[openvino-langchain] Fix embeddings (#977)

Details:
Fixed OpenVINOEmbeddings work with the embedding models.
Expanded list of tested embedding models
Updated env variable for embedding test

---------

Signed-off-by: Kirill Suvorov <[email protected]>
Co-authored-by: Alicja Miloszewska <[email protected]>

Author: Retribution98

modules/openvino-langchain/README.md

@@ -21,11 +21,17 @@ npm install openvino-langchain
 
 In order to use OpenVINO, you need to convert and compress the text generation model into the [OpenVINO IR format](https://docs.openvino.ai/2025/documentation/openvino-ir-format.html).
 
-Tested compatibility with:
-  - BAAI/bge-small-en-v1.5 (Embeddings model)
-  - openlm-research/open_llama_7b_v2
-  - meta-llama/Llama-2-13b-chat-hf
-  - microsoft/Phi-3.5-mini-instruct
+The following models are tested:
+  * Embeddings models:
+    - BAAI/bge-small-en-v1.5
+    - intfloat/multilingual-e5-large
+    - sentence-transformers/all-MiniLM-L12-v2
+    - sentence-transformers/all-mpnet-base-v2
+  * Large language models:  
+    - openlm-research/open_llama_7b_v2
+    - meta-llama/Llama-2-13b-chat-hf
+    - microsoft/Phi-3.5-mini-instruct
+    - deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
 
 #### Use HuggingFace Hub
 

modules/openvino-langchain/sample/README.md

@@ -19,5 +19,3 @@ cd sample/
 npm install
 node index.js *path_to_llm_model_dir* *path_to_embeddings_model_dir*
 ```
-
-The sample model `Phi-3.5-mini-instruct` should include both LLM and embeddings models.

modules/openvino-langchain/src/embeddings.ts

@@ -3,6 +3,9 @@ import { addon as ov } from 'openvino-node';
 // CVS-146344
 type CompiledModel = {
   createInferRequest: () => InferRequest;
+  output: (index: number) => any;
+  outputs: any[];
+  inputs: any[];
 };
 type InferRequest = {
   inferAsync: (inputs: any) => Promise<{ [outputName: string]: Tensor }>;
@@ -100,30 +103,25 @@ export class OpenVINOEmbeddings extends Embeddings {
 
     const inputTensor = new ov.Tensor([text]);
     const tokenizedInput = await irTokenizer.inferAsync([inputTensor]);
-    const inputShape = tokenizedInput['input_ids'].getShape();
-
-    // The current openvino-node version does not officially support Int64Array
-    // TODO: Remove any after adding official Int64Array support
-    const positionArray : any = BigInt64Array.from(
-      {length: inputShape[1]},
-      (_x, i) => BigInt(i),
-    );
-    const positionIds = new ov.Tensor(
-      ov.element.i64,
-      [1, inputShape[1]],
-      positionArray,
-    );
-    const beamIdx = new ov.Tensor(ov.element.i32, [1], new Int32Array([0]));
 
     const modelCompiled = await this.modelCompiled;
+
+    const tokenizerOutputs = tokenizerModelCompiled.outputs
+      .map(o => o.getAnyName())
+      .sort();
+    const embeddingInputs = modelCompiled.inputs
+      .map(i => i.getAnyName())
+      .sort();
+    if (!tokenizerOutputs.every(
+      (value, index) => value === embeddingInputs[index],
+    )) {
+      throw Error('Embedding model is incorrect. Tokenizer outputs should be the same as embedding model inputs.');
+    }
     const ir = modelCompiled.createInferRequest();
-    const embeddings = await ir.inferAsync({
-      'input_ids': tokenizedInput['input_ids'],
-      'attention_mask': tokenizedInput['attention_mask'],
-      'position_ids': positionIds,
-      'beam_idx': beamIdx,
-    });
-
-    return embeddings.logits.data;
+    const embeddings = await ir.inferAsync(tokenizedInput);
+
+    const outputName = modelCompiled.output(0);
+
+    return embeddings[outputName].data;
   }
 }

modules/openvino-langchain/src/tests/embeddings.int.test.ts

@@ -1,8 +1,8 @@
 import { test, expect } from '@jest/globals';
 import { OpenVINOEmbeddings } from '../embeddings.js';
 
-const modelPath = process.env.MODEL_PATH;
-if (modelPath === undefined) throw new Error('MODEL_PATH doest not defined');
+const modelPath = process.env.EMBEDDING_MODEL_PATH;
+if (modelPath === undefined) throw new Error('EMBEDDING_MODEL_PATH doest not defined');
 
 test('Test embedQuery', async () => {
   const embeddings = new OpenVINOEmbeddings({