fix: isolate Vertex AI credentials within each embedding client

AF
8h ago
+141-2

Summary

VertexAIEmbeddingConfig.get_client() currently writes every configuration's service-account key to the same file and replaces the process-wide GOOGLE_APPLICATION_CREDENTIALS setting. Creating one embedding client can therefore replace another component's application-default credential selection, and separate configurations share a mutable credential file.

Parse the provided service-account JSON into a Google Auth credentials object and pass that object and its project directly to VertexAIEmbeddings. Each client retains its own account, and construction no longer changes the environment or creates a key file. The existing explicit register_application_credentials() method remains available.

Testing

  • On unchanged ac932052e6e0ba3feab8a6e96b6d4480218e009f, the seven new regression cases produce 6 failures and 1 passing control. With the fix, the new cases and the neighboring Vertex encoder test produce 8 passes.
  • Tests construct the real Vertex AI SDK client with locally generated RSA service-account keys. They check credential identity and project, verify a real Google Auth RSA signature, construct two independent clients in both orders, preserve an existing application-default credentials setting, reject malformed JSON, and ensure no credential file is created. They make no embedding or authentication requests.
  • Verified with Python 3.12.16, langchain-google-vertexai 3.3.0, google-auth 2.62.2 and google-genai 2.34.0. SDK-dependent regression cases skip when the optional modern SDK is unavailable; existing tests retain their normal coverage.
  • Changed-file Ruff checks, Ruff formatting and source integrity checks pass.

Reproducible fork validation: https://github.com/Afloat16/unstructured/actions/runs/43843777142

The full repository suite and hosted Vertex embedding service were not run.