Can not use custom embedding model

[hanwendie]

I followed the instructions and used doubao embedding model with LangchainEmbeddingsWrapper, but an error occurred

(<class 'openai.BadRequestError'>, BadRequestError("Error code: 400 - {'error': {'code': 'InvalidParameter', 'message': 'The parameter input[0] specified in the request are not valid: expected a string, but got [4599 574 279 1176 2307 19763 30] instead. Request id: 021754361159016ee5eb03f9031c9451997bdb14cb7183a6316d9', 'param': 'input[0]', 'type': 'BadRequest'}}"), <traceback object at 0x134f2a300>)

here is the code

def do():
    evaluator_llm = ChatOpenAI(
        openai_api_base="https://ark.cn-beijing.volces.com/api/v3",
        openai_api_key="xxxx",  # app_key
        model_name="ep-20241017183117-br5f4",  # 推理接入点
    )
    evaluator = LangchainLLMWrapper(evaluator_llm)

    # embedding_model = HuggingFaceEmbeddings(model_name="bert-base-uncased")
    embedding_model_doubao = OpenAIEmbeddings(
        base_url="https://ark.cn-beijing.volces.com/api/v3",
        api_key="xxx",  # app_key
        model="ep-20250805102725-7m5n4",  # 推理接入点
    )
    embedding = LangchainEmbeddingsWrapper(embedding_model_doubao)

    sample = SingleTurnSample(
        user_input="When was the first super bowl?",
        response="The first superbowl was held on Jan 15, 1967",
        retrieved_contexts=[
            "The First AFL–NFL World Championship Game was an American football game played on January 15, 1967, at the Los Angeles Memorial Coliseum in Los Angeles."
        ]
    )
    scorer = ResponseRelevancy(llm=evaluator, embeddings=embedding)
    result = scorer.single_turn_score(sample)
    print(result)

[onestardao]

This error matches Problem Map No 2 (Interpretation Collapse) — the request payload type doesn’t align with what the embedding API expects. In your case, the wrapper is handing off a list where a string is required, so the model rejects it even though the config looks fine.

Quick check: log the raw input before it’s sent, and confirm it’s flattened into a single string per call. That usually resolves this mismatch.

If you’d like, I can share the full set of mapped failure modes + fixes (semantic firewall style, no infra changes). Just let me know.

[anistark]

The wrapper is NOT the problem - it correctly passes strings to the underlying LangChain OpenAIEmbeddings object.

The error message clearly shows:
expected a string, but got [4599 574 279 1176 2307 19763 30]

The Doubao API received a tokenized list (token IDs) instead of raw text strings.

Try using ragas's modern OpenAIEmbeddings provider instead of LangchainEmbeddingsWrapper.

[github-actions]

It seems the issue was answered, closing this now.