@massa-ai/mcp-client 1.58.0 → 1.59.0

package/dist/config-cli.js

@@ -444,10 +444,33 @@ function parseLmStudioModelList(body) {
 function inferenceProviderList() {
   return LOCAL_INFERENCE_IDS.map((id) => INFERENCE_PROVIDERS[id]);
 }
-var LOCAL_INFERENCE_IDS, INFERENCE_PROVIDERS;
+function deriveInferenceBaseUrls(providerId, explicitBaseUrl) {
+  const spec = INFERENCE_PROVIDERS[providerId];
+  if (!explicitBaseUrl) {
+    return {
+      embeddingBaseUrl: spec.defaultEmbeddingBaseUrl,
+      llmBaseUrl: spec.defaultLlmBaseUrl
+    };
+  }
+  let end = explicitBaseUrl.length;
+  while (end > 0 && explicitBaseUrl.charCodeAt(end - 1) === SLASH)
+    end--;
+  const base = explicitBaseUrl.slice(0, end);
+  const suffix = spec.defaultLlmBaseUrl.startsWith(spec.defaultEmbeddingBaseUrl) ? spec.defaultLlmBaseUrl.slice(spec.defaultEmbeddingBaseUrl.length) : "";
+  return {
+    embeddingBaseUrl: base,
+    llmBaseUrl: `${base}${suffix}`
+  };
+}
+var LOCAL_INFERENCE_IDS, INFERENCE_ROLE_DEFAULTS, INFERENCE_PROVIDERS, SLASH = 47;
 var init_inference_providers = __esm(() => {
   init_embedding_dimensions();
   LOCAL_INFERENCE_IDS = ["ollama", "lmstudio"];
+  INFERENCE_ROLE_DEFAULTS = {
+    embedding: { contextWindow: 8192 },
+    instruct: { contextWindow: 16384, temperature: 0.2 },
+    coding: { contextWindow: 32768, temperature: 0 }
+  };
   INFERENCE_PROVIDERS = {
     ollama: {
       id: "ollama",
@@ -459,6 +482,13 @@ var init_inference_providers = __esm(() => {
         dimensions: "OLLAMA_EMBEDDING_DIMENSIONS"
       },
       knownDimensions: KNOWN_EMBEDDING_DIMENSIONS,
+      defaultModels: {
+        embedding: "qwen3-embedding:0.6b",
+        instruct: "qwen3-vl:8b",
+        coding: "qwen2.5-coder:7b"
+      },
+      appliesContextPerRequest: true,
+      embedBatchSize: 64,
       supportsOllamaVersionProbe: true,
       injectsDisableThink: true,
       requiresChatCompletionsApi: false,
@@ -474,8 +504,16 @@ var init_inference_providers = __esm(() => {
         dimensions: "LMSTUDIO_EMBEDDING_DIMENSIONS"
       },
       knownDimensions: {
-        "text-embedding-nomic-embed-text-v1.5": 768
+        "text-embedding-nomic-embed-text-v1.5": 768,
+        "text-embedding-qwen3-embedding-0.6b": 1024
       },
+      defaultModels: {
+        embedding: "text-embedding-qwen3-embedding-0.6b",
+        instruct: "qwen3-vl-8b-instruct",
+        coding: "qwen2.5-coder-7b-instruct"
+      },
+      appliesContextPerRequest: false,
+      embedBatchSize: 64,
       supportsOllamaVersionProbe: false,
       injectsDisableThink: false,
       requiresChatCompletionsApi: true,
@@ -490,6 +528,7 @@ var API_PROVIDER_IDS, EMBEDDING_PROVIDER_IDS, SCHEDULER_JOB_KINDS, MAX_MATCH_WOR
 var init_massa_ai_config = __esm(() => {
   init_xdg();
   init_inference_providers();
+  init_embedding_dimensions();
   API_PROVIDER_IDS = ["mistral", "openai", "google", "cohere"];
   EMBEDDING_PROVIDER_IDS = [
     ...LOCAL_INFERENCE_IDS,
@@ -556,9 +595,9 @@ var init_massa_ai_config = __esm(() => {
     },
     embedding: {
       provider: "ollama",
-      model: "qwen3-embedding:4b",
+      model: INFERENCE_PROVIDERS.ollama.defaultModels.embedding,
       baseURL: "http://localhost:11434",
-      dimensions: 2560
+      dimensions: knownEmbeddingDimensions(INFERENCE_PROVIDERS.ollama.defaultModels.embedding) ?? 768
     },
     compression: {
       defaultStrategy: "code_structure",
@@ -597,12 +636,15 @@ var init_massa_ai_config = __esm(() => {
       enabled: false,
       baseUrl: "http://localhost:11434/v1",
       apiKey: "ollama",
-      model: "qwen2.5:7b-instruct",
-      codeModel: "qwen2.5-coder:7b",
+      model: INFERENCE_PROVIDERS.ollama.defaultModels.instruct,
+      codeModel: INFERENCE_PROVIDERS.ollama.defaultModels.coding,
       temperature: 0.2,
       maxOutputTokens: 8000,
       timeoutMs: 90000,
-      disableThink: true
+      disableThink: true,
+      contextWindow: INFERENCE_ROLE_DEFAULTS.instruct.contextWindow,
+      codeContextWindow: INFERENCE_ROLE_DEFAULTS.coding.contextWindow,
+      codeTemperature: INFERENCE_ROLE_DEFAULTS.coding.temperature
     },
     memory: {
       decay: {
@@ -1157,6 +1199,13 @@ function getGlobalDataDir() {
     return fileConfig.dataDir;
   return path4.join(getConfigDir(), "data");
 }
+function activeInferenceProviderId() {
