v2 GenAI handlers regression: cached_content requests now include system_instruction/tools/tool_config → 400 INVALID_ARGUMENT

[lskrajny-marcura]

Summary

In 1.15.1, the GenAI handlers omitted system_instruction, tools, and tool_config from the request when the caller passed cached_content (the Gemini API rejects requests that set them alongside cached content — they must live in the cache). See handle_genai_tools / handle_genai_structured_outputs in instructor/providers/gemini/utils.py @ 1.15.1, which guard on user_cached_content is None with the comment "Adding them causes 400 INVALID_ARGUMENT".

In 1.16.0, from_genai routes to the v2 handlers (instructor/v2/providers/genai/handlers.py), and both GenAIToolsHandler.prepare_request and GenAIStructuredOutputsHandler.prepare_request build base_config unconditionally — the cached-content guard is gone. Every structured call that passes config={"cached_content": ...} now fails with:

400 INVALID_ARGUMENT: Tool config, tools and system instruction should not be set in the request when using cached content.

Repro (no network needed — inspect the prepared request)

from instructor.mode import Mode
from instructor.utils.providers import Provider
from instructor.v2.core.registry import mode_registry
import instructor.v2.providers.genai.handlers  # noqa: F401
from pydantic import BaseModel

class Out(BaseModel):
    x: int

for mode in (Mode.JSON, Mode.TOOLS):
    prepare = mode_registry.get_handler(Provider.GENAI, mode)
    _, prepared = prepare(Out, {
        "messages": [{"role": "user", "content": "hi"}],
        "config": {"cached_content": "cachedContents/abc"},
    })
    cfg = prepared["config"]
    print(mode, {
        "system_instruction": cfg.system_instruction is not None,
        "tools": cfg.tools is not None,
        "tool_config": cfg.tool_config is not None,
        "cached_content": cfg.cached_content,
    })

1.16.0 output: JSON mode sets system_instruction (instructor's own JSON instruction); TOOLS mode sets system_instruction, tools, and tool_config — all alongside cached_content. Sending any of them with cached content is a guaranteed 400 from the Gemini API.

Expected

When cached_content is present in the user config, prepare_request should omit system_instruction, tools, and tool_config, as 1.15.1 did (https://ai.google.dev/gemini-api/docs/caching).

Environment

• instructor 1.16.0, google-genai 1.75.0, Python 3.12

[princethummar1]

@lskrajny-marcura