@devilsdev/rag-connectors 3.4.0 → 3.5.0

package/dist/index.cjs

@@ -363,6 +363,7 @@ var require_anthropic_connector = __commonJS({
     var loadAxios = () => require("axios");
     var { BaseConnector } = require_base_connector();
     var {
+      cleanAxiosErrors,
       missingApiKeyError,
       raceAbort,
       resolveApiKey,
@@ -374,6 +375,38 @@ var require_anthropic_connector = __commonJS({
     var { componentCacheKey } = require_component_keys();
     var API_KEY_ENV = "ANTHROPIC_API_KEY";
     var API_DEFAULT_TEMPERATURE = 1;
+    var NO_TEMPERATURE = /* @__PURE__ */ new Set();
+    var warnedTemperature = /* @__PURE__ */ new Set();
+    function isTemperatureRejection(error) {
+      const response = error && error.response;
+      if (!response || response.status !== 400) return false;
+      const data = response.data;
+      if (!data || typeof data.pipe === "function") return true;
+      const message = data.error && data.error.message || (typeof data === "string" ? data : "");
+      return /temperature/i.test(message);
+    }
+    function warnTemperatureDropped(model, temperature) {
+      if (warnedTemperature.has(model)) return;
+      warnedTemperature.add(model);
+      process.emitWarning(
+        `${model} does not accept temperature ${temperature}; sent without it (the API default). Claude 4.7 and later models accept only the default.`,
+        { code: "RAG_ANTHROPIC_TEMPERATURE" }
+      );
+    }
+    function responseText(data) {
+      const blocks = Array.isArray(data && data.content) ? data.content : [];
+      const text = blocks.filter(
+        (b) => b && typeof b.text === "string" && (b.type === "text" || b.type === void 0)
+      ).map((b) => b.text).join("");
+      if (text) return text;
+      const stop = data && data.stop_reason;
+      const error = new Error(
+        stop === "refusal" ? "Claude declined to answer (stop_reason: refusal)" : stop === "max_tokens" ? "Claude reached maxTokens before writing an answer (stop_reason: max_tokens); raise maxTokens" : `Claude returned no text (stop_reason: ${stop || "unknown"})`
+      );
+      error.code = stop === "refusal" ? "MODEL_REFUSED" : "NO_TEXT";
+      error.stopReason = stop || null;
+      throw error;
+    }
     var DEFAULT_CONFIG = {
       model: "claude-sonnet-5-5",
       baseURL: "https://api.anthropic.com",
@@ -431,6 +464,7 @@ var require_anthropic_connector = __commonJS({
             "content-type": "application/json"
           }
         });
+        cleanAxiosErrors(this.client, "Anthropic");
       }
       /**
        * @throws {Error} code MISSING_API_KEY when no key is configured
@@ -466,7 +500,7 @@ var require_anthropic_connector = __commonJS({
           messages: [{ role: "user", content: prompt.user }]
         };
         const temperature = options.temperature ?? this.config.temperature;
-        if (temperature !== void 0 && temperature !== API_DEFAULT_TEMPERATURE) {
+        if (temperature !== void 0 && temperature !== API_DEFAULT_TEMPERATURE && !NO_TEMPERATURE.has(body.model)) {
           body.temperature = temperature;
         }
         if (prompt.system) body.system = prompt.system;
@@ -479,26 +513,69 @@ var require_anthropic_connector = __commonJS({
        * @param {Array<Object|string>} [context=[]] - Retrieved documents for RAG
        * @param {Object} [options={}] - Override model, temperature, maxTokens per-call
        * @param {AbortSignal} [options.signal] - Cancels the request
-       * @returns {Promise<string>} The generated text
+       * @returns {Promise<string>} The answer: every text block of the reply,
+       *   joined. Rejects (code MODEL_REFUSED or NO_TEXT) when the reply has no
+       *   text, such as a refusal or a reply that ran out of maxTokens while
+       *   thinking.
        */
       async generate(query, context = [], options = {}) {
+        return (await this.generateWithUsage(query, context, options)).text;
+      }
+      /**
+       * Like generate(), with the reply's token usage. The LLM judge uses it so
+       * `check.judge.usage` reports what the judge cost.
+       *
+       * @param {string} query
+       * @param {Array<Object|string>} [context=[]]
+       * @param {Object} [options={}] - As for generate()
+       * @returns {Promise<{text: string, usage: {inputTokens: number, outputTokens: number, totalTokens: number}}>}
+       */
+      async generateWithUsage(query, context = [], options = {}) {
         const { signal } = options;
         throwIfAborted(signal);
         this._requireApiKey();
         const body = this._buildBody(query, context, options);
-        return this.withRetry(
-          async () => {
-            const response = await this.client.post(
-              "/v1/messages",
-              body,
-              signal ? { signal } : void 0
-            );
-            return response.data.content[0].text;
-          },
+        const data = await this._post(body, signal);
+        const text = responseText(data);
+        const u = data.usage || {};
+        const inputTokens = (u.input_tokens || 0) + (u.cache_creation_input_tokens || 0) + (u.cache_read_input_tokens || 0);
+        const outputTokens = u.output_tokens || 0;
+        return {
+          text,
+          usage: {
+            inputTokens,
+            outputTokens,
+            totalTokens: inputTokens + outputTokens
+          }
+        };
+      }
+      /**
+       * POST a Messages request with retries; a model that rejects
+       * `temperature` gets the request again without it.
+       * @private
+       */
+      async _post(body, signal) {
+        const send = (b) => this.withRetry(
+          async () => (await this.client.post(
+            "/v1/messages",
+            b,
+            signal ? { signal } : void 0
+          )).data,
           this.config.maxRetries,
           this.config.retryDelay,
           { signal }
         );
+        try {
+          return await send(body);
+        } catch (error) {
+          if (body.temperature === void 0 || !isTemperatureRejection(error)) {
+            throw error;
+          }
+          NO_TEMPERATURE.add(body.model);
+          warnTemperatureDropped(body.model, body.temperature);
+          const { temperature: _t, ...rest } = body;
+          return send(rest);
+        }
       }
       /**
        * Stream a response from Claude token-by-token via SSE.
@@ -513,14 +590,27 @@ var require_anthropic_connector = __commonJS({
         const { signal } = options;
         throwIfAborted(signal);
         this._requireApiKey();
-        const body = { ...this._buildBody(query, context, options), stream: true };
-        const response = await raceAbort(
-          this.client.post("/v1/messages", body, {
+        let body = { ...this._buildBody(query, context, options), stream: true };
+        const open = (b) => raceAbort(
+          this.client.post("/v1/messages", b, {
             responseType: "stream",
             ...signalOptions(signal)
           }),
           signal
         );
+        let response;
+        try {
+          response = await open(body);
+        } catch (error) {
+          if (body.temperature === void 0 || !isTemperatureRejection(error)) {
+            throw error;
+          }
+          NO_TEMPERATURE.add(body.model);
+          warnTemperatureDropped(body.model, body.temperature);
+          const { temperature: _t, ...rest } = body;
+          body = rest;
+          response = await open(body);
+        }
         let buffer = "";
         for await (const text of textChunks(response.data)) {
           throwIfAborted(signal);
@@ -560,7 +650,7 @@ var require_anthropic_connector = __commonJS({
         return buildRagPrompt(query, context).user;
       }
     };
-    module2.exports = { AnthropicConnector, DEFAULT_CONFIG };
+    module2.exports = { AnthropicConnector, DEFAULT_CONFIG, responseText };
   }
 });
 
