Add KDB Router

llm_inference/model_inference.py

@@ -191,6 +191,11 @@ def _get_provider(self, model_name: str) -> str:
             "xiaomi/mimo-v2-flash:free": "openrouter",
             "openai/gpt-oss-120b": "openrouter",
             "qwen/qwen3-235b-a22b-2507": "openrouter",
+            "openai/gpt-6-luna": "openrouter",
+            "deepseek/deepseek-v4.1-flash": "openrouter",
+            "deepseek/deepseek-v4-flash-0731": "openrouter",
+            "google/gemma-4-31b-it": "openrouter",
+            "google/gemini-3-flash-preview": "openrouter",
             "qwen/qwen3-next-80b-a3b-instruct": "openrouter",
             "Qwen/Qwen3-Coder-Next": "openrouter",
             "deepseek/deepseek-v4-flash": "openrouter",

model_cost/model_cost.json

@@ -386,5 +386,17 @@
   "google/gemma-4-31b-it": {
     "input_token_price_per_million": 0.08,
     "output_token_price_per_million": 0.35
+  },
+  "openai/gpt-6-luna": {
+    "input_token_price_per_million": 0.1,
+    "output_token_price_per_million": 0.5
+  },
+  "deepseek/deepseek-v4.1-flash": {
+    "input_token_price_per_million": 0.03,
+    "output_token_price_per_million": 0.6
+  },
+  "deepseek/deepseek-v4-flash-0731": {
+    "input_token_price_per_million": 0.021,
+    "output_token_price_per_million": 0.32
   }
 }

router_inference/config/kdb-router.json

@@ -0,0 +1,14 @@
+{
+  "pipeline_params": {
+    "router_name": "kdb-router",
+    "router_cls_name": "KDBRouter",
+    "models": [
+      "deepseek/deepseek-v4-flash-0731",
+      "deepseek/deepseek-v4.1-flash",
+      "google/gemini-3-flash-preview",
+      "google/gemma-4-31b-it",
+      "openai/gpt-6-luna"
+    ],
+    "description": "KDB Router: content-only, cost-first routing (question body -> content category via MiniLM + logistic regression trained on an external calibration set -> model / reasoning setting). No RouterArena templates, config files, labels or proportions used."
+  }
+}

router_inference/router/__init__.py

@@ -11,6 +11,7 @@
 from router_inference.router.llm_router import LLMRouter
 from router_inference.router.lynkr_router import LynkrRouter
 from router_inference.router.cruq_sc_router import CruqSCRouter
+from router_inference.router.kdb_router import KDBRouter
 
 __all__ = [
     "BaseRouter",
@@ -21,4 +22,5 @@
     "ChuzomSoloV32Router",
     "LynkrRouter",
     "CruqSCRouter",
+    "KDBRouter",
 ]

router_inference/router/kdb_router.py

@@ -0,0 +1,58 @@
+# SPDX-FileCopyrightText: Copyright contributors to the RouterArena project
+# SPDX-License-Identifier: Apache-2.0
+
+"""KDB Router: content-only, cost-first routing.
+
+The question body (first/last paragraph dropped; line labels, option letters and "None" placeholders stripped) is
+embedded with all-MiniLM-L6-v2 and classified into one of 11 content categories by a logistic regression trained on
+an external calibration set (no RouterArena data). Each category maps to a model, optionally with a reasoning setting
+(`model@low` = reasoning effort low, `model@off` = reasoning off); the prediction is the base model name.
+Weights and table: config/kdb-router-model.json.
+"""
+
+import json
+import os
+import re
+
+import numpy as np
+
+from router_inference.router.base_router import BaseRouter
+
+_CFG = os.path.join(os.path.dirname(__file__), "..", "config", "kdb-router-model.json")
+_LABEL = re.compile(r"(?m)^[ \t]*[A-Za-z][A-Za-z0-9 _\-\"']{0,30}:[ \t]*")
+_OPTION = re.compile(r"(?m)^[ \t]*\(?[A-Ja-j][\).:\]][ \t]+")
+_PLACEHOLDER = re.compile(r"(?mi)^[ \t]*(none|null|n/a)[ \t]*$")
+
+
+def routing_text(prompt: str) -> str:
+    paras = [p for p in prompt.strip().split("\n\n") if p.strip()]
+    body = prompt.strip() if len(paras) < 3 else "\n\n".join(paras[1:-1]).strip()
+    text = _PLACEHOLDER.sub("", _OPTION.sub("", _LABEL.sub("", body)))
+    return re.sub(r"\n{2,}", "\n", text).strip()
+
+
+class KDBRouter(BaseRouter):
+    def __init__(self, router_name: str):
+        super().__init__(router_name)
+        from sentence_transformers import SentenceTransformer
+
+        with open(_CFG) as f:
+            cfg = json.load(f)
+        self._embed = SentenceTransformer(
+            cfg["embed_model"], revision=cfg["embed_revision"]
+        )
+        self._classes = cfg["classes"]
+        self._coef = np.array(cfg["coef"])
+        self._intercept = np.array(cfg["intercept"])
+        self._policy = cfg["policy"]
+        self._default = cfg["default"]
+
+    def route_option(self, query: str) -> str:
+        x = self._embed.encode(
+            [routing_text(query)], normalize_embeddings=True, show_progress_bar=False
+        )[0]
+        cat = self._classes[int(np.argmax(self._coef @ x + self._intercept))]
+        return self._policy.get(cat, self._default)
+
+    def _get_prediction(self, query: str) -> str:
+        return self.route_option(query).split("@")[0]

universal_model_names.py

@@ -138,6 +138,10 @@
     "deepseek-ai/DeepSeek-R1-0528-Qwen3-8B",
     # KT-ModelRouter pool additions
     "google/gemma-4-31b-it",
+    # KDB Router pool additions (OpenRouter / OpenAI-served)
+    "openai/gpt-6-luna",
+    "deepseek/deepseek-v4.1-flash",
+    "deepseek/deepseek-v4-flash-0731",
 ]