[Bug]: GoogleCV (Google Vertex vision) provider crashes with AttributeError when key is a JSON non-object (same pattern as #17204 / #17373 / #17389 / #17456 / #17458)

[Harsh23Kashyap]

Self Checks

• I have searched for existing issues, including closed ones.
• I confirm that I am using English to submit this report.
• Non-English title submissions will be closed directly.
• I have not modified the template and have filled in all the required fields.

RAGFlow workspace code commit ID

main @ 9b0719fa9 (also reproducible on any v0.26.x).

RAGFlow image version

v0.26.4 (also reproducible on the current dev image).

Other environment information

• Provider: Google factory (covers AnthropicVertex and GeminiVertex via Google Cloud Vertex AI)
• Affected class: GoogleCV (the CV/vision side of the Google Vertex provider) — rag/llm/cv_model.py:1295-1322
• Companion class (already actively being worked on by another contributor, NOT in scope for this issue): GoogleChat — rag/llm/chat_model.py:1258-1289. Open PR fix(llm): implement async chat methods for GoogleChat (Vertex AI Gemini) #15994 by @glu000 addresses the async chat methods on this class. The JSON-decode bug at chat_model.py:1284 is documented in the issue but the chat side is intentionally left out of this fix to avoid a conflicting PR.
• Also related: BaiduYiyanEmbed (rag/llm/embedding_model.py:1014) has the same shape and will be filed separately.

Actual behavior

The GoogleCV.__init__ method parses its API key as JSON to extract the Google Cloud service-account fields:

# rag/llm/cv_model.py:1319-1322
key = json.loads(key)
access_token = json.loads(base64.b64decode(key.get("google_service_account_key", "")))
project_id = key.get("google_project_id", "")
region = key.get("google_region", "")

There is no try/except around the json.loads(key) call. A user pasting any input that is NOT a JSON object — for example:

• key = '[1, 2, 3]' parses to a list → .get("google_service_account_key", "") raises AttributeError: 'list' object has no attribute 'get'
• key = '"a plain string"' parses to a string → same AttributeError
• key = 'null' parses to None → same AttributeError
• key = '42' parses to int → same AttributeError
• key = 'true' parses to bool → same AttributeError
• key = '3.14' parses to float → same AttributeError

In every case the call site raises an unhandled AttributeError that bubbles up as a 500 in the API and a stack trace in the server log. This is the same class of bug that PR #17215 fixed for Azure, PR #17377 fixed for Bedrock, PR #17390 fixed for BaiduYiyan, PR #17457 fixed for VolcEngine/Ark, and PR #17459 fixed for OpenRouter.

The Google Cloud Vertex AI provider REQUIRES a JSON object key (per the schema in api/apps/llm_app.py:291 and the test in test/testcases/test_web_api/test_llm_app/test_llm_list_unit.py:635-640):

{
  "google_project_id": "<project-id>",
  "google_region": "us-central1",
  "google_service_account_key": "<base64-encoded-service-account-json>"
}

A user pasting a plain string or any other JSON type gets a AttributeError from inside rag/llm internals — no indication of what they did wrong.

Expected behavior

A user pasting any non-Google-Vertex-shaped key (a plain string, a JSON list, a JSON number, a malformed JSON object) should get a clear, actionable error at provider-init time, not a 500. The provider REQUIRES a JSON object, so a non-object input should raise a clear ModelException(retryable=False) with a message that names the required fields (google_project_id, google_region, google_service_account_key) and points at the schema in api/apps/llm_app.py:291.

Steps to reproduce

1. Start RAGFlow v0.26.4 (any deployment).
2. Try to add a Google Cloud Vertex model provider (Gemini or Claude via Vertex) for vision use via the admin UI (or POST /api/v1/llm/factories).
3. Paste any of these into the API key field:
   • A JSON array like ["a", "b"] — crashes with AttributeError: 'list' object has no attribute 'get'.
   • A JSON number like 42 — same crash.
   • A JSON string like "a plain string" (with quotes) — same crash.
   • A JSON null null — same crash.
   • A JSON bool true — same crash.
4. The server log shows an unhandled AttributeError. The admin UI shows a generic 500 with no actionable hint.

Expected error surface

After the fix, a JSON non-object should raise the same clear error as the other provider helpers (e.g. ModelException(retryable=False) with a message like: "Google Vertex requires a JSON object with 'google_project_id', 'google_region', and 'google_service_account_key'; got <type>").

Out of scope

• The chat counterpart GoogleChat at rag/llm/chat_model.py:1284-1289 has the same shape of bug, but PR fix(llm): implement async chat methods for GoogleChat (Vertex AI Gemini) #15994 is open by @glu000 actively working on that class. The chat-side JSON-decode fix should be coordinated with that contributor to avoid conflicting changes — NOT included in this PR.
• BaiduYiyanEmbed at rag/llm/embedding_model.py:1014 has the same shape and will be filed in a separate issue.

Additional information

Affected files and line numbers (in commit 9b0719fa9):

File	Class	Lines	Current state
rag/llm/cv_model.py	GoogleCV	1319-1322	CRASHES on non-JSON or JSON non-object

The fix should be a single helper in rag/llm/key_utils.py (following the established pattern for _resolve_bedrock_credentials, _resolve_qianfan_credentials, _resolve_volcengine_credentials, and _resolve_openrouter_credentials) with regression tests covering:

1. Plain string key → ModelException (the provider REQUIRES a JSON object; no plain-key fallback)
2. JSON dict with all fields → return parsed fields
3. JSON dict missing google_project_id / google_region / google_service_account_key → default to empty strings (preserves tolerance)
4. JSON list, number, string, null, bool → ModelException (the new fix)
5. Bad JSON string → ModelException (the new fix)
6. Helper propagates the right fields to GoogleCV.__init__ so the class gets exactly what it needs

A PR fixing this is in progress.

[Harsh23Kashyap]

Fix opened in #17464.

Summary of the fix (matches the established pattern from #17215 / #17377 / #17390 / #17457 / #17459):

• Added _resolve_google_service_account_key(key) to rag/llm/key_utils.py. Unlike the Bedrock/BaiduYiyan/VolcEngine/OpenRouter helpers, this one does NOT have a plain-string-key fallback -- the Google Vertex provider genuinely requires a full JSON object.
• Wired it into GoogleCV.__init__ in rag/llm/cv_model.py:1319.
• 19 p0 regression tests in test/unit_test/rag/llm/test_googlecv_json_key_fallback.py (16 on the helper directly, 3 on the constructor with stubbed Google SDK modules).

The helper raises ModelException(retryable=False) on every non-JSON-object input (plain string, malformed JSON, JSON list/number/string/null/bool) with a message that names the required fields (google_project_id, google_region, google_service_account_key) and points at the schema in api/apps/llm_app.py:291.

Out of scope (intentionally):

• The chat counterpart GoogleChat at chat_model.py:1284 -- PR fix(llm): implement async chat methods for GoogleChat (Vertex AI Gemini) #15994 by @glu000 is actively working on that class. The chat-side JSON-decode fix should be coordinated with that PR to avoid a conflicting change.
• The pre-existing access_token = json.loads(base64.b64decode(...)) line still raises JSONDecodeError if google_service_account_key is missing. That's a separate missing-field bug.
• BaiduYiyanEmbed at embedding_model.py:1014 has the same shape and will be filed separately.

Closing this once the PR lands.