+  const providerId = fileConfig.embedding?.provider;
+  if (providerId && LOCAL_INFERENCE_IDS.includes(providerId)) {
+    return providerId;
+  }
+  return "ollama";
+}
 
 class Config {
   config;
@@ -1293,11 +1342,12 @@ class Config {
     this.config[key] = value;
   }
 }
-var DEFAULT_LLM_MODEL = "qwen2.5:7b-instruct", DEFAULT_LLM_CODE_MODEL = "qwen2.5-coder:7b", SCHEDULER_JOB_ENV, DEFAULT_ALLOWED_EXTENSIONS, fileConfig, fileCacheL1Bytes, fileCacheL2Bytes, resolvedDataDir, defaultConfig, config;
+var SCHEDULER_JOB_ENV, DEFAULT_ALLOWED_EXTENSIONS, fileConfig, DEFAULT_LLM_MODEL, DEFAULT_LLM_CODE_MODEL, fileCacheL1Bytes, fileCacheL2Bytes, resolvedDataDir, defaultConfig, config;
 var init_config = __esm(() => {
   init_env();
   init_config_loader();
   init_massa_ai_config();
+  init_inference_providers();
   init_massa_ai_config();
   init_massa_ai_config();
   init_config_loader();
@@ -1368,6 +1418,8 @@ var init_config = __esm(() => {
     ".hs"
   ];
   fileConfig = loadConfigSafe();
+  DEFAULT_LLM_MODEL = INFERENCE_PROVIDERS[activeInferenceProviderId()].defaultModels.instruct;
+  DEFAULT_LLM_CODE_MODEL = INFERENCE_PROVIDERS[activeInferenceProviderId()].defaultModels.coding;
   fileCacheL1Bytes = fileConfig.cache?.l1MaxSizeMB ? fileConfig.cache.l1MaxSizeMB * 1024 * 1024 : undefined;
   fileCacheL2Bytes = fileConfig.cache?.l2MaxSizeMB ? fileConfig.cache.l2MaxSizeMB * 1024 * 1024 : undefined;
   resolvedDataDir = getGlobalDataDir();
@@ -1410,7 +1462,10 @@ var init_config = __esm(() => {
       temperature: envNum("MASSA_AI_LLM_TEMPERATURE", fileConfig.llm?.temperature ?? 0.2),
       maxOutputTokens: envNum("MASSA_AI_LLM_MAX_OUTPUT_TOKENS", fileConfig.llm?.maxOutputTokens ?? 8000),
       timeoutMs: envNum("MASSA_AI_LLM_TIMEOUT_MS", fileConfig.llm?.timeoutMs ?? 90000),
-      disableThink: envBool("MASSA_AI_LLM_DISABLE_THINK", fileConfig.llm?.disableThink ?? true)
+      disableThink: envBool("MASSA_AI_LLM_DISABLE_THINK", fileConfig.llm?.disableThink ?? true),
+      contextWindow: fileConfig.llm?.contextWindow ?? INFERENCE_ROLE_DEFAULTS.instruct.contextWindow,
+      codeContextWindow: fileConfig.llm?.codeContextWindow ?? INFERENCE_ROLE_DEFAULTS.coding.contextWindow,
+      codeTemperature: envNum("MASSA_AI_LLM_CODE_TEMPERATURE", fileConfig.llm?.codeTemperature ?? INFERENCE_ROLE_DEFAULTS.coding.temperature)
     },
     memory: {
       decay: {
@@ -9346,7 +9401,7 @@ var require_ignore = __commonJS((exports, module) => {
   var REGEX_REPLACE_LEADING_EXCAPED_HASH = /^\\#/;
   var REGEX_SPLITALL_CRLF = /\r?\n/g;
   var REGEX_TEST_INVALID_PATH = /^\.*\/|^\.+$/;
-  var SLASH = "/";
+  var SLASH2 = "/";
   var TMP_KEY_IGNORE = "node-ignore";
   if (typeof Symbol !== "undefined") {
     TMP_KEY_IGNORE = Symbol.for("node-ignore");
@@ -9557,13 +9612,13 @@ var require_ignore = __commonJS((exports, module) => {
         return cache[path16];
       }
       if (!slices) {
-        slices = path16.split(SLASH);
+        slices = path16.split(SLASH2);
       }
       slices.pop();
       if (!slices.length) {
         return cache[path16] = this._testOne(path16, checkUnignored);
       }
-      const parent = this._t(slices.join(SLASH) + SLASH, cache, checkUnignored, slices);
+      const parent = this._t(slices.join(SLASH2) + SLASH2, cache, checkUnignored, slices);
       return cache[path16] = parent.ignored ? parent : this._testOne(path16, checkUnignored);
     }
     ignores(path16) {
@@ -52395,20 +52450,26 @@ function isLlmEnabled() {
 function _setLlmEnabledForTesting(flag) {
   testEnabledOverride = flag;
 }
-function getLlmConfig(opts) {
-  const cfg = config.get("llm");
-  const role = opts?.modelRole ?? "instruct";
-  const model = role === "code" ? cfg?.codeModel ?? cfg?.model ?? DEFAULT_LLM_MODEL : cfg?.model ?? DEFAULT_LLM_MODEL;
+function _resolveLlmConfig(cfg, role, baseUrlOverride) {
+  const baseUrl = baseUrlOverride ?? cfg?.baseUrl ?? INFERENCE_PROVIDERS.ollama.defaultLlmBaseUrl;
+  const spec = resolveInferenceSpec(baseUrl);
+  const model = role === "code" ? cfg?.codeModel ?? spec.defaultModels.coding : cfg?.model ?? spec.defaultModels.instruct;
+  const temperature = role === "code" ? cfg?.codeTemperature ?? INFERENCE_ROLE_DEFAULTS.coding.temperature : cfg?.temperature ?? INFERENCE_ROLE_DEFAULTS.instruct.temperature;
