langchain-google-genai

langchain-google-genai is the officially maintained bridge between LangChain and Google’s unified google-genai SDK, giving Python applications a drop-in LangChain interface to Gemini models. It ships three primary classes — ChatGoogleGenerativeAI, GoogleGenerativeAIEmbeddings, and GoogleGenerativeAI — that translate LangChain’s message, tool-call, and structured-output conventions into Gemini’s request/response schema and back, so a Gemini-backed chain looks and behaves like any other LangChain chat model.

Beyond basic chat completion, the package handles the harder integration surface: multimodal input (images, audio, video, PDFs) converted into Gemini’s Part/Blob types, function calling translated to and from Gemini’s FunctionDeclaration format including Pydantic-schema-based structured output, safety-setting and generation-config passthrough, code execution and thinking-config support, and context caching helpers. It lives inside the langchain-ai/langchain-google monorepo alongside the community and vertexai integration packages, and is the recommended path for developers using the Gemini Developer API (as opposed to the Vertex AI-specific package).

What You Get

• ChatGoogleGenerativeAI — a LangChain BaseChatModel implementation for Gemini’s chat/completion API, with streaming, async, and tool-calling support
• GoogleGenerativeAIEmbeddings — LangChain-compatible embeddings backed by Gemini’s embedding models
• GoogleGenerativeAI — a simpler LLM-interface wrapper for non-chat text generation use cases
• Multimodal input handling — automatic conversion of images, audio, video, and PDF bytes/URIs into Gemini’s Part and Blob types
• Function calling and structured output — Pydantic models, dicts, and callables converted to Gemini FunctionDeclarations, with parsed structured responses returned as typed objects
• Context caching helper (create_context_cache) for reusing large prompt prefixes across requests

Common Use Cases

• Swapping Gemini into an existing LangChain chain or LangGraph agent without changing surrounding orchestration code
• Building multimodal chat applications that mix text, images, and audio in a single Gemini conversation
• Using Gemini’s function calling for tool-using agents with Pydantic-defined tool schemas
• Generating embeddings with Gemini’s embedding models for retrieval-augmented generation pipelines
• Requesting structured (schema-validated) JSON output from Gemini via LangChain’s with_structured_output pattern

Under The Hood

Architecture The package centers on chat_models.py, a single large module implementing ChatGoogleGenerativeAI by extending LangChain’s BaseChatModel and overriding its generation, streaming, and async hooks; nearly all Gemini-specific translation logic — message history parsing, tool-call conversion, safety settings, generation config — is concentrated there, with embeddings.py and llms.py as thinner sibling wrappers for the embedding and plain-LLM interfaces respectively. Supporting internal modules (prefixed with an underscore: _function_utils.py for tool/schema conversion, _image_utils.py for multimodal Part construction, _enums.py mirroring Gemini’s HarmCategory/Modality enums, _common.py and _compat.py for shared helpers and version-compat shims) keep the public surface small while isolating Gemini API quirks. The package sits alongside langchain-google-community and langchain-google-vertexai in the same monorepo, sharing conventions but each independently versioned and published.

Tech Stack Built for Python 3.10+ using hatchling as the build backend and uv for dependency locking. Core runtime dependencies are langchain-core (the LangChain abstraction layer being implemented against), google-genai (Google’s unified SDK covering both the Gemini Developer API and Vertex AI), pydantic v2 for schema validation and structured-output parsing, and filetype for MIME-type sniffing of multimodal inputs. Dev tooling includes ruff configured against a near-complete rule set with narrowly scoped, explained ignores, and mypy in strict mode.

Code Quality Test coverage is extensive relative to the package’s size: unit tests cover chat models, embeddings, LLMs, function-utils conversion, image-utils conversion, content conversion, and import surface, plus a standard-interface conformance suite (test_standard.py) built on LangChain’s shared langchain-tests harness — the same compliance suite other LangChain model integrations are tested against. Snapshot testing (syrupy) covers serialization-shape regressions, pytest-socket blocks accidental live network calls in unit tests, and a separate integration_tests/ directory holds tests that hit the real Gemini API. mypy strict mode with disallow_untyped_defs is enabled, and CI in the parent monorepo runs dedicated lint/test/integration-test workflows per package.

API Design The integration’s value is almost entirely in fidelity and ergonomics rather than novel mechanics: it normalizes Gemini’s request/response shapes (Content/Part/FunctionDeclaration/GenerateContentConfig) into LangChain’s common message and tool-call conventions closely enough that Gemini-backed chains are largely interchangeable with other LangChain chat models, while still exposing Gemini-specific capabilities — code execution, thinking/reasoning config, context caching, safety settings — as opt-in parameters rather than hiding them. Getting started requires only an API key and the standard LangChain chat-model constructor pattern, with no Gemini-specific boilerplate for common flows.