@massa-ai/tools-api 1.58.0 → 1.59.0

package/dist/index.js

@@ -444,10 +444,15 @@ function parseLmStudioModelList(body) {
 function inferenceProviderList() {
   return LOCAL_INFERENCE_IDS.map((id) => INFERENCE_PROVIDERS[id]);
 }
-var LOCAL_INFERENCE_IDS, INFERENCE_PROVIDERS;
+var LOCAL_INFERENCE_IDS, INFERENCE_ROLE_DEFAULTS, INFERENCE_PROVIDERS;
 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 +464,13 @@ 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 +486,16 @@ 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 +510,7 @@ 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 +577,9 @@ 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 +618,15 @@ 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: {
@@ -1098,6 +1122,10 @@ function validatePartial(partial) {
       details.push("embedding.apiKey must be a string");
     if (e.dimensions !== undefined && !checkNumber(e.dimensions, 1))
       details.push("embedding.dimensions must be a positive number");
+    if (e.contextWindow !== undefined && !checkNumber(e.contextWindow, 1))
+      details.push("embedding.contextWindow must be a positive number");
+    if (e.batchSize !== undefined && !checkNumber(e.batchSize, 1))
+      details.push("embedding.batchSize must be a positive number");
   }
   if (partial.compression !== undefined) {
     const c = partial.compression;
@@ -1181,6 +1209,12 @@ function validatePartial(partial) {
       details.push("llm.timeoutMs must be a positive number");
     if (!checkBoolean(l.disableThink))
       details.push("llm.disableThink must be a boolean");
+    if (!checkNumber(l.contextWindow, 1))
+      details.push("llm.contextWindow must be a positive number");
+    if (!checkNumber(l.codeContextWindow, 1))
+      details.push("llm.codeContextWindow must be a positive number");
+    if (!checkNumber(l.codeTemperature))
+      details.push("llm.codeTemperature must be a number");
   }
   if (partial.memory !== undefined) {
     const m = partial.memory;
@@ -1619,6 +1653,13 @@ function getGlobalDataDir() {
     return fileConfig.dataDir;
   return path6.join(getConfigDir(), "data");
 }
+function activeInferenceProviderId() {
+  const providerId = fileConfig.embedding?.provider;
+  if (providerId && LOCAL_INFERENCE_IDS.includes(providerId)) {
+    return providerId;
+  }
+  return "ollama";
+}
 
 class Config {
   config;
@@ -1755,11 +1796,12 @@ 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();
@@ -1830,6 +1872,8 @@ 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();
@@ -1872,7 +1916,10 @@ 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: {
@@ -52615,20 +52662,26 @@ 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);
@@ -52672,12 +52725,34 @@ 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);
 }
@@ -95425,6 +95500,9 @@ 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));
 }
@@ -95467,7 +95545,7 @@ 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();
@@ -95479,8 +95557,8 @@ 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;
@@ -95642,7 +95720,7 @@ 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}`);
@@ -95733,7 +95811,7 @@ var init_provider = __esm(() => {
             const response = await this.ollamaFetch("/api/embed", {
               model: this.model,
               input: texts.map((t2) => this.sanitizeText(this.truncateText(t2))),
-              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}`);
@@ -107020,7 +107098,7 @@ var init_config2 = __esm(() => {
     })(),
     ollama: (() => {
       const file3 = fileFor("ollama");
-      const model = process.env.OLLAMA_EMBEDDING_MODEL || file3?.model || "qwen3-embedding:4b";
+      const model = process.env.OLLAMA_EMBEDDING_MODEL || file3?.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, file3?.dimensions, envDimensions);
@@ -107126,7 +107204,7 @@ var init_config2 = __esm(() => {
     })(),
     lmstudio: (() => {
       const file3 = fileFor("lmstudio");
-      const model = process.env.LMSTUDIO_EMBEDDING_MODEL || file3?.model || "text-embedding-nomic-embed-text-v1.5";
+      const model = process.env.LMSTUDIO_EMBEDDING_MODEL || file3?.model || INFERENCE_PROVIDERS.lmstudio.defaultModels.embedding;
       return {
         provider: "custom",
         model,
@@ -109478,6 +109556,12 @@ 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;
@@ -109590,6 +109674,8 @@ 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;
@@ -109815,7 +109901,7 @@ 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) {
@@ -181069,9 +181155,14 @@ var analyticsRoutes = new Elysia({ prefix: "/api/v1/analytics" }).post("/", asyn
 
 // src/routes/system.ts
 init_dist();
+init_config();
+init_inference_providers();
 import path37 from "path";
 import fs24 from "fs";
 import os9 from "os";
+function resolveConfiguredOllamaEmbeddingModel() {
+  return process.env.OLLAMA_EMBEDDING_MODEL || loadRawUserConfig().embedding?.model || INFERENCE_PROVIDERS.ollama.defaultModels.embedding;
+}
 function databaseUrlParts() {
   const url2 = new URL(process.env.DATABASE_URL);
   return {
@@ -181185,7 +181276,7 @@ var systemRoutes = new Elysia({ prefix: "/api/v1/system" }).get("/info", async (
   return {
     ...ollamaStatus,
     models,
-    configuredModel: process.env.OLLAMA_EMBEDDING_MODEL || "qwen3-embedding:4b",
+    configuredModel: resolveConfiguredOllamaEmbeddingModel(),
     baseUrl: process.env.OLLAMA_BASE_URL || "http://localhost:11434"
   };
 }, {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@massa-ai/tools-api",
-  "version": "1.58.0",
+  "version": "1.59.0",
   "author": "luizgmassa",
   "description": "massa-ai REST API server - Semantic code search, memory, and context compression",
   "type": "module",
@@ -21,8 +21,8 @@
     "test": "bun scripts/run-tests-isolated.ts"
   },
   "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",
     "elysia": "^1.2.25",
     "@elysiajs/swagger": "^1.2.0",
     "@elysiajs/cors": "^1.2.0",