+  const contextWindow = role === "code" ? cfg?.codeContextWindow ?? INFERENCE_ROLE_DEFAULTS.coding.contextWindow : cfg?.contextWindow ?? INFERENCE_ROLE_DEFAULTS.instruct.contextWindow;
   return {
-    baseUrl: testBaseUrlOverride ?? cfg?.baseUrl ?? "http://localhost:11434/v1",
-    apiKey: cfg?.apiKey ?? "ollama",
+    baseUrl,
+    apiKey: cfg?.apiKey ?? spec.id,
     model,
-    temperature: cfg?.temperature ?? 0.2,
+    temperature,
+    contextWindow,
     maxOutputTokens: cfg?.maxOutputTokens ?? 8000,
     timeoutMs: cfg?.timeoutMs ?? 90000,
-    disableThink: cfg?.disableThink ?? true
+    disableThink: cfg?.disableThink ?? spec.injectsDisableThink
   };
 }
+function getLlmConfig(opts) {
+  return _resolveLlmConfig(config.get("llm"), opts?.modelRole ?? "instruct", testBaseUrlOverride);
+}
 function hostPort(url2) {
   try {
     const u = new URL(url2);
@@ -52452,12 +52513,34 @@ function _wrapFetchDisableThink(baseFetch) {
   };
   return wrapped;
 }
+function _wrapFetchContextWindow(baseFetch, contextWindow) {
+  const wrapped = async (input, init) => {
+    try {
+      if (init?.body && typeof init.body === "string") {
+        const parsed = JSON.parse(init.body);
+        if (parsed && typeof parsed === "object") {
+          parsed.options = { ...parsed.options, num_ctx: contextWindow };
+          init = { ...init, body: JSON.stringify(parsed) };
+        }
+      }
+    } catch {}
+    return baseFetch(input, init);
+  };
+  return wrapped;
+}
 function buildProvider(llm) {
   const spec = resolveInferenceSpec(llm.baseUrl);
+  let fetchImpl;
+  if (spec.appliesContextPerRequest) {
+    fetchImpl = _wrapFetchContextWindow(fetchImpl ?? globalThis.fetch, llm.contextWindow);
+  }
+  if (llm.disableThink && spec.injectsDisableThink) {
+    fetchImpl = _wrapFetchDisableThink(fetchImpl ?? globalThis.fetch);
+  }
   const openai2 = createOpenAI({
     baseURL: llm.baseUrl,
     apiKey: llm.apiKey,
-    ...llm.disableThink && spec.injectsDisableThink ? { fetch: _wrapFetchDisableThink(globalThis.fetch) } : {}
+    ...fetchImpl ? { fetch: fetchImpl } : {}
   });
   return spec.requiresChatCompletionsApi ? openai2.chat(llm.model) : openai2(llm.model);
 }
@@ -95205,6 +95288,9 @@ var init_local_transformers = __esm(() => {
 });
 
 // ../../packages/core/dist/services/embeddings/provider.js
+function _resolveEmbedContextWindow(embeddingConfig) {
+  return parsePositiveIntEnv(process.env.OLLAMA_EMBEDDING_NUM_CTX, embeddingConfig?.contextWindow ?? INFERENCE_ROLE_DEFAULTS.embedding.contextWindow);
+}
 function sleep(ms) {
   return new Promise((resolve4) => setTimeout(resolve4, ms));
 }
@@ -95247,7 +95333,7 @@ function createProvider(config3, providerId) {
   }
   return new AISDKEmbeddingProvider(config3, providerId);
 }
-var OLLAMA_EMBED_NUM_CTX, DimensionMismatchError, AISDKEmbeddingProvider;
+var DimensionMismatchError, AISDKEmbeddingProvider;
 var init_provider = __esm(() => {
   init_dist6();
   init_dist7();
@@ -95259,8 +95345,8 @@ var init_provider = __esm(() => {
   init_rate_limiter2();
   init_dist();
   init_config();
+  init_inference_providers();
   init_local_transformers();
-  OLLAMA_EMBED_NUM_CTX = parsePositiveIntEnv(process.env.OLLAMA_EMBEDDING_NUM_CTX, 8192);
   DimensionMismatchError = class DimensionMismatchError extends Error {
     providerId;
     expected;
@@ -95422,7 +95508,7 @@ var init_provider = __esm(() => {
               const response = await this.ollamaFetch("/api/embed", {
                 model: this.model,
                 input: inputText,
-                options: { num_ctx: OLLAMA_EMBED_NUM_CTX }
+                options: { num_ctx: _resolveEmbedContextWindow(loadConfigSafe().embedding) }
               });
               if (!response.ok) {
                 throw new Error(`Ollama API error: ${response.status} ${response.statusText}`);
@@ -95513,7 +95599,7 @@ var init_provider = __esm(() => {
             const response = await this.ollamaFetch("/api/embed", {
               model: this.model,
               input: texts.map((t) => this.sanitizeText(this.truncateText(t))),
-              options: { num_ctx: OLLAMA_EMBED_NUM_CTX }
+              options: { num_ctx: _resolveEmbedContextWindow(loadConfigSafe().embedding) }
             });
             if (!response.ok) {
               throw new Error(`Ollama batch API error: ${response.status} ${response.statusText}`);
