langchain-core==1.5.4 リリースlangchain-core==1.5.4

LangChain has published langchain-core&tag=langchain-core&entry=152874e59b1278f2">langchain-core 1.5.4, a maintenance release that concentrates on bug fixes rather than new capabilities. For teams building agents and language-model applications on the LangChain stack, patch releases like this one matter because langchain-core&tag=langchain-core&entry=152874e59b1278f2">langchain-core is the foundational package that most other LangChain libraries depend on. Small regressions in its runnables, prompt objects, and tool-handling logic can propagate widely, so incremental fixes help keep downstream projects stable.

The most prominent change in this release is a compatibility fix for Pydantic 2.14 (#39328). Pydantic is the data-validation library that LangChain relies on heavily to define schemas for tool arguments, structured outputs, and configuration objects. Because langchain-core&tag=langchain-core&entry=152874e59b1278f2">langchain-core supports a broad range of Pydantic versions, each new Pydantic release can surface incompatibilities that break validation or schema generation. The fix is intended to keep langchain-core&tag=langchain-core&entry=152874e59b1278f2">langchain-core functioning correctly against Pydantic 2.14, which is important given how quickly the Pydantic maintainers ship point releases. Users who upgrade their environment and pull in a newer Pydantic could otherwise encounter errors that originate in that dependency rather than in their own code.

A second fix stops StructuredPrompt from mutating caller kwargs (#39174). StructuredPrompt is a prompt construction utility used when an application needs the model to return data conforming to a defined structure. The reported issue appears to have been that the object modified the keyword-argument dictionary passed into it, which is a side effect that can produce hard-to-trace bugs when the same arguments are reused elsewhere in a program. Preventing in-place mutation of caller-supplied data is a common defensive-programming practice, and this change brings StructuredPrompt in line with that expectation.

The release also preserves the flat tool args schema for RootModel runnables (#39307). In LangChain, tools expose an argument schema so that a model can understand what inputs a function expects and generate a correctly shaped call. RootModel is a Pydantic construct used to wrap a single value or a non-dictionary root type. The fix ensures that when a runnable is defined around a RootModel, the resulting tool argument schema remains flat rather than being nested or otherwise restructured, which helps models produce valid tool calls and keeps behavior consistent with other tool definitions. The source notes indicate at least one additional fix beyond these, though the changelog entry is truncated, so the full set of adjustments should be confirmed against the official release notes.

For readers less familiar with the ecosystem, langchain-core&tag=langchain-core&entry=152874e59b1278f2">langchain-core sits at the bottom of the LangChain architecture. It defines the core abstractions such as runnables, messages, prompts, output parsers, and the tool interface, while higher-level packages like langchain, the various provider integrations, and the LangGraph library for building stateful agent workflows build on top of it. Because these pieces are versioned somewhat independently, keeping the core package robust against dependency shifts is a recurring maintenance task. The version number, 1.5.4, follows semantic-versioning conventions in which the final digit signals a patch focused on fixes rather than new interfaces, so upgrading from 1.5.3 is likely to be low risk for most projects.

The Pydantic-related work in particular reflects a broader reality of the Python AI tooling landscape: frameworks that lean on validation libraries must track their release cadence closely. Pydantic's v2 line, which was rewritten with a Rust-based core for performance, has continued to evolve, and libraries that depend on it periodically need compatibility patches to avoid breakage. LangChain is not alone here; many agent and data-pipeline frameworks pin or test against specific Pydantic ranges for the same reason. Adjacent projects such as instructor and various structured-output tools face similar pressures.

Practically, developers running langchain-core&tag=langchain-core&entry=152874e59b1278f2">langchain-core should review the linked pull requests and their own dependency pins before upgrading, especially if they use StructuredPrompt or define tools around RootModel types. Those who have deferred moving to Pydantic 2.14 may find this release removes a blocker. As always with patch releases, the changes are targeted and incremental, and teams with comprehensive test suites are best positioned to verify that the fixes behave as expected in their specific configurations before deploying to production.