@ai-for-dev/providers 0.1.0 → 0.2.0

package/README.md

@@ -31,10 +31,13 @@ Vous pouvez aussi fournir la clé par variable d'environnement : `ILAAS_API_KEY`
 
 `extensions/index.ts` contient, pour chaque serveur, une table des modèles connus : taille du contexte, sortie maximale, lecture d'images et gestion du raisonnement. Ces valeurs ont été mesurées sur les serveurs. Le raisonnement peut être :
 
-- activable (`toggle`) : Pi propose `off` et `high`, et passe par `chat_template_kwargs.enable_thinking` ;
+- réglable par niveau : Pi propose `off`, `low`, `medium` et `high`, et envoie `enable_thinking` et `reasoning_effort` dans `chat_template_kwargs`. La table donne la valeur à envoyer pour chaque niveau, car chaque modèle a son vocabulaire (`xhigh` pour Qwen 3.8, `max` pour DeepSeek V4.1) ;
+- activable (`toggle`) : le modèle ne sait que réfléchir ou non. Pi propose `off` et `high`, et passe par `chat_template_kwargs.enable_thinking` ;
 - permanent (`always`) : Pi ne propose que `high` ;
 - absent (`none`) : Pi n'envoie aucun paramètre de raisonnement.
 
+Seuls `qwen-3.8-27b` et `ARES/DeepSeek-V4.1-Flash` lisent un niveau d'effort dans leur modèle de chat. Ceux de Qwen 3.6 et de Gemma 4 ne connaissent que `enable_thinking`.
+
 Un modèle marqué `hidden: true` n'apparaît pas, même si le serveur le liste. Un modèle qui n'est pas dans la table reste utilisable. Il prend la taille de contexte donnée par le serveur (128 000 tokens à défaut), 32 768 tokens de sortie, sans images ni raisonnement.
 
 ## Limites

package/extensions/index.ts

@@ -13,11 +13,19 @@ type ChatModel = Model<"openai-completions">;
 
 /**
  * How a model handles reasoning, measured against the real servers:
- * - "toggle": off by default, switched on with chat_template_kwargs.enable_thinking (vLLM).
+ * - "toggle": chat_template_kwargs.enable_thinking switches reasoning on and off, nothing finer.
  * - "always": always reasons and ignores every switch.
  * - "none": no reasoning; any reasoning parameter makes the server fail.
+ * - Efforts: the chat template also reads a reasoning_effort, given here as the value
+ *   to send for each of Pi's low, medium and high levels.
  */
-type Thinking = "toggle" | "always" | "none";
+type Thinking = "toggle" | "always" | "none" | Efforts;
+
+interface Efforts {
+	low: string;
+	medium: string;
+	high: string;
+}
 
 interface ModelTuning {
 	contextWindow?: number;
@@ -61,8 +69,10 @@ const ENDPOINTS: Endpoint[] = [
 		baseUrl: "https://ia4dev-owui.univ-grenoble-alpes.fr/api",
 		envVar: "IA4DEV_API_KEY",
 		models: {
-			"qwen-3.8-27b": { vision: true, thinking: "toggle" },
-			"ARES/DeepSeek-V4.1-Flash": { vision: true, thinking: "toggle" },
+			// The chat template accepts low, medium and xhigh (its default).
+			"qwen-3.8-27b": { vision: true, thinking: { low: "low", medium: "medium", high: "xhigh" } },
+			// The deployed chat template accepts low, high (its default) and max, but not medium.
+			"ARES/DeepSeek-V4.1-Flash": { vision: true, thinking: { low: "low", medium: "high", high: "max" } },
 			"gemma-4-31b": { vision: true, thinking: "toggle" },
 		},
 	},
@@ -72,11 +82,34 @@ const DEFAULT_CONTEXT_WINDOW = 128_000;
 const DEFAULT_MAX_TOKENS = 32_768;
 const FETCH_TIMEOUT_MS = 10_000;
 
-// Effort levels mean nothing to these servers: reasoning is either on or off.
-const THINKING_LEVELS: Record<Exclude<Thinking, "none">, ChatModel["thinkingLevelMap"]> = {
-	toggle: { minimal: null, low: null, medium: null },
-	always: { off: null, minimal: null, low: null, medium: null },
-};
+type Reasoning = Pick<ChatModel, "reasoning" | "thinkingLevelMap"> & { compat: ChatModel["compat"] };
+
+// Pi's levels without a counterpart on the server are mapped to null, which hides them.
+function reasoningFor(thinking: Thinking): Reasoning {
+	if (thinking === "none") return { reasoning: false, compat: {} };
+	if (thinking === "always") {
+		return { reasoning: true, thinkingLevelMap: { off: null, minimal: null, low: null, medium: null }, compat: {} };
+	}
+	if (thinking === "toggle") {
+		return {
+			reasoning: true,
+			thinkingLevelMap: { minimal: null, low: null, medium: null },
+			compat: { thinkingFormat: "qwen-chat-template" },
+		};
+	}
+	return {
+		reasoning: true,
+		thinkingLevelMap: { minimal: null, ...thinking },
+		compat: {
+			thinkingFormat: "chat-template",
+			chatTemplateKwargs: {
+				enable_thinking: { $var: "thinking.enabled" },
+				preserve_thinking: true,
+				reasoning_effort: { $var: "thinking.effort", omitWhenOff: true },
+			},
+		},
+	};
+}
 
 interface RemoteModel {
 	id: string;
@@ -90,7 +123,7 @@ interface RemoteModel {
 
 function toModel(endpoint: Endpoint, remote: RemoteModel): ChatModel {
 	const tuning = endpoint.models[remote.id] ?? {};
-	const thinking = tuning.thinking ?? "none";
+	const reasoning = reasoningFor(tuning.thinking ?? "none");
 	const contextWindow =
 		tuning.contextWindow ?? remote.max_model_len ?? remote.max_input_tokens ?? DEFAULT_CONTEXT_WINDOW;
 	const maxTokens = tuning.maxTokens ?? remote.max_output_tokens ?? DEFAULT_MAX_TOKENS;
@@ -101,8 +134,8 @@ function toModel(endpoint: Endpoint, remote: RemoteModel): ChatModel {
 		provider: endpoint.id,
 		baseUrl: endpoint.baseUrl,
 		input: tuning.vision ? ["text", "image"] : ["text"],
-		reasoning: thinking !== "none",
-		thinkingLevelMap: thinking === "none" ? undefined : THINKING_LEVELS[thinking],
+		reasoning: reasoning.reasoning,
+		thinkingLevelMap: reasoning.thinkingLevelMap,
 		cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
 		contextWindow,
 		maxTokens: Math.min(maxTokens, contextWindow),
@@ -110,9 +143,10 @@ function toModel(endpoint: Endpoint, remote: RemoteModel): ChatModel {
 			// These gateways reject the newer OpenAI-only fields.
 			supportsDeveloperRole: false,
 			supportsStore: false,
-			// reasoning_effort is ignored by vLLM chat templates and crashes the Mistral models.
+			// A top-level reasoning_effort is ignored by vLLM chat templates and crashes the
+			// Mistral models; models that read an effort get it through the chat template.
 			supportsReasoningEffort: false,
-			...(thinking === "toggle" ? { thinkingFormat: "qwen-chat-template" as const } : {}),
+			...reasoning.compat,
 		},
 	};
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ai-for-dev/providers",
-  "version": "0.1.0",
+  "version": "0.2.0",
   "description": "Pi providers for the AI4Dev course models (ILAAS and IA4Dev)",
   "keywords": [
     "pi",