@@ -572,6 +662,7 @@ var require_cohere_connector = __commonJS({
     var { componentCacheKey } = require_component_keys();
     var { BaseConnector } = require_base_connector();
     var {
+      cleanAxiosErrors,
       missingApiKeyError,
       resolveApiKey,
       throwIfAborted
@@ -627,6 +718,7 @@ var require_cohere_connector = __commonJS({
             "Content-Type": "application/json"
           }
         });
+        cleanAxiosErrors(this.client, "Cohere");
       }
       /**
        * @throws {Error} code MISSING_API_KEY when no key is configured
@@ -764,6 +856,7 @@ var require_ollama_connector = __commonJS({
     var loadAxios = () => require("axios");
     var { BaseConnector } = require_base_connector();
     var {
+      cleanAxiosErrors,
       raceAbort,
       signalOptions,
       textChunks,
@@ -822,6 +915,7 @@ var require_ollama_connector = __commonJS({
           baseURL: this.config.baseURL,
           timeout: this.config.timeout
         });
+        cleanAxiosErrors(this.client, "Ollama");
       }
       /**
        * Connect to the Ollama server by verifying it is reachable.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@devilsdev/rag-connectors",
-  "version": "3.4.0",
+  "version": "3.5.0",
   "description": "Embedders, LLMs and vector stores for @devilsdev/rag-core: OpenAI, Anthropic, Cohere, Ollama, the AI SDK, Qdrant and pgvector.",
   "keywords": [
     "rag",
@@ -43,7 +43,7 @@
     "access": "public"
   },
   "dependencies": {
-    "@devilsdev/rag-core": "3.4.0",
+    "@devilsdev/rag-core": "3.5.0",
     "axios": "^1.21.0"
   },
   "peerDependencies": {