@juspay/neurolink 12.0.3 → 12.0.4

package/dist/neurolink.js

@@ -5548,7 +5548,15 @@ Current user's request: ${currentInput}`;
                 : "empty",
             hasCustomSystemPrompt: !!options.systemPrompt,
         });
-        const conversationMessages = (await getConversationMessages(this.conversationMemory, options));
+        // Caller-supplied conversationMessages win — mirroring
+        // directProviderGeneration. getConversationMessages() returns [] whenever
+        // no memory manager / session context is configured, which silently
+        // dropped an inline history on the MCP-first path (the default path for
+        // any bare `new NeuroLink()` with tools enabled) while the direct path
+        // honored it.
+        const conversationMessages = (options.conversationMessages?.length
+            ? options.conversationMessages
+            : await getConversationMessages(this.conversationMemory, options));
         this.logMCPConversationSummary(requestId, conversationMessages);
         logger.debug("[Observability] Available tools for LLM", {
             requestId,

package/dist/rag/ragIntegration.js

@@ -235,9 +235,47 @@ async function _prepareRAGToolInner(ragConfig, fallbackProvider) {
     const EMBEDDING_DIMENSION = 128;
     const vectorStore = new InMemoryVectorStore();
     const indexName = "rag-index";
+    // When the caller configured an embedding provider/model, embed BOTH the
+    // index chunks and (below) the queries through that provider — previously
+    // those config fields had no runtime effect and retrieval always used the
+    // deterministic hash embedding, which is a lexical fingerprint rather than
+    // a semantic space. Index and query must share one embedding space, so the
+    // provider path replaces the hash path wholesale; any provider failure
+    // falls back to the hash for both sides.
+    const wantProviderEmbeddings = Boolean(embeddingProvider || embeddingModel);
+    const embedProviderName = embeddingProvider || fallbackProvider || "vertex";
+    const embedModelName = embeddingModel || "gemini-2.5-flash";
+    let embedFn = (text) => Promise.resolve(generateSimpleEmbedding(text, EMBEDDING_DIMENSION));
+    if (wantProviderEmbeddings) {
+        try {
+            const { AIProviderFactory } = await import("../core/factory.js");
+            const embedderProvider = (await AIProviderFactory.createProvider(embedProviderName, embedModelName));
+            if (typeof embedderProvider.embed === "function") {
+                const providerEmbed = embedderProvider.embed.bind(embedderProvider);
+                embedFn = (text) => providerEmbed(text, embedModelName);
+            }
+            else {
+                logger.warn(`[RAG] Embedding provider '${embedProviderName}' has no embed(); falling back to hash embeddings`);
+            }
+        }
+        catch (error) {
+            logger.warn("[RAG] Failed to create embedding provider; falling back to hash embeddings", { error: error instanceof Error ? error.message : String(error) });
+        }
+    }
+    let chunkVectors;
+    try {
+        chunkVectors = await Promise.all(allChunks.map((chunk) => embedFn(chunk.text)));
+    }
+    catch (error) {
+        // One failed chunk must not leave a mixed-space index — flip the whole
+        // index AND all queries back to the hash space together.
+        logger.warn("[RAG] Provider embedding failed mid-index; falling back to hash embeddings for index and queries", { error: error instanceof Error ? error.message : String(error) });
+        embedFn = (text) => Promise.resolve(generateSimpleEmbedding(text, EMBEDDING_DIMENSION));
+        chunkVectors = allChunks.map((chunk) => generateSimpleEmbedding(chunk.text, EMBEDDING_DIMENSION));
+    }
     const items = allChunks.map((chunk, i) => ({
         id: `rag-chunk-${i}`,
-        vector: generateSimpleEmbedding(chunk.text, EMBEDDING_DIMENSION),
+        vector: chunkVectors[i],
         metadata: {
             text: chunk.text,
             ...chunk.metadata,
@@ -274,9 +312,9 @@ async function _prepareRAGToolInner(ragConfig, fallbackProvider) {
                     "rag.top_k": topK ?? 5,
                 },
             }, async (span) => {
-                // For the in-memory store with simple embeddings,
-                // generate a query embedding using the same method
-                const queryEmbedding = generateSimpleEmbedding(query, EMBEDDING_DIMENSION);
+                // Query through the same embedding space the index was built in —
+                // provider embeddings when configured (and healthy), else the hash.
+                const queryEmbedding = await embedFn(query).catch(() => generateSimpleEmbedding(query, EMBEDDING_DIMENSION));
                 // Fetch more candidates than needed so diversity can select across files
                 const fetchK = fileContents.length > 1 ? topK * 3 : topK;
                 const rawResults = await vectorStore.query({

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@juspay/neurolink",
-  "version": "12.0.3",
+  "version": "12.0.4",
   "packageManager": "pnpm@10.15.1",
   "description": "TypeScript AI SDK with 24+ LLM providers behind one consistent API. MCP-native (connect any MCP server), voice TTS/STT/realtime, RAG, agents, memory, context compaction. OpenAI · Anthropic · Gemini · Bedrock · Azure · Ollama · DeepSeek · NVIDIA NIM and more.",
   "author": {