@@ -108159,7 +108245,7 @@ var init_config2 = __esm(() => {
     })(),
     ollama: (() => {
       const file2 = fileFor("ollama");
-      const model = process.env.OLLAMA_EMBEDDING_MODEL || file2?.model || "qwen3-embedding:4b";
+      const model = process.env.OLLAMA_EMBEDDING_MODEL || file2?.model || INFERENCE_PROVIDERS.ollama.defaultModels.embedding;
       const rawEnvDimensions = Number(process.env.OLLAMA_EMBEDDING_DIMENSIONS);
       const envDimensions = Number.isInteger(rawEnvDimensions) && rawEnvDimensions > 0 ? rawEnvDimensions : undefined;
       const resolvedDimensions = resolveEmbeddingDimensions(model, file2?.dimensions, envDimensions);
@@ -108265,7 +108351,7 @@ var init_config2 = __esm(() => {
     })(),
     lmstudio: (() => {
       const file2 = fileFor("lmstudio");
-      const model = process.env.LMSTUDIO_EMBEDDING_MODEL || file2?.model || "text-embedding-nomic-embed-text-v1.5";
+      const model = process.env.LMSTUDIO_EMBEDDING_MODEL || file2?.model || INFERENCE_PROVIDERS.lmstudio.defaultModels.embedding;
       return {
         provider: "custom",
         model,
@@ -110617,6 +110703,12 @@ var init_base_vector_store = __esm(() => {
 });
 
 // ../../packages/core/dist/data/vector/postgres-vector-store.js
+function _resolveEmbedBatchSize(embeddingConfig) {
+  const providerId = embeddingConfig?.provider;
+  const spec = providerId && LOCAL_INFERENCE_IDS.includes(providerId) ? INFERENCE_PROVIDERS[providerId] : INFERENCE_PROVIDERS.ollama;
+  return embeddingConfig?.batchSize ?? spec.embedBatchSize;
+}
+
 class PostgresVectorCollection {
   pool;
   name;
@@ -110729,6 +110821,8 @@ var init_postgres_vector_store = __esm(() => {
   init_base_vector_store();
   init_dist();
   init_dist();
+  init_config();
+  init_inference_providers();
   init_identity_guard_installer();
   PostgresVectorStore = class PostgresVectorStore extends BaseVectorStore {
     pool = null;
@@ -110954,7 +111048,7 @@ var init_postgres_vector_store = __esm(() => {
       if (documents.length === 0)
         return;
       const pool = await this.ensureInitialized();
-      const EMBED_SUB_BATCH_SIZE = 8;
+      const EMBED_SUB_BATCH_SIZE = _resolveEmbedBatchSize(loadConfigSafe().embedding);
       let totalInserted = 0;
       let totalFailed = 0;
       for (let i = 0;i < documents.length; i += EMBED_SUB_BATCH_SIZE) {
@@ -154152,7 +154246,7 @@ Examples:
   massa-ai-config init
   massa-ai-config init --lmstudio
   massa-ai-config init --mistral your-api-key
-  massa-ai-config use ollama --model qwen3-embedding:4b
+  massa-ai-config use ollama --model qwen3-embedding:0.6b
   massa-ai-config use lmstudio
   massa-ai-config use mistral --api-key your-key
   massa-ai-config set embedding.dimensions 1024
@@ -154234,7 +154328,7 @@ async function runCli(argv) {
         console.log("\u2713 Configured for OpenAI embeddings");
       } else if (options.lmstudio) {
         const config3 = loadConfig();
-        const model = "text-embedding-nomic-embed-text-v1.5";
+        const model = INFERENCE_PROVIDERS.lmstudio.defaultModels.embedding;
         config3.embedding = {
           provider: "lmstudio",
           model,
@@ -154242,6 +154336,8 @@ async function runCli(argv) {
           dimensions: INFERENCE_PROVIDERS.lmstudio.knownDimensions[model] ?? 768
         };
         config3.llm.baseUrl = INFERENCE_PROVIDERS.lmstudio.defaultLlmBaseUrl;
+        config3.llm.model = INFERENCE_PROVIDERS.lmstudio.defaultModels.instruct;
+        config3.llm.codeModel = INFERENCE_PROVIDERS.lmstudio.defaultModels.coding;
         saveConfig(config3);
         console.log("\u2713 Configured for LM Studio (local) embeddings");
       } else {
@@ -154294,21 +154390,29 @@ Using defaults:`);
       }
       const config3 = loadConfig();
       if (provider === "ollama") {
+        const model = options.model || INFERENCE_PROVIDERS.ollama.defaultModels.embedding;
+        const urls = deriveInferenceBaseUrls("ollama", options["base-url"]);
         config3.embedding = {
           provider: "ollama",
-          model: options.model || "qwen3-embedding:4b",
-          baseURL: options["base-url"] || "http://localhost:11434",
-          dimensions: 2560
+          model,
+          baseURL: urls.embeddingBaseUrl,
