AzureOpenAIStructuredLLM: constructor crashes on None/dict/BaseLlmConfig (AttributeError/TypeError) because azure_kwargs is accessed without AzureOpenAIConfig conversion #7428

Component

Python SDK

Summary

AzureOpenAIStructuredLLM.__init__ crashes on every config form except an already-built AzureOpenAIConfig: the class's own declared default (config: Optional[BaseLlmConfig] = None), any dict config, and any plain BaseLlmConfig all raise — AttributeError: 'BaseLlmConfig' object has no attribute 'azure_kwargs' (for dicts containing the documented azure_kwargs key: TypeError: BaseLlmConfig.__init__() got an unexpected keyword argument 'azure_kwargs').

The constructor passes the raw config to super().__init__() (mem0/llms/azure_openai_structured.py:18), where LLMBase builds a plain BaseLlmConfig from None/dict (mem0/llms/base.py:19-25), and then immediately reads self.config.azure_kwargs (mem0/llms/azure_openai_structured.py:24) — an attribute only AzureOpenAIConfig has. The sibling class AzureOpenAILLM converts None/dict/BaseLlmConfig to AzureOpenAIConfig before calling super() (mem0/llms/azure_openai.py:20-38), so the exact same inputs work there.

Steps to Reproduce

No Azure credentials or network needed — every crash happens before the AzureOpenAI client is built (the api_key="k" value also steers away from the DefaultAzureCredential branch).

from mem0.configs.llms.base import BaseLlmConfig
from mem0.llms.azure_openai import AzureOpenAILLM
from mem0.llms.azure_openai_structured import AzureOpenAIStructuredLLM

AZURE_DICT = {
    "model": "gpt-4o",
    "api_key": "k",
    "azure_kwargs": {
        "azure_deployment": "dep",
        "azure_endpoint": "https://x.openai.azure.com",
        "api_version": "2024-10-21",
    },
}

for label, arg in [
    ("no-arg (declared default None)", None),
    ("dict with azure_kwargs", AZURE_DICT),
    ("dict without azure_kwargs", {"model": "gpt-4o", "api_key": "k"}),
    ("plain BaseLlmConfig", BaseLlmConfig(model="gpt-4o", api_key="k")),
]:
    try:
        AzureOpenAIStructuredLLM(arg)
        print(f"AzureOpenAIStructuredLLM({label}): OK")
    except Exception as e:
        print(f"AzureOpenAIStructuredLLM({label}): CRASH: {type(e).__name__}: {e}")

# sibling contrast: same dict works for AzureOpenAILLM
try:
    AzureOpenAILLM(AZURE_DICT)
    print("AzureOpenAILLM(dict with azure_kwargs): OK")
except Exception as e:
    print(f"AzureOpenAILLM(dict with azure_kwargs): CRASH: {type(e).__name__}: {e}")

Expected Behavior

AzureOpenAIStructuredLLM should accept the config forms its own signature and its base class declare, converting them to AzureOpenAIConfig the way the sibling AzureOpenAILLM does:

• AzureOpenAIStructuredLLM() — the declared default config: Optional[BaseLlmConfig] = None (mem0/llms/azure_openai_structured.py:17) should not crash on its own default.
• dict configs — LLMBase.__init__ documents and implements dict configs: "Handle dict-based configuration (backward compatibility): self.config = BaseLlmConfig(**config)" with the docstring "config: LLM configuration option class or dict, defaults to None" (mem0/llms/base.py:16-23). AzureOpenAIConfig itself is a dict-friendly dataclass carrying azure_kwargs, and azure_kwargs is the documented key for Azure options.
• plain BaseLlmConfig — accepted by the declared signature's type annotation.

Concrete basis: the sibling class in the same provider family, AzureOpenAILLM.__init__ (mem0/llms/azure_openai.py:20-38):

if config is None:
    config = AzureOpenAIConfig()
elif isinstance(config, dict):
    config = AzureOpenAIConfig(**config)
elif isinstance(config, BaseLlmConfig) and not isinstance(config, AzureOpenAIConfig):
    config = AzureOpenAIConfig(model=..., temperature=..., ...)

With that conversion in place, the identical dict constructs AzureOpenAILLM successfully (verified below).

Actual Behavior

Verifier 1:

AzureOpenAIStructuredLLM(no-arg): CRASH AttributeError: 'BaseLlmConfig' object has no attribute 'azure_kwargs'  [azure_openai_structured.py:24 in __init__]
AzureOpenAIStructuredLLM(dict with azure_kwargs): CRASH TypeError: BaseLlmConfig.__init__() got an unexpected keyword argument 'azure_kwargs'
AzureOpenAIStructuredLLM(dict without azure_kwargs): CRASH AttributeError: 'BaseLlmConfig' object has no attribute 'azure_kwargs'
AzureOpenAIStructuredLLM(plain BaseLlmConfig): CRASH AttributeError: 'BaseLlmConfig' object has no attribute 'azure_kwargs'
AzureOpenAILLM(dict with azure_kwargs): OK

Verifier 2:

AzureOpenAIStructuredLLM(no-arg (declared default None)): CRASH: AttributeError: 'BaseLlmConfig' object has no attribute 'azure_kwargs'
AzureOpenAIStructuredLLM(dict with azure_kwargs): CRASH: TypeError: BaseLlmConfig.__init__() got an unexpected keyword argument 'azure_kwargs'
AzureOpenAIStructuredLLM(plain BaseLlmConfig): CRASH: AttributeError: 'BaseLlmConfig' object has no attribute 'azure_kwargs'
AzureOpenAILLM(dict with azure_kwargs): OK
AzureOpenAIStructuredLLM(full AzureOpenAIConfig): OK, model = gpt-4o

