infiniflow/ragflow · error · Exception

Datasets use different embedding models: {[kb.embd_id for kb

Error message

Datasets use different embedding models: {[kb.embd_id for kb in kbs]}

What it means

Raised (as generic Exception) from get_models when all datasets in a dialog have embedding models but they are not the same model after normalization (tenant_model references and legacy model@instance@provider composites are resolved to base names before comparing). Cross-dataset retrieval requires one shared embedding space, so mismatched models abort model resolution.

Source

Thrown at api/db/services/dialog_service.py:359

                yield {"answer": "", "reference": {}, "audio_binary": None, "prompt": "", "created_at": time.time(), "final": False, **flags}
                continue
            yield {"answer": value, "reference": {}, "audio_binary": tts(tts_mdl, value), "prompt": "", "created_at": time.time(), "final": False}
    else:
        if model_config["model_type"] == "chat":
            answer = await chat_mdl.async_chat(system_prompt, msg, dialog.llm_setting)
        else:
            answer = await chat_mdl.async_chat(system_prompt, msg, dialog.llm_setting, images=image_files)
        user_content = msg[-1].get("content", "[content not available]")
        logging.debug("User: {}|Assistant: {}".format(user_content, answer))
        yield {"answer": answer, "reference": {}, "audio_binary": tts(tts_mdl, answer), "prompt": "", "created_at": time.time()}


def get_models(dialog, trace_context=None, langfuse_session_id=None):
    embd_mdl, chat_mdl, rerank_mdl, tts_mdl = None, None, None, None
    kbs = KnowledgebaseService.get_by_ids(dialog.kb_ids)
    err = validate_dataset_embedding_models(kbs)
    if err:
        raise Exception(err)

    if kbs and kbs[0].embd_id:
        embd_owner_tenant_id = kbs[0].tenant_id
        embd_model_config = resolve_model_config(embd_owner_tenant_id, LLMType.EMBEDDING, kbs[0].embd_id)
        embd_mdl = LLMBundle(embd_owner_tenant_id, embd_model_config, trace_context=trace_context, langfuse_session_id=langfuse_session_id)
        if not embd_mdl:
            raise LookupError("Embedding model(%s) not found" % kbs[0].embd_id)

    if dialog.llm_id:
        if dialog.tenant_llm_id:
            try:
                chat_model_config = get_model_config_by_id(dialog.tenant_id, LLMType.CHAT, dialog.tenant_llm_id)
            except LookupError:
                chat_model_config = resolve_model_config(dialog.tenant_id, LLMType.CHAT, dialog.llm_id)
        else:
            chat_model_config = resolve_model_config(dialog.tenant_id, LLMType.CHAT, dialog.llm_id)
    else:
        chat_model_config = get_tenant_default_model_by_type(dialog.tenant_id, LLMType.CHAT)

View on GitHub (pinned to 554fb1133a)

Solutions

1. Re-embed the outlier dataset(s) with the embedding model used by the others, then re-add them to the dialog.
2. Split the dialog so each dialog only groups datasets sharing one embedding model.
3. Check the message's listed embd_id values in the error to identify exactly which datasets disagree.

Example fix

# before
dialog.kb_ids = [kb_bge.id, kb_gte.id]  # different embd_id
# after: rebuild kb_gte with kb_bge's embedding model
kb_gte.embd_id = kb_bge.embd_id; kb_gte.save(); # then re-parse/re-embed its documents

embd_refs = {(kb.embd_id, getattr(kb, 'tenant_embd_id', None)) for kb in kbs}
# rely on the same validator the server uses:
err = validate_dataset_embedding_models(kbs)
if err:
    raise ValueError(err)  # shows exactly which embd_ids disagree

try:
    chat(dialog, msg)
except Exception as e:
    if 'different embedding models' in str(e):
        highlight_mismatched_datasets(dialog.kb_ids)  # parse embd list from message
    else:
        raise

When it happens

Trigger: Chat/ask request against a dialog whose kb_ids contain datasets created with different embd_id values (e.g. one BAAI/bge-large-zh, one maidapark/gte), including cases where composite refs point at different base models.

Common situations: Adding a dataset created later under a different default embedding model to an old dialog; tenant switched embedding providers and only some datasets were rebuilt; combining shared and private datasets with different models.

Related errors

• Cannot search across datasets where some have embedding mode
• dataset_id (or kb_id/knowledge_id) is required for retrieval
• Embedding model(%s) not found
• Miss parameter: {key}
• main() must be defined or exported.

AI-assisted analysis of infiniflow/ragflow@554fb1133a (2026-08-15). Data as JSON: /api/errors/ace1bb83387bf6a7. Report an issue: GitHub.