Enable num_ctx to match defined chunk length in ollama embedder (#3129)

* Enable `num_ctx` to match defined chunk length in ollama embedder

* remove console

server/utils/EmbeddingEngines/ollama/index.js

@@ -1,4 +1,5 @@
 const { maximumChunkLength } = require("../../helpers");
+const { Ollama } = require("ollama");
 
 class OllamaEmbedder {
   constructor() {
@@ -7,21 +8,27 @@ class OllamaEmbedder {
     if (!process.env.EMBEDDING_MODEL_PREF)
       throw new Error("No embedding model was set.");
 
-    this.basePath = `${process.env.EMBEDDING_BASE_PATH}/api/embeddings`;
+    this.basePath = process.env.EMBEDDING_BASE_PATH;
     this.model = process.env.EMBEDDING_MODEL_PREF;
     // Limit of how many strings we can process in a single pass to stay with resource or network limits
     this.maxConcurrentChunks = 1;
     this.embeddingMaxChunkLength = maximumChunkLength();
+    this.client = new Ollama({ host: this.basePath });
+    this.log(
+      `initialized with model ${this.model} at ${this.basePath}. num_ctx: ${this.embeddingMaxChunkLength}`
+    );
   }
 
   log(text, ...args) {
     console.log(`\x1b[36m[${this.constructor.name}]\x1b[0m ${text}`, ...args);
   }
 
+  /**
+   * Checks if the Ollama service is alive by pinging the base path.
+   * @returns {Promise<boolean>} - A promise that resolves to true if the service is alive, false otherwise.
+   */
   async #isAlive() {
-    return await fetch(process.env.EMBEDDING_BASE_PATH, {
+    return await fetch(this.basePath)
-      method: "HEAD",
-    })
       .then((res) => res.ok)
       .catch((e) => {
         this.log(e.message);
@@ -40,6 +47,13 @@ class OllamaEmbedder {
    * This function takes an array of text chunks and embeds them using the Ollama API.
    * chunks are processed sequentially to avoid overwhelming the API with too many requests
    * or running out of resources on the endpoint running the ollama instance.
+   *
+   * We will use the num_ctx option to set the maximum context window to the max chunk length defined by the user in the settings
+   * so that the maximum context window is used and content is not truncated.
+   *
+   * We also assume the default keep alive option. This could cause issues with models being unloaded and reloaded
+   * on load memory machines, but that is simply a user-end issue we cannot control. If the LLM and embedder are
+   * constantly being loaded and unloaded, the user should use another LLM or Embedder to avoid this issue.
    * @param {string[]} textChunks - An array of text chunks to embed.
    * @returns {Promise<Array<number[]>>} - A promise that resolves to an array of embeddings.
    */
@@ -48,7 +62,6 @@ class OllamaEmbedder {
       throw new Error(
         `Ollama service could not be reached. Is Ollama running?`
       );
-
     this.log(
       `Embedding ${textChunks.length} chunks of text with ${this.model}.`
     );
@@ -58,15 +71,17 @@ class OllamaEmbedder {
 
     for (const chunk of textChunks) {
       try {
-        const res = await fetch(this.basePath, {
+        const res = await this.client.embeddings({
-          method: "POST",
+          model: this.model,
-          body: JSON.stringify({
+          prompt: chunk,
-            model: this.model,
+          options: {
-            prompt: chunk,
+            // Always set the num_ctx to the max chunk length defined by the user in the settings
-          }),
+            // so that the maximum context window is used and content is not truncated.
+            num_ctx: this.embeddingMaxChunkLength,
+          },
         });
 
-        const { embedding } = await res.json();
+        const { embedding } = res;
         if (!Array.isArray(embedding) || embedding.length === 0)
           throw new Error("Ollama returned an empty embedding for chunk!");