Re-run on c7ee362aff94a369af70f13f2b4f853f6793ff4c before filing — same result:

AzureOpenAIStructuredLLM(no-arg (declared default None)): CRASH: AttributeError: 'BaseLlmConfig' object has no attribute 'azure_kwargs'
AzureOpenAIStructuredLLM(dict with azure_kwargs): CRASH: TypeError: BaseLlmConfig.__init__() got an unexpected keyword argument 'azure_kwargs'
AzureOpenAIStructuredLLM(dict without azure_kwargs): CRASH: AttributeError: 'BaseLlmConfig' object has no attribute 'azure_kwargs'
AzureOpenAIStructuredLLM(plain BaseLlmConfig): CRASH: AttributeError: 'BaseLlmConfig' object has no attribute 'azure_kwargs'
AzureOpenAILLM(dict with azure_kwargs): OK

Tracebacks (same re-run) — the two distinct failure points:

File "mem0/llms/azure_openai_structured.py", line 24, in __init__
    api_key = self.config.azure_kwargs.api_key or os.getenv("LLM_AZURE_OPENAI_API_KEY")
              ^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: 'BaseLlmConfig' object has no attribute 'azure_kwargs'

File "mem0/llms/base.py", line 23, in __init__
    self.config = BaseLlmConfig(**config)
                  ^^^^^^^^^^^^^^^^^^^^^^^
TypeError: BaseLlmConfig.__init__() got an unexpected keyword argument 'azure_kwargs'

Only passing an already-built AzureOpenAIConfig works.

Environment

• mem0 version: 2.0.20 (pyproject.toml), commit c7ee362aff94a369af70f13f2b4f853f6793ff4c
• Python 3.12.14
• macOS, Darwin 25.6.0
• No credentials / offline — the crash precedes any client construction or network call

Root Cause

mem0/llms/azure_openai_structured.py:17-24 — __init__ forwards the raw config to super().__init__(config) without converting it to AzureOpenAIConfig. LLMBase.__init__ then turns None/dict into a plain BaseLlmConfig (mem0/llms/base.py:19-25), which has no azure_kwargs attribute, so the first access at line 24 raises AttributeError; a dict that does contain azure_kwargs instead fails earlier inside super() with TypeError, because BaseLlmConfig(**config) does not accept that keyword. The sibling AzureOpenAILLM (mem0/llms/azure_openai.py:20-38) does the conversion first and is unaffected.

How You Verified This

What I Ran

The script in Steps to Reproduce against a checkout of c7ee362aff94a369af70f13f2b4f853f6793ff4c (editable install, Python 3.12.14, azure-identity installed for the module imports), offline. Independently reproduced by two verifiers on the same commit with the same script.

What I Saw

The verbatim outputs in Actual Behavior: all four documented/declared config forms crash AzureOpenAIStructuredLLM.__init__ (AttributeError at azure_openai_structured.py:24, or TypeError at base.py:23 for a dict carrying azure_kwargs), while the identical dict constructs AzureOpenAILLM fine and a full AzureOpenAIConfig constructs AzureOpenAIStructuredLLM fine.

Why This Is a Bug

The constructor violates its own declared contract: the signature is config: Optional[BaseLlmConfig] = None (azure_openai_structured.py:17), so AzureOpenAIStructuredLLM() must work, and LLMBase explicitly documents and supports dict configs "backward compatibility" (base.py:16-23). The direct fix already exists in the same provider family — AzureOpenAILLM.__init__ (azure_openai.py:20-38) converts None → AzureOpenAIConfig(), dict → AzureOpenAIConfig(**config), and plain BaseLlmConfig → AzureOpenAIConfig(...); AzureOpenAIStructuredLLM is the only Azure class missing that conversion. The class is a registered provider ("azure_openai_structured": (..., AzureOpenAIConfig) at mem0/utils/factory.py:52), so anyone instantiating it directly — the documented pattern for the LLM classes — hits the crash.

What I Ruled Out

• The Memory/factory path — LlmFactory.create pre-builds the provider config class before instantiating (mem0/utils/factory.py:80-88), so from_config-style usage works; only direct construction with the class's own declared signature crashes.
• Missing environment variables — the os.getenv fallbacks at azure_openai_structured.py:24-27 are only reached after the self.config.azure_kwargs access; no env setting can avoid the AttributeError.
• Test coverage — tests/llms/test_azure_openai_structured.py only passes a DummyConfig duck-type that happens to carry an azure_kwargs attribute (tests/llms/test_azure_openai_structured.py:26-49), so None/dict/BaseLlmConfig inputs are never exercised.
• PR fix(llms): honor reasoning-model params in AzureOpenAIStructuredLLM #5548 (merged, touches this file) — fixes reasoning-model params in generate_response; different function, different defect.

Related

Searched issues/PRs for azure_kwargs, AzureOpenAIStructuredLLM, azure_openai_structured: nothing covers this constructor crash. Nearest, none duplicates: #4164 (closed; AzureOpenAIStructuredLLM missing _parse_response method — different defect), merged PR #5548 (same file, generate_response reasoning-model params — different function), PR #2914 (closed; original structured Azure LLM addition).

Happy to open a PR mirroring the sibling conversion from azure_openai.py:20-38 at the top of __init__ (None → AzureOpenAIConfig(), dict → AzureOpenAIConfig(**config), plain BaseLlmConfig → AzureOpenAIConfig(...)) plus tests for the three currently-crashing forms — say the word (per CONTRIBUTING I'll wait for the accepted label first).

AI Assistance

AI helped me find it, and I reproduced it myself afterwards (the re-run in Actual Behavior is my own; two independent verifiers reproduced it on the same commit before filing).