+          dimensions: knownEmbeddingDimensions(model) ?? 768
         };
+        config3.llm.baseUrl = urls.llmBaseUrl;
+        config3.llm.model = INFERENCE_PROVIDERS.ollama.defaultModels.instruct;
+        config3.llm.codeModel = INFERENCE_PROVIDERS.ollama.defaultModels.coding;
       } else if (provider === "lmstudio") {
-        const model = options.model || "text-embedding-nomic-embed-text-v1.5";
+        const model = options.model || INFERENCE_PROVIDERS.lmstudio.defaultModels.embedding;
+        const urls = deriveInferenceBaseUrls("lmstudio", options["base-url"]);
         config3.embedding = {
           provider: "lmstudio",
           model,
-          baseURL: options["base-url"] || INFERENCE_PROVIDERS.lmstudio.defaultEmbeddingBaseUrl,
+          baseURL: urls.embeddingBaseUrl,
           dimensions: INFERENCE_PROVIDERS.lmstudio.knownDimensions[model] ?? 768
         };
-        config3.llm.baseUrl = options["base-url"] || INFERENCE_PROVIDERS.lmstudio.defaultLlmBaseUrl;
+        config3.llm.baseUrl = urls.llmBaseUrl;
+        config3.llm.model = INFERENCE_PROVIDERS.lmstudio.defaultModels.instruct;
+        config3.llm.codeModel = INFERENCE_PROVIDERS.lmstudio.defaultModels.coding;
       } else if (provider === "mistral") {
         if (!options["api-key"]) {
           console.error("Error: --api-key required for Mistral");

package/dist/index.js

@@ -25441,10 +25441,33 @@ function parseLmStudioModelList(body) {
 function inferenceProviderList() {
   return LOCAL_INFERENCE_IDS.map((id) => INFERENCE_PROVIDERS[id]);
 }
-var LOCAL_INFERENCE_IDS, INFERENCE_PROVIDERS;
+function deriveInferenceBaseUrls(providerId, explicitBaseUrl) {
+  const spec = INFERENCE_PROVIDERS[providerId];
+  if (!explicitBaseUrl) {
+    return {
+      embeddingBaseUrl: spec.defaultEmbeddingBaseUrl,
+      llmBaseUrl: spec.defaultLlmBaseUrl
+    };
+  }
+  let end = explicitBaseUrl.length;
+  while (end > 0 && explicitBaseUrl.charCodeAt(end - 1) === SLASH)
+    end--;
+  const base = explicitBaseUrl.slice(0, end);
+  const suffix = spec.defaultLlmBaseUrl.startsWith(spec.defaultEmbeddingBaseUrl) ? spec.defaultLlmBaseUrl.slice(spec.defaultEmbeddingBaseUrl.length) : "";
+  return {
+    embeddingBaseUrl: base,
+    llmBaseUrl: `${base}${suffix}`
+  };
+}
+var LOCAL_INFERENCE_IDS, INFERENCE_ROLE_DEFAULTS, INFERENCE_PROVIDERS, SLASH = 47;
 var init_inference_providers = __esm(() => {
   init_embedding_dimensions();
   LOCAL_INFERENCE_IDS = ["ollama", "lmstudio"];
+  INFERENCE_ROLE_DEFAULTS = {
+    embedding: { contextWindow: 8192 },
+    instruct: { contextWindow: 16384, temperature: 0.2 },
+    coding: { contextWindow: 32768, temperature: 0 }
+  };
   INFERENCE_PROVIDERS = {
     ollama: {
       id: "ollama",
@@ -25456,6 +25479,13 @@ var init_inference_providers = __esm(() => {
         dimensions: "OLLAMA_EMBEDDING_DIMENSIONS"
       },
       knownDimensions: KNOWN_EMBEDDING_DIMENSIONS,
+      defaultModels: {
+        embedding: "qwen3-embedding:0.6b",
+        instruct: "qwen3-vl:8b",
+        coding: "qwen2.5-coder:7b"
+      },
+      appliesContextPerRequest: true,
+      embedBatchSize: 64,
       supportsOllamaVersionProbe: true,
       injectsDisableThink: true,
       requiresChatCompletionsApi: false,
@@ -25471,8 +25501,16 @@ var init_inference_providers = __esm(() => {
         dimensions: "LMSTUDIO_EMBEDDING_DIMENSIONS"
       },
       knownDimensions: {
-        "text-embedding-nomic-embed-text-v1.5": 768
+        "text-embedding-nomic-embed-text-v1.5": 768,
+        "text-embedding-qwen3-embedding-0.6b": 1024
       },
+      defaultModels: {
+        embedding: "text-embedding-qwen3-embedding-0.6b",
+        instruct: "qwen3-vl-8b-instruct",
+        coding: "qwen2.5-coder-7b-instruct"
+      },
+      appliesContextPerRequest: false,
+      embedBatchSize: 64,
       supportsOllamaVersionProbe: false,
       injectsDisableThink: false,
       requiresChatCompletionsApi: true,
@@ -25487,6 +25525,7 @@ var API_PROVIDER_IDS, EMBEDDING_PROVIDER_IDS, SCHEDULER_JOB_KINDS, MAX_MATCH_WOR
 var init_massa_ai_config = __esm(() => {
   init_xdg();
   init_inference_providers();
+  init_embedding_dimensions();
   API_PROVIDER_IDS = ["mistral", "openai", "google", "cohere"];
   EMBEDDING_PROVIDER_IDS = [
     ...LOCAL_INFERENCE_IDS,
@@ -25553,9 +25592,9 @@ var init_massa_ai_config = __esm(() => {
