We don't publish
your competitive advantage.

IFFastAPI 0.113.0 or later breaks vLLM's OpenAI-compatible API due to a Pydantic compatibility issue with TypedDicts.

THENDowngrade FastAPI to version 0.111.0 or 0.112.2, or upgrade Pydantic to version 2.9.0 or later. Alternatively, modify the request handler to use `Annotated[dict, ChatCompletionRequest]` instead of the direct type hint.

IFllama-index-vector-stores-chroma 0.0.1 causes a ResolutionImpossible error because it pins llama-index-core<0.10.0 and >=0.9.32, conflicting with other packages that require llama-index-core>=0.10.1.

THENUpgrade llama-index-vector-stores-chroma to version 0.1.0 or later, which depends on llama-index-core==0.10.0, or adjust requirements to use a compatible set of versions. For example, replace 'llama-index-vector-stores-chroma==0.0.1' with 'llama-index-vector-stores-chroma>=0.1.0' in your requirements file.

IFInstalling flash-attn (FlashAttention-2) with ChatTTS results in slower text-to-speech generation compared to not using it.

THENDo not install flash-attn when using ChatTTS. If already installed, uninstall it via pip uninstall flash-attn. The standard attention mechanism in ChatTTS is more performant for this model.

IFResolving llama-index package dependencies fails with 'ResolutionImpossible' due to incompatible version requirements, e.g., llama-index-vector-stores-chroma 0.0.1 requires llama-index-core<0.10.0 while other packages require >=0.10.1.

THENUpgrade llama-index-vector-stores-chroma to version 0.1.0 or later, which depends on llama-index-core==0.10.0, and ensure all llama-index packages use compatible version ranges. Consider using a requirements file with loosely pinned versions or allow pip to resolve conflicts by not pinning sub-dependencies.

IFWhen using transformers 4.52.2 with PyTorch <2.5, importing any model (e.g., RobertaModel) fails with NameError: name 'Replicate' is not defined.

THENUpgrade PyTorch to version 2.5 or later, or downgrade transformers to version 4.52.1. The error occurs because `Replicate` is conditionally imported only for torch >=2.5 but is used unconditionally in the tensor_parallel module.

IFBuild fails when installing litellm >= 1.77.0 due to missing 'madoka' wheel which requires C++ compilation (e.g., error 'g++: No such file or directory' or 'stdexcept' not found).

THENPin litellm to version <1.77.0 (e.g., 1.76.x) in your requirements. For pip: `litellm>=1.76.0,<1.77.0`. For uv: `uv add litellm==1.76.1`. If you must use a newer version, install C++ build tools (e.g., build-essential on Linux, Xcode CLI tools on macOS).

IFDocArrayInMemorySearch fails with pydantic ValidationError — missing required fields 'text' and 'metadata' in DocArrayDoc when using pydantic v2.

THENPin pydantic to version 1.10.x to avoid the validation error. Run `pip install pydantic==1.10.8` to force a compatible version until LangChain updates its pydantic v2 support.

IFWhen using ChatVertexAI from langchain-google-vertexai with pydantic version 2.10 or higher, a PydanticUndefinedAnnotation error occurs because 'SafetySetting' is not defined.

THENDowngrade pydantic to version 2.9.0 to resolve the forward reference error. Alternatively, wait for a fix in the langchain-google-vertexai package as the issue stems from an upstream incompatibility.