Issues using Google Gemini with graphiti-core==0.9.0

[CaliLuke]

Environment:

• graphiti-core Version: 0.9.0 (latest release on PyPI as of 2025-04-08)
• google-generativeai Version: 0.8.4 (installed via pip install "graphiti-core[google-genai]" and also explicitly)
• Python Version: 3.13.2
• Operating System: macOS Sequoia
• Installation Method: uv sync / pip install within a virtual environment

Description: When attempting to initialize graphiti_core.Graphiti with GeminiClient and GeminiEmbedder according to the README instructions, several issues arise within the graphiti-core==0.9.0 library code itself.

Issue 1: Incorrect google.generativeai Import Statements

• Files:
  • graphiti_core/llm_client/gemini_client.py
  • graphiti_core/embedder/gemini.py
• Problem: Both files use incorrect import statements:

  from google import genai  # Incorrect
  from google.genai import types # Incorrect
  

• Error: This leads to ImportError: cannot import name 'genai' from 'google' when trying to import GeminiClient or GeminiEmbedder.
• Expected: The imports should likely be:

  import google.generativeai as genai
  from google.generativeai import types
  

• Workaround: Temporarily patching these import lines in the installed library files resolves this specific error.

Issue 2: Incorrect google.generativeai Client Initialization

• Files:
  • graphiti_core/llm_client/gemini_client.py (__init__ method)
  • graphiti_core/embedder/gemini.py (__init__ method)
• Problem: After patching the imports, the code fails when trying to initialize the Google client using self.client = genai.Client(...).
• Error: This leads to AttributeError: module 'google.generativeai' has no attribute 'Client'.
• Likely Cause: The google-generativeai library (v0.8.4) likely uses a different initialization pattern (e.g., genai.configure(api_key=...) followed by genai.GenerativeModel(...) or specific embedding functions) than what graphiti-core==0.9.0 expects.
• Workaround: Temporarily patching the __init__ methods to use genai.configure(api_key=...) allows initialization to proceed further, but subsequent calls within the methods might still fail.

Issue 3: Required OpenAI Key for Reranker

• File: graphiti_core/graphiti.py (__init__ method)
• Problem: The Graphiti class constructor defaults to initializing OpenAIRerankerClient() if no cross_encoder is explicitly provided. This initialization requires an OpenAI API key.
• Error: If only a GEMINI_API_KEY is set in the environment, initializing Graphiti (even with GeminiClient and GeminiEmbedder) fails with openai.OpenAIError: The api_key client option must be set....
• Workaround: Provide both GEMINI_API_KEY and OPENAI_API_KEY environment variables, or pass a custom (potentially dummy) cross_encoder instance during Graphiti initialization.

Summary: Using Gemini with graphiti-core==0.9.0 currently requires patching the library's internal code due to incorrect imports and client initialization calls, and also requires providing an OpenAI API key for the default reranker.

[danielchalef]

Thanks for reporting. The import issue is resolved here: https://github.com/getzep/graphiti/pull/336/

[danielchalef]

Duplicate of https://github.com/getzep/graphiti/issues/333