     },
     embedding: {
       provider: "ollama",
-      model: "qwen3-embedding:4b",
+      model: INFERENCE_PROVIDERS.ollama.defaultModels.embedding,
       baseURL: "http://localhost:11434",
-      dimensions: 2560
+      dimensions: knownEmbeddingDimensions(INFERENCE_PROVIDERS.ollama.defaultModels.embedding) ?? 768
     },
     compression: {
       defaultStrategy: "code_structure",
@@ -25594,12 +25633,15 @@ var init_massa_ai_config = __esm(() => {
       enabled: false,
       baseUrl: "http://localhost:11434/v1",
       apiKey: "ollama",
-      model: "qwen2.5:7b-instruct",
-      codeModel: "qwen2.5-coder:7b",
+      model: INFERENCE_PROVIDERS.ollama.defaultModels.instruct,
+      codeModel: INFERENCE_PROVIDERS.ollama.defaultModels.coding,
       temperature: 0.2,
       maxOutputTokens: 8000,
       timeoutMs: 90000,
-      disableThink: true
+      disableThink: true,
+      contextWindow: INFERENCE_ROLE_DEFAULTS.instruct.contextWindow,
+      codeContextWindow: INFERENCE_ROLE_DEFAULTS.coding.contextWindow,
+      codeTemperature: INFERENCE_ROLE_DEFAULTS.coding.temperature
     },
     memory: {
       decay: {
@@ -26154,6 +26196,13 @@ function getGlobalDataDir() {
     return fileConfig.dataDir;
   return path4.join(getConfigDir(), "data");
 }
+function activeInferenceProviderId() {
+  const providerId = fileConfig.embedding?.provider;
+  if (providerId && LOCAL_INFERENCE_IDS.includes(providerId)) {
+    return providerId;
+  }
+  return "ollama";
+}
 
 class Config {
   config;
@@ -26290,11 +26339,12 @@ class Config {
     this.config[key] = value;
   }
 }
-var DEFAULT_LLM_MODEL = "qwen2.5:7b-instruct", DEFAULT_LLM_CODE_MODEL = "qwen2.5-coder:7b", SCHEDULER_JOB_ENV, DEFAULT_ALLOWED_EXTENSIONS, fileConfig, fileCacheL1Bytes, fileCacheL2Bytes, resolvedDataDir, defaultConfig, config2;
+var SCHEDULER_JOB_ENV, DEFAULT_ALLOWED_EXTENSIONS, fileConfig, DEFAULT_LLM_MODEL, DEFAULT_LLM_CODE_MODEL, fileCacheL1Bytes, fileCacheL2Bytes, resolvedDataDir, defaultConfig, config2;
 var init_config = __esm(() => {
   init_env();
   init_config_loader();
   init_massa_ai_config();
+  init_inference_providers();
   init_massa_ai_config();
   init_massa_ai_config();
   init_config_loader();
@@ -26365,6 +26415,8 @@ var init_config = __esm(() => {
     ".hs"
   ];
   fileConfig = loadConfigSafe();
+  DEFAULT_LLM_MODEL = INFERENCE_PROVIDERS[activeInferenceProviderId()].defaultModels.instruct;
+  DEFAULT_LLM_CODE_MODEL = INFERENCE_PROVIDERS[activeInferenceProviderId()].defaultModels.coding;
   fileCacheL1Bytes = fileConfig.cache?.l1MaxSizeMB ? fileConfig.cache.l1MaxSizeMB * 1024 * 1024 : undefined;
   fileCacheL2Bytes = fileConfig.cache?.l2MaxSizeMB ? fileConfig.cache.l2MaxSizeMB * 1024 * 1024 : undefined;
   resolvedDataDir = getGlobalDataDir();
@@ -26407,7 +26459,10 @@ var init_config = __esm(() => {
       temperature: envNum("MASSA_AI_LLM_TEMPERATURE", fileConfig.llm?.temperature ?? 0.2),
       maxOutputTokens: envNum("MASSA_AI_LLM_MAX_OUTPUT_TOKENS", fileConfig.llm?.maxOutputTokens ?? 8000),
       timeoutMs: envNum("MASSA_AI_LLM_TIMEOUT_MS", fileConfig.llm?.timeoutMs ?? 90000),
-      disableThink: envBool("MASSA_AI_LLM_DISABLE_THINK", fileConfig.llm?.disableThink ?? true)
+      disableThink: envBool("MASSA_AI_LLM_DISABLE_THINK", fileConfig.llm?.disableThink ?? true),
+      contextWindow: fileConfig.llm?.contextWindow ?? INFERENCE_ROLE_DEFAULTS.instruct.contextWindow,
+      codeContextWindow: fileConfig.llm?.codeContextWindow ?? INFERENCE_ROLE_DEFAULTS.coding.contextWindow,
+      codeTemperature: envNum("MASSA_AI_LLM_CODE_TEMPERATURE", fileConfig.llm?.codeTemperature ?? INFERENCE_ROLE_DEFAULTS.coding.temperature)
     },
     memory: {
       decay: {
@@ -34343,7 +34398,7 @@ var require_ignore = __commonJS((exports, module) => {
   var REGEX_REPLACE_LEADING_EXCAPED_HASH = /^\\#/;
   var REGEX_SPLITALL_CRLF = /\r?\n/g;
   var REGEX_TEST_INVALID_PATH = /^\.*\/|^\.+$/;
-  var SLASH = "/";
+  var SLASH2 = "/";
   var TMP_KEY_IGNORE = "node-ignore";
   if (typeof Symbol !== "undefined") {
     TMP_KEY_IGNORE = Symbol.for("node-ignore");
@@ -34554,13 +34609,13 @@ var require_ignore = __commonJS((exports, module) => {
         return cache[path16];
       }
       if (!slices) {
-        slices = path16.split(SLASH);
+        slices = path16.split(SLASH2);
       }
       slices.pop();
       if (!slices.length) {
         return cache[path16] = this._testOne(path16, checkUnignored);
       }
-      const parent = this._t(slices.join(SLASH) + SLASH, cache, checkUnignored, slices);
+      const parent = this._t(slices.join(SLASH2) + SLASH2, cache, checkUnignored, slices);
       return cache[path16] = parent.ignored ? parent : this._testOne(path16, checkUnignored);
     }
     ignores(path16) {
@@ -58954,20 +59009,26 @@ function isLlmEnabled() {
 function _setLlmEnabledForTesting(flag) {
   testEnabledOverride = flag;
 }
-function getLlmConfig(opts) {
-  const cfg = config2.get("llm");
-  const role = opts?.modelRole ?? "instruct";
-  const model = role === "code" ? cfg?.codeModel ?? cfg?.model ?? DEFAULT_LLM_MODEL : cfg?.model ?? DEFAULT_LLM_MODEL;
+function _resolveLlmConfig(cfg, role, baseUrlOverride) {
+  const baseUrl = baseUrlOverride ?? cfg?.baseUrl ?? INFERENCE_PROVIDERS.ollama.defaultLlmBaseUrl;
+  const spec = resolveInferenceSpec(baseUrl);
+  const model = role === "code" ? cfg?.codeModel ?? spec.defaultModels.coding : cfg?.model ?? spec.defaultModels.instruct;
+  const temperature = role === "code" ? cfg?.codeTemperature ?? INFERENCE_ROLE_DEFAULTS.coding.temperature : cfg?.temperature ?? INFERENCE_ROLE_DEFAULTS.instruct.temperature;
+  const contextWindow = role === "code" ? cfg?.codeContextWindow ?? INFERENCE_ROLE_DEFAULTS.coding.contextWindow : cfg?.contextWindow ?? INFERENCE_ROLE_DEFAULTS.instruct.contextWindow;
   return {
-    baseUrl: testBaseUrlOverride ?? cfg?.baseUrl ?? "http://localhost:11434/v1",
-    apiKey: cfg?.apiKey ?? "ollama",
+    baseUrl,
+    apiKey: cfg?.apiKey ?? spec.id,
     model,
-    temperature: cfg?.temperature ?? 0.2,
+    temperature,
+    contextWindow,
     maxOutputTokens: cfg?.maxOutputTokens ?? 8000,
     timeoutMs: cfg?.timeoutMs ?? 90000,
-    disableThink: cfg?.disableThink ?? true
+    disableThink: cfg?.disableThink ?? spec.injectsDisableThink
   };
 }
+function getLlmConfig(opts) {
+  return _resolveLlmConfig(config2.get("llm"), opts?.modelRole ?? "instruct", testBaseUrlOverride);
+}
 function hostPort(url2) {
   try {
     const u = new URL(url2);
@@ -59011,12 +59072,34 @@ function _wrapFetchDisableThink(baseFetch) {
   };
   return wrapped;
 }
+function _wrapFetchContextWindow(baseFetch, contextWindow) {
+  const wrapped = async (input, init) => {
+    try {
+      if (init?.body && typeof init.body === "string") {
+        const parsed = JSON.parse(init.body);
+        if (parsed && typeof parsed === "object") {
+          parsed.options = { ...parsed.options, num_ctx: contextWindow };
+          init = { ...init, body: JSON.stringify(parsed) };
+        }
+      }
+    } catch {}
+    return baseFetch(input, init);
+  };
+  return wrapped;
+}
 function buildProvider(llm) {
   const spec = resolveInferenceSpec(llm.baseUrl);
+  let fetchImpl;
+  if (spec.appliesContextPerRequest) {
+    fetchImpl = _wrapFetchContextWindow(fetchImpl ?? globalThis.fetch, llm.contextWindow);
+  }
+  if (llm.disableThink && spec.injectsDisableThink) {
+    fetchImpl = _wrapFetchDisableThink(fetchImpl ?? globalThis.fetch);
+  }
   const openai2 = createOpenAI({
     baseURL: llm.baseUrl,
     apiKey: llm.apiKey,
-    ...llm.disableThink && spec.injectsDisableThink ? { fetch: _wrapFetchDisableThink(globalThis.fetch) } : {}
+    ...fetchImpl ? { fetch: fetchImpl } : {}
   });
   return spec.requiresChatCompletionsApi ? openai2.chat(llm.model) : openai2(llm.model);
 }
@@ -101764,6 +101847,9 @@ var init_local_transformers = __esm(() => {
 });
 
 // ../../packages/core/dist/services/embeddings/provider.js
+function _resolveEmbedContextWindow(embeddingConfig) {
+  return parsePositiveIntEnv(process.env.OLLAMA_EMBEDDING_NUM_CTX, embeddingConfig?.contextWindow ?? INFERENCE_ROLE_DEFAULTS.embedding.contextWindow);
+}
 function sleep(ms) {
   return new Promise((resolve4) => setTimeout(resolve4, ms));
 }
@@ -101806,7 +101892,7 @@ function createProvider(config3, providerId) {
   }
   return new AISDKEmbeddingProvider(config3, providerId);
 }
-var OLLAMA_EMBED_NUM_CTX, DimensionMismatchError, AISDKEmbeddingProvider;
+var DimensionMismatchError, AISDKEmbeddingProvider;
 var init_provider = __esm(() => {
   init_dist6();
   init_dist7();
@@ -101818,8 +101904,8 @@ var init_provider = __esm(() => {
   init_rate_limiter2();
   init_dist();
   init_config();
+  init_inference_providers();
   init_local_transformers();
-  OLLAMA_EMBED_NUM_CTX = parsePositiveIntEnv(process.env.OLLAMA_EMBEDDING_NUM_CTX, 8192);
   DimensionMismatchError = class DimensionMismatchError extends Error {
     providerId;
     expected;
@@ -101981,7 +102067,7 @@ var init_provider = __esm(() => {
               const response = await this.ollamaFetch("/api/embed", {
                 model: this.model,
                 input: inputText,
-                options: { num_ctx: OLLAMA_EMBED_NUM_CTX }
+                options: { num_ctx: _resolveEmbedContextWindow(loadConfigSafe().embedding) }
               });
               if (!response.ok) {
                 throw new Error(`Ollama API error: ${response.status} ${response.statusText}`);
@@ -102072,7 +102158,7 @@ var init_provider = __esm(() => {
             const response = await this.ollamaFetch("/api/embed", {
               model: this.model,
               input: texts.map((t) => this.sanitizeText(this.truncateText(t))),
-              options: { num_ctx: OLLAMA_EMBED_NUM_CTX }
+              options: { num_ctx: _resolveEmbedContextWindow(loadConfigSafe().embedding) }
             });
             if (!response.ok) {
               throw new Error(`Ollama batch API error: ${response.status} ${response.statusText}`);
@@ -114718,7 +114804,7 @@ var init_config2 = __esm(() => {
     })(),
     ollama: (() => {
       const file2 = fileFor("ollama");
-      const model = process.env.OLLAMA_EMBEDDING_MODEL || file2?.model || "qwen3-embedding:4b";
+      const model = process.env.OLLAMA_EMBEDDING_MODEL || file2?.model || INFERENCE_PROVIDERS.ollama.defaultModels.embedding;
       const rawEnvDimensions = Number(process.env.OLLAMA_EMBEDDING_DIMENSIONS);
       const envDimensions = Number.isInteger(rawEnvDimensions) && rawEnvDimensions > 0 ? rawEnvDimensions : undefined;
       const resolvedDimensions = resolveEmbeddingDimensions(model, file2?.dimensions, envDimensions);
@@ -114824,7 +114910,7 @@ var init_config2 = __esm(() => {
     })(),
     lmstudio: (() => {
       const file2 = fileFor("lmstudio");
-      const model = process.env.LMSTUDIO_EMBEDDING_MODEL || file2?.model || "text-embedding-nomic-embed-text-v1.5";
+      const model = process.env.LMSTUDIO_EMBEDDING_MODEL || file2?.model || INFERENCE_PROVIDERS.lmstudio.defaultModels.embedding;
       return {
         provider: "custom",
         model,
@@ -117176,6 +117262,12 @@ var init_base_vector_store = __esm(() => {
 });
 
 // ../../packages/core/dist/data/vector/postgres-vector-store.js
+function _resolveEmbedBatchSize(embeddingConfig) {
+  const providerId = embeddingConfig?.provider;
+  const spec = providerId && LOCAL_INFERENCE_IDS.includes(providerId) ? INFERENCE_PROVIDERS[providerId] : INFERENCE_PROVIDERS.ollama;
+  return embeddingConfig?.batchSize ?? spec.embedBatchSize;
+}
+
 class PostgresVectorCollection {
   pool;
   name;
@@ -117288,6 +117380,8 @@ var init_postgres_vector_store = __esm(() => {
   init_base_vector_store();
   init_dist();
   init_dist();
+  init_config();
+  init_inference_providers();
   init_identity_guard_installer();
   PostgresVectorStore = class PostgresVectorStore extends BaseVectorStore {
     pool = null;
@@ -117513,7 +117607,7 @@ var init_postgres_vector_store = __esm(() => {
       if (documents.length === 0)
         return;
       const pool = await this.ensureInitialized();
-      const EMBED_SUB_BATCH_SIZE = 8;
+      const EMBED_SUB_BATCH_SIZE = _resolveEmbedBatchSize(loadConfigSafe().embedding);
       let totalInserted = 0;
       let totalFailed = 0;
       for (let i = 0;i < documents.length; i += EMBED_SUB_BATCH_SIZE) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@massa-ai/mcp-client",
-  "version": "1.58.0",
+  "version": "1.59.0",
   "description": "massa-ai MCP server - Semantic code search, memory, and context compression via Model Context Protocol",
   "author": "luizgmassa",
   "type": "module",
@@ -20,8 +20,8 @@
     "type-check": "tsc --noEmit"
   },
   "dependencies": {
-    "@massa-ai/core": "^1.58.0",
-    "@massa-ai/shared": "^1.58.0",
+    "@massa-ai/core": "^1.59.0",
+    "@massa-ai/shared": "^1.59.0",
     "@modelcontextprotocol/sdk": "^1.0.0"
   },
   "devDependencies": {