result = []
+    for row in rows:
+        r = dict(row)
+        eval_obj = r.get("eval")
+        if eval_obj is not None:
+            eval_dict = dict(eval_obj) if not isinstance(eval_obj, dict) else eval_obj
+            if isinstance(eval_dict.get("criteria"), str):
+                eval_dict["criteria"] = json.loads(eval_dict["criteria"])
+            r["eval"] = eval_dict
+        result.append(r)
+    return result