feat(agent): Gemini 3.5 Flash on Vertex via service account

echo/agent/agent.py

@@ -5,7 +5,8 @@
 from copilotkit.langgraph import CopilotKitState
 from langchain_core.messages import HumanMessage, SystemMessage
 from langchain_core.tools import tool
-from langchain_google_genai import ChatGoogleGenerativeAI
+from google.oauth2 import service_account
+from langchain_google_vertexai import ChatVertexAI
 from langgraph.checkpoint.memory import MemorySaver
 from langgraph.graph import END, StateGraph
 from langgraph.prebuilt import ToolNode
@@ -94,14 +95,29 @@
 )
 
 
-def _build_llm() -> ChatGoogleGenerativeAI:
+VERTEX_AUTH_SCOPES = ["https://www.googleapis.com/auth/cloud-platform"]
+
+
+def _build_llm() -> ChatVertexAI:
     settings = get_settings()
-    if not settings.gemini_api_key:
-        raise ValueError("GEMINI_API_KEY is required")
 
-    return ChatGoogleGenerativeAI(
-        model=settings.llm_model,
-        google_api_key=settings.gemini_api_key,
+    credentials = None
+    project = settings.vertex_project or None
+    credentials_payload = settings.vertex_credentials or settings.gcp_sa_json
+    if credentials_payload:
+        credentials = service_account.Credentials.from_service_account_info(
+            credentials_payload,
+            scopes=VERTEX_AUTH_SCOPES,
+        )
+        if not project:
+            project = credentials_payload.get("project_id")
+
+    return ChatVertexAI(
+        model_name=settings.llm_model,
+        project=project,
+        location=settings.vertex_location,
+        api_endpoint=settings.vertex_api_endpoint or None,
+        credentials=credentials,
     )
 
 

echo/agent/pyproject.toml

@@ -6,7 +6,8 @@ requires-python = ">=3.11"
 dependencies = [
     "copilotkit>=0.1.77",
     "langgraph>=0.2",
-    "langchain-google-genai>=2.0",
+    "langchain-google-vertexai>=3.2.2",
+    "google-auth>=2.35",
     "fastapi>=0.115",
     "uvicorn[standard]>=0.30",
     "httpx>=0.27",

echo/agent/settings.py

@@ -1,13 +1,30 @@
+import json
 from functools import lru_cache
+from typing import Any, Optional
 
-from pydantic import Field
+from pydantic import Field, field_validator
 from pydantic_settings import BaseSettings
 
 
 class Settings(BaseSettings):
     echo_api_url: str = Field(default="http://localhost:8000/api", alias="ECHO_API_URL")
-    gemini_api_key: str = Field(default="", alias="GEMINI_API_KEY")
-    llm_model: str = Field(default="gemini-3-pro-preview", alias="LLM_MODEL")
+    # Vertex model id (no provider prefix). Gemini 3.x is served on the
+    # global Vertex host only; see vertex_api_endpoint below.
+    llm_model: str = Field(default="gemini-3.5-flash", alias="LLM_MODEL")
+    vertex_location: str = Field(default="eu", alias="VERTEX_LOCATION")
+    # Pinning the global host while keeping locations/<region> in the request
+    # path mirrors the server's LiteLLM config (validated in production):
+    # the regional eu-aiplatform host 404s for gemini-3.x models.
+    vertex_api_endpoint: str = Field(
+        default="aiplatform.googleapis.com", alias="VERTEX_API_ENDPOINT"
+    )
+    vertex_project: str = Field(default="", alias="VERTEX_PROJECT")
+    # Service-account JSON blob. VERTEX_CREDENTIALS wins over GCP_SA_JSON;
+    # with neither set, Application Default Credentials apply.
+    vertex_credentials: Optional[dict[str, Any]] = Field(
+        default=None, alias="VERTEX_CREDENTIALS"
+    )
+    gcp_sa_json: Optional[dict[str, Any]] = Field(default=None, alias="GCP_SA_JSON")
     agent_graph_recursion_limit: int = Field(
         default=80,
         alias="AGENT_GRAPH_RECURSION_LIMIT",
@@ -17,6 +34,18 @@ class Settings(BaseSettings):
         alias="AGENT_CORS_ORIGINS",
     )
 
+    @field_validator("vertex_credentials", "gcp_sa_json", mode="before")
+    @classmethod
+    def _parse_service_account_json(cls, value: Any) -> Optional[dict[str, Any]]:
+        if value is None or isinstance(value, dict):
+            return value
+        if isinstance(value, str):
+            stripped = value.strip()
+            if not stripped:
+                return None
+            return json.loads(stripped)
+        raise ValueError("Expected a JSON object or JSON string")
+
     class Config:
         env_file = ".env"
         extra = "ignore"

echo/agent/tests/test_settings.py

@@ -4,15 +4,18 @@
 def test_settings_reads_env(monkeypatch):
     get_settings.cache_clear()
     monkeypatch.setenv("ECHO_API_URL", "http://example.test/api")
-    monkeypatch.setenv("GEMINI_API_KEY", "test-key")
+    monkeypatch.setenv("GCP_SA_JSON", '{"type": "service_account", "project_id": "proj-1"}')
+    monkeypatch.setenv("VERTEX_LOCATION", "europe-west4")
     monkeypatch.setenv("LLM_MODEL", "gemini-test")
     monkeypatch.setenv("AGENT_GRAPH_RECURSION_LIMIT", "64")
     monkeypatch.setenv("AGENT_CORS_ORIGINS", "http://localhost:1111,http://localhost:2222")
 
     settings = get_settings()
 
     assert settings.echo_api_url == "http://example.test/api"
-    assert settings.gemini_api_key == "test-key"
+    assert settings.gcp_sa_json == {"type": "service_account", "project_id": "proj-1"}
+    assert settings.vertex_location == "europe-west4"
+    assert settings.vertex_api_endpoint == "aiplatform.googleapis.com"
     assert settings.llm_model == "gemini-test"
     assert settings.agent_graph_recursion_limit == 64
     assert settings.agent_cors_origins == "http://localhost:1111,http://